Procedural Generation in

Game Design 
Procedural Generation in
Game Design 

Edited by
Tanya X. Short and Tarn Adams
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

©  2017 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-1-4987-9919-5 (paperback)

978-1-138-74331-1 (hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts
have been made to publish reliable data and information, but the author and publisher cannot assume
responsibility for the validity of all materials or the consequences of their use. The authors and publishers
have attempted to trace the copyright holders of all material reproduced in this publication and apologize to
copyright holders if permission to publish in this form has not been obtained. If any copyright material has
not been acknowledged please write and let us know so we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmit-
ted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented,
including photocopying, microfilming, and recording, or in any information storage or retrieval system,
without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.
com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood
Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and
registration for a variety of users. For organizations that have been granted a photocopy license by the CCC,
a separate system of payment has been arranged.

Trademark Notice:  Product or corporate names may be trademarks or registered trademarks, and are used
only for identification and explanation without intent to infringe.

Library of Congress Cataloging-in-Publication Data 

Names: Short, Tanya X., author. | Adams, Tarn, author.

Title: Procedural generation in game design / authors, Tanya X. Short and
Tarn Adams.
Description: Boca Raton : Taylor & Francis, CRC Press, 2017. | Includes
bibliographical references and index.
Identifiers: LCCN 2016054480| ISBN 9781498799195 (pbk. : alk. paper) | ISBN
9781138743311 (hardback : alk. paper)
Subjects: LCSH: Level design (Computer science) | Compter games--Design.
Classification: LCC QA76.76.C672 S543 2017 | DDC 794.8/1536--dc23
LC record available at https://lccn.loc.gov/2016054480

Visit the Taylor & Francis Web site at 

http://www.taylorandfrancis.com 
and the CRC Press Web site at 
http://www.crcpress.com 
Contents

Preface, xvii

Section I  Procedural Generation

Chapter 1   ◾   When and Why to Use Procedural Generation 3

Darren Grey
PLANNING WHEN TO USE PROCEDURAL GENERATION 3
Integral 3
Drafting Content 4
Modal 5
Segmented 5
WHEN PCG IS A BAD IDEA 5
Quality Assurance 6
Time Restrictions 6
Authored Experience 7
Multiplayer 7
Just Random 8
Overreliance on PCG 8
WHY USE PROCEDURAL GENERATION? 8
Utilitarian 8
Unique 10

Chapter 2   ◾   Managing Output: Boredom versus Chaos 13

Kenny Backus

v
vi   ◾    Contents

Chapter 3   ◾   Aesthetics in Procedural Generation 23

Liam Welton
ATMOSPHERE IN SUNLESS SEA  24
ESTABLISHING THE RULES 25
BLURRING THE BOUNDARIES 26
THE RESULT 27

Chapter 4   ◾   Designing for Modularity 29

Jason Grinblat
MODULES AND GESTALTS 29
ASSEMBLY MECHANISMS AND GESTALT SPACES 30
ENABLING PLAY 33
Mechanics as Shared Substrates 33
Orthogonality 35
Equivalence of Impact 37
PLOTTING DESIRABLE GESTALTS 38
INSERTING MEMORABLE ASYMMETRY 40
REFERENCE  41

Chapter 5   ◾   Ethical Procedural Generation 43

Dr. Michael Cook
TALKING IN CODE 44
THE BIG WIDE WORLD 46
YOU ARE WHAT YOU EAT 50
TALKING THE TALK 51
THE FUTURE 54

Section II  Procedural Content

Chapter 6   ◾   Level Design I: Case Study 57

Chris Chung
OVERVIEW 58
THE RULES 59
 Contents    ◾    vii

HOW IT WORKS 60
CONCLUSION 62

Chapter 7   ◾   Level Design II: Handcrafted Integration 63

Jim Shepard
STANDARD DUNGEONS 63
CRYPT GENERATION 66
BEST PRACTICES 68
DUNGEONMANS DUNGEON GENERATION PSEUDOCODE  69
DUNGEONMANS CRYPT GENERATION PSEUDOCODE  70

Chapter 8   ◾   Level Design III: Architecture and Destruction 73

Evan Hahn
ARCHITECTURE GENERATION 74
Step 1: Calculate the Bounding Box 75
Step 2: Split the Box into Regions 75
Step 3: Skim Perimeter Regions 76
Step 4: Place Connections 78
Step 5: Assign Region Types 80
Step 6: Make Adjustments 81
Step 7: Generate the Regions 81
DISCUSSION 82

Chapter 9   ◾   Cyclic Generation 83

Dr. Joris Dormans
CYCLES 84
USING GRAPHS TO EXPRESS CYCLES 84
PATTERNS 87
IMPLEMENTATION 89
ASIDE: LOCK-AND-KEY ATTRIBUTES 91
Locks Might Be Conditional, Dangerous, or Uncertain 91
Locks Are Permanent, Reversible, Temporary, or Collapsing 92
Locks Might Be Valves or Asymmetrical 92
Locks and Keys Can Be Safe or Unsafe 92
viii   ◾    Contents

Keys Can Be Single Purpose or Multipurpose 92

Keys Are Particular or Nonparticular 93
Keys Might Be Consumed or Persistent 93
Keys Might Be Fixed in Place 93
TILEMAPS 93
DISCUSSION 95

Chapter 10   ◾   Worlds 97
Dr. Mark R. Johnson
BRIEF HISTORY OF WORLD GENERATION 98
WHY MAKE WORLDS? 101
Exploration 101
Expansive or Complex Worlds 102
Gameplay Variation 103
QUALITATIVE PROCEDURAL GENERATION 103

Chapter 11   ◾   Puzzles 107

Danny Day
PROCEDURALLY GENERATING PUZZLES 107
Puzzle-Spaces 108
Desired Outputs 108
PUZZLE GENERATION APPROACHES 109
Random Start State 109
Backward from Goal State 110
Heuristics 110
Extra Bonus: Permutations 111
DESKTOP DUNGEONS, THE PUZZLE ROGUELIKE 112
More Puzzle than Roguelike? 112
More Roguelike than Puzzle? 112
PLAYER HOPE AS A RESOURCE 113
Guaranteeing Solvability 114
Generating Hope 115
CONCLUSION 117
 Contents    ◾    ix

Chapter 12   ◾   Procedural Logic 119

Ben K ane
BACKGROUND 119
USUAL APPROACH TO PROCEDURAL GENERATION 120
A DIFFERENT APPLICATION: PROCEDURAL LOGIC 121
HOW THE RULE LOGIC IS GENERATED 122
Trivial Case 122
Countable Problems 123
Not So Trivial: Procedural Logic 124
Generating Rules 125
Queries: Asking Questions 125
Solutions: Taking Action 126
IMPROVING THE PROCESS 127
Better Queries 127
Compound Queries 127
Query Contexts 127
Better Solutions 128
RIGGING THE DECK: RANDOM THAT FEELS GOOD 129
Dealing with Degenerates 129
Avoiding the Impossible 130
PUTTING IT ALL TOGETHER 131
CONCLUSION 132

Chapter 13   ◾   Artificial Intelligence 133

Dr. Mark R. Johnson
UNPREDICTABILITY AND ARTIFICIAL INTELLIGENCE 134
MOVEMENT AND COMBAT 135
AMBIENT BEHAVIOR 137
EMERGENT PHENOMENA 137
CONVERSATIONS 138
Dialects 138
Conversation System 140
CONCLUSION 141
x   ◾    Contents

Chapter 14   ◾   Procedural Enemy Waves 143

Wyatt Cheng
METHOD 1: SPAWN BY TIMER 144
Spawn by Timer Pseudocode 144
Commentary 144
Key Characteristics 146
METHOD 2: SPAWN ON COMPLETION 146
Spawn on Completion Pseudocode 146
Commentary 147
Key Characteristics 147
METHOD 3: CONTINUOUSLY ESCALATING TOTAL 148
Continuously Escalating Total Pseudocode 148
Commentary 149
Key Characteristics 150
METHOD 4: HITPOINT PROGRESSION 150
Hitpoint Progression Pseudocode 150
Commentary 151
Key Characteristics 151
CONCLUSION 152

Chapter 15   ◾   Generative Artwork 153

Loren Schmidt
TECHNIQUES 154
PERCEPTION OF INTENT 158

Chapter 16   ◾   Generative Art Toys 161

K ate Compton
BUILDING ART TOYS, FOR EXPERTS AND NOVICES 161
EXPERIENCE OF ART TOYS 163
TRADING CONTROL FOR POWER 164
DESIGN AND CONSTRUCTION OF ART TOYS 166
Inputs 166
Data and Transformations 168
 Contents    ◾    xi

Points and Rotations 168

Connectivity and Meshes 168
Gestural Curves 168
Forces and Acceleration 169
Rendering 169
OUTSIDE THE GENERATOR: JUDGMENT, SHARING,
AND CURATION 172
CONCLUSIONS: CREATIVITY FOR ANYONE 173

Chapter 17   ◾   Audio and Composition 175

Bronson Zgeb
PROCEDURAL AUDIO IN SKIPPING STONES 175
Sampling 175
Pitch 176
Implementation 177
PROCEDURAL COMPOSITION IN SKIPPING STONES 177
Beat 178
Scale 179
Steps 179
Chords 181
Motif and Repetition 181
CONCLUSION 183

Section III  Procedural Narrative

Chapter 18   ◾   Story and Plot Generation 187

Ben Kybartas
GRAMMARS AND STORY GRAMMARS 188
GAME WORLD 190
STORY MODEL 191
RULE DESIGN 192
SECONDARY REWRITE RULES 194
GAME WORLD SIMULATION 195
xii   ◾    Contents

METRIC-GUIDED GENERATION 196

CONCLUSIONS 197

Chapter 19   ◾   Emergent Narratives and Story Volumes 199

Jason Grinblat
MOTIVATION 199
THE VOICE IN THE MACHINE 201
FIASCO: A CASE STUDY 203
UNPACKING YOUR THEMES 205
GRIST FOR THE NARRATIVE MILL 207

Chapter 20   ◾   Poetry Generation 209

Harry Tuffs

Chapter 21   ◾   Characters and Personalities 215

Emily Short
INTRODUCTION 215
SOURCE MATERIAL 216
REALIZATION 217
Selecting Dialogue 217
Layering Dialogue Features 218
Character and World Interaction 220
CREATION 220
Orthogonal 221
Mechanically Significant 221
Easy to Communicate 222
Meaningful in Combination 222
RECURRING STRATEGIES 223
Combining Output from Several Layers of Simulation or
Gameplay 223
Bringing Character into Every Interaction 223
Juxtaposing Events and Interpretation 224
Callbacks to Earlier Events 225
 Contents    ◾    xiii

PITFALLS 225
Overgeneralization 225
Overrealism 226
Untamed Simulation 227

Section IV  The Procedural Future

Chapter 22   ◾   Understanding the Generated 231

Gillian Smith
EXPRESSIVE RANGE AND GENERATIVE SPACES 232
QUALITIES OF THE GENERATED 232
Types of Qualities 232
Example 233
Formalizing Qualities into Metrics 235
Example 235
Metrics versus Requirements 237
QUALITIES OF THE GENERATOR 238
VISUALIZING EXPRESSIVE RANGE 239
Histograms 239
Example 240
Distance-Based Clustering 242
CONCLUSION 243
REFERENCES  243

Chapter 23   ◾   Content Tools Case Study 245

Kepa Auwae
SYSTEM OVERVIEW 247
EXAMPLE ROOM 248

Chapter 24   ◾   Automated Game Tuning 251

A aron Isaksen
STEP 1: SET PARAMETERS 253
STEP 2: GENERATE THE GAME LEVEL 255
xiv   ◾    Contents

STEP 3: SIMULATE THE GAME 255

STEP 4: ANALYZE THE DATA 258
STEP 5: VISUALIZE THE DATA AND MAKE ADJUSTMENTS 261
CONCLUSION 263
ACKNOWLEDGMENTS 264

Chapter 25   ◾   Generating Rules 265

Dr. Michael Cook
MIX AND MATCH 266
GAME DESIGN 101 267
THROWING OUT THE RULEBOOK 269

Chapter 26   ◾   Algorithms and Approaches 271

Brian Bucklew
RANDOM NUMBERS 271
Pseudorandom Number Generators 271
Making Use of Repeatable Series 272
Seeds and Hashing 273
Rolling Dice 274
Normal Distributions 275
Weighted Distributions 277
HEIGHTMAPS 279
Box Linear Filters 279
Midpoint Displacement 280
Perlin and Simplex Noise 281
SEQUENCE GENERATION 283
Lindenmayer Systems (L-Systems) 283
Markov Chains 284
FILLING SPACE 286
Random Walks 286
One-Dimensional Random Walks 286
Two-Dimensional Random Walks 287
 Contents    ◾    xv

Cellular Automata 288

Settling 291
Wang Tiles 291
PARTITIONING SPACE 293
Binary Space Partition 293
Voronoi Diagrams 295
Dijkstra Maps 296
Tree Mapping 297
PUTTING IT ALL TOGETHER 298

Chapter 27   ◾   Meaning 301

Dr. Mark R. Johnson
MEANING IN GAMES 302
MEANINGLESSNESS IN GAMES 304
DESIGNER AND PLAYER MEANING 306
MEANING IN QUALITATIVE PROCEDURAL GENERATION 308
CONCLUSION 311

INDEX, 313
Preface

The term procedural generation  sounds cold and uncaring—it invokes

images of robots building things for practical use rather than for fun,
aesthetics, or enlightenment. In practice, however, I’ ve found that hand-
made content can feel far more artificial than what an algorithm can pro-
duce. When game designers have total control over their creations, there’ s
a strong temptation to overdesign and lead players around by the hand,
making them feel less like a hero on an adventure and more like a child
on a scavenger hunt. In that sense, a little bit of indifference toward the
player can be a good thing. The real world, after all, doesn’ t care if there’ s
anything of interest to us at our destination or even if we have a destina-
tion, and that makes it all the more exciting when something interesting
does happen.
My first experience with procedural generation was with Hack , one of
the first roguelikes released after Rogue . The game’ s ASCII graphics were
primitive looking even by 1982’ s standards, but its smart design based
around randomly generated dungeons made it deep enough to last a life-
time. Fast-forward to 2006: brothers Tarn and Zach Adams released the
first version of Dwarf Fortress , a hybrid simulation/ roguelike that proce-
durally generated not just dungeons but entire continents, populated by
characters with histories stretching back centuries. Visually, the game
was only a few steps above Hack , but the promise of never-ending fantasy
worlds more than made up for it. At the time, I was sorely disappointed
with the beautiful but restrictive titles that were coming out of the main-
stream AAA. Games like Dwarf Fortress  and Hack  were so inspiring to
me that I began experimenting with random-level generation myself, in a
side-scrolling platform game called Spelunky .
I quickly realized how fun it is to build levels with procedural generation.
Making small changes behind the scenes could lead to a cascade of new
possibilities, and the randomization that made roguelikes so replayable

xvii
xviii   ◾    Preface

to players also kept the game fresh for me even after years of working on
it. My experience with Spelunky  also taught me how much handiwork is
involved in making algorithms that generate enjoyable levels. Using pro-
cedural generation doesn’ t mean that there’ s no human hand involved in
the design, but only that the hand is less direct in its control. It’ s similar to
how a painter might hold the paintbrush closer to the end of the handle to
create looser, more painterly brush strokes.
That intersection between handiwork and procedure, art and sci-
ence, still feels vast and wild. It took many years before the importance
of roguelikes became widely accepted. Similarly, it will take some time
before developers begin to fully realize the potential for procedural gen-
eration beyond dungeon building, for things like graphics, music, and sto-
rytelling. Thankfully, many of the creators who are thinking about those
things are in this book, sharing their knowledge and experience.

Derek Yu 
 I
Procedural Generation

P rocedural generation is a powerful tool and a great way to ruin

your game design. A human artist wants to create compelling land-
scapes that aren’ t bland or messy. A human author wants to write an
engaging story that isn’ t repetitive or nonsensical. Even better, they can
do these things in concert, respecting the overall vision of the project.
Procedural content can be difficult to control, and even when the output
is tamed, it might add up to a costly result no better than a scripted one.
Yet the draw of procedural generation is undeniable. Unimagined
visions, replayable levels, reactive music, and unique player-guided nar-
ratives are all within reach. Players can have their own personal journey
but still have enough common experience to share their tales with others,
working together to define what any given procedural game is and what
it can become. Unexplored frontiers are somehow close at hand at every
turn, and even beginning developers can jump right in to see what they
can find.
In this first section, we discuss when to use procedural generation, and
the many concerns that need to be addressed when using these techniques,
whether game balance, object management, adhering to a vision, or han-
dling broader obligations to the community that experiences what you
create. Later sections offer specific techniques and solutions, while here
we will consider the purpose and utility of generation as a tool, providing
a foundation for future explorations.

1
 Chapter 1

When and Why to Use

Procedural Generation
Darren Grey
Roguelike Radio 

P rocedural generation (PCG) techniques are now used widely

through game design and development, from content drafting to on-
the-fly generation, graphical effects to responsive narrative. This chapter
examines the motivations behind including PCG in your development
process, as well as important issues that may hold you back and the risks
when deciding to delve into the depths of generative processes.

PLANNING WHEN TO USE PROCEDURAL GENERATION

Making a game can be an intensive process, and if not planned properly, it
can easily run hugely over budget. A worry with any project is scheduling,
and at what point in your process to instigate risky endeavors like proce-
durally generated content. In game development, this partly comes down
to what way you’ re using PCG in your game’ s design.

Integral
In the most well-known cases, the choice to make use of PCG is wrapped
up in the initial design of the game. Most of the games that make extensive
use of PCG also rely on it entirely to be the type of games they are. Games
with significant random content, such as Rogue  and Spelunky , depend on
PCG to deliver their core gameplay. Games using PCG to include a vast
amount of content, such as Elite  and No Man’ s Sky , are similarly dependent

3
4   ◾    Procedural Generation in Game Design

on PCG to achieve their vision. The decision to use procedural content was
integral to the decision to make that type of game.
It is vital in these cases to consider the scope of PCG involvement in
achieving the aims of the game. The amount of time it takes to produce the
majority of game content by procedural means cannot be underestimated.
This can have a massive impact on project planning. Any changes to your
project direction (which are far from unusual in game development) will
have vast implications for your codebase.
Your algorithms will need to be drafted and redrafted multiple times,
with the rest of the game’ s content built around them. Work on this
will continue throughout the project life cycle, and this should be care-
fully considered when planning resources for development. The debug-
ging and tweaking of the game’ s procedural systems will continue after
the game’ s release, and with the nature of these games, every little
change can have wide-reaching impact on every aspect of the player’ s
experience.
Holistic use of procedural content needs tight development from cradle
to grave.

Drafting Content
PCG is often used in game development to produce a large amount of
content that is then polished later. The typical example of this is an open-
world game, such as Skyrim , which has a vast explorable map. Major
sections of this are produced using procedural techniques and are then
adjusted and polished by hand later. Other examples include puzzle
games where a generator can produce hundreds of examples of solvable
puzzles, and then a human manually picks out or sorts the more inter-
esting ones.
In these cases, the decision to use PCG comes early, but isn’ t needed
from the project inception. Typically, the code used doesn’ t need a huge
amount of finesse, and doesn’ t need to be as carefully honed to fit the rest
of the project. A landscape generator for an RPG just needs to produce
some rough drafts, with the important detail that the player really inter-
acts with filled in by hand. Because of this, the project timescales are more
predictable.
The use of PCG in these areas both starts and finishes early. Once the
generator has done its job, it is locked down, and the flesh-based designers
get to work on honing the outputs to their needs.
 When and Why to Use Procedural Generation    ◾    5

Modal
Some games rely little on PCG, but have a special mode of play that makes
use of procedurally generated content. This is sometimes called an “ infin-
ity mode”  and is a special piece of side content to the main game. Examples
include the Ancient Cave in Lufia II  and procedural maps in Rust .
Since these are separate modes, they can be created later in the develop-
ment process, perhaps released as downloadable content (DLC) after the
game’ s initial release. It can even be something the game’ s modding com-
munity produces, if your game allows them such flexibility.
A procedural mode that mimics the core game may prove to be less
resource-intensive to develop, as the designer already has a clear vision of
how the output should look and how the mechanics will work in this envi-
ronment, and many of the assets will already be in place. However, this
also tends to restrict the potential for interesting and innovative content,
leading to content that is more random than procedural. Many games that
implement this mode of play end up simply relying on waves of random-
ized enemies rather than putting the extra work into having engaging,
procedurally driven gameplay.

Segmented
Certain segments of a game can be squared off from the rest of the game’ s
design to use PCG. A linear, hand-designed game could still have proce-
dural music, for instance. A certain graphical effect could have interesting
randomized elements. A specific room in a single area of the game may
have some special procedural rules in place.
These individual elements of the game can vary widely in scope, but they
tend to be segmented from the rest of the game design. Thus, the “ when” 
of using PCG comes down purely to when this feature needs implement-
ing. And because of the lack of knock-on implications from these sections,
they typically don’ t run a heavy risk of overrun on time. If there is a prob-
lem implementing PCG in these instances, then as a contingency you can
simply revert to hand-generated content instead.

WHEN PCG IS A BAD IDEA

PCG comes with a warning label. It can be a black hole, a slippery slope,
a project risk, and a dark abyss. It’ s important to understand these risks
when looking to integrate PCG into your game, and to realize all the
implications and impacts of PCG across your project.
6   ◾    Procedural Generation in Game Design

Quality Assurance
One of the most prominent blockers to the use of PCG in mainstream
games, in particular AAA games, is quality assurance (QA). A PCG game
may not be guaranteed to work as intended 100% of the time. There is no
way of having testers go through every single iteration of the procedural
content. The generator may have serious bugs in 0.1% of its levels that are
never picked up until it gets released to a wide playerbase.
I have heard multiple first-hand accounts of developers making a pro-
cedural generator for a game, showing their boss how it can make 100 lev-
els in an instant, and the boss saying, “ Just send those 100 levels to the QA
team, but we’ re not having the generator live in the game.”  The perceived
risk of losing control to a procedural system is too great.
The risks can be mitigated with various controls and automated tests.
Connectivity tests for level generators are vital, for instance, but one can
also include minimum difficulty tests or put in further constraints to have
a more controlled experience. This is usually good design too! However,
there is a chance you lose the variety PCG brings to the game, with over-
constrained, “ samey”  levels. It’ s also a lot more work.
One area of PCG that tends to pass this QA test is quilted-content PCG
using premade blocks of content, meshed together on the fly for a varied
experience. This is how the Diablo  games make their levels. But it often
produces the least varied experience, as the player gets to recognize the
content blocks and the patterns produced by the generator. Over repeated
plays, it can produce repetitive and stale gameplay.
Concern over QA is one of the reasons we see the most experimenta-
tion in this field from independent developers. Indies often have no one to
report to, little reputation to sacrifice, and are more willing to take risks.
But indies are also often the ones with the least resources and time to build
well-constrained systems or automated testing scripts.

Time Restrictions
PCG is often touted as a time-saver, but there is no guarantee this is the
case. Designing the generator may prove harder and more nuanced than
initially expected, especially in an area for which there isn’ t a lot of exist-
ing code to rely on. Work on it may also continue after completion, as
constant tweaking is needed to adapt to other developments.
It’ s important to assess how much a reliable time frame is needed for
your project. A handcrafted process takes a more predictable amount of
 When and Why to Use Procedural Generation    ◾    7

time. You might hope that making a level generator will take less time to
make than hand-designing 10 levels, but you don’ t know  that, and in the
end, it might be that 10 levels is just fine for your game.
If money and time are particularly restricted for your game, then think
twice about using significant amounts of PCG, and factor in plenty of con-
tingency to your budget should it not go according to plan.

Authored Experience
In many games, the designer wishes to produce a highly authored experi-
ence. This is particularly notable in story-based games, where you wish the
player to encounter various parts of the game in a linear way. These types
of games benefit less from PCG, and may be made worse if the generated
content doesn’ t fit neatly to the authored experience the developer has in
mind.
This can  be made to work, though. In particular, a thorough and well-
crafted procedural game can take into account the player’ s actions and
integrate their impact into the content produced. This sort of integration
can create a far more effective feeling of an authored experience than a
traditional linear game. It also requires a lot of work and great design to
get right.
Another authored experience where PCG is unsuitable is reaction-
based games that present specific timed and placed challenges. The likes of
Super Meat Boy  expect the player to fail and repeat the same content over
and over until they master that authored part of the game. They can also
include “ gotcha”  moments or clever pieces of design that take the player’ s
thought processes into account, which are harder to reproduce through
procedural means.

Multiplayer
Control of game balance in PCG isn’ t always easy, but it becomes an espe-
cially difficult issue in competitive multiplayer games. In particular, the
likes of maps in real-time strategy games can present challenges, with
starting positions, resources, threats, and topography all affecting how
each player performs. A small difference in balance in a generator can give
one side sufficient advantage to win far more often. This is unacceptable in
games with highly competitive audiences.
Imbalance can be mitigated by having symmetrical maps, but this isn’ t
always desired for aesthetic or other reasons.
8   ◾    Procedural Generation in Game Design

Just Random
If you don’ t have the time and resources to make interesting procedural
content, then typically you end up with something just random. As noted
in Chapter  2, this is often boring and repetitive and doesn’ t make for a
good play experience. Bland, samey content should be avoided, and if
you’ re not able to do PCG properly, then reconsider its use in your game.

Overreliance on PCG
PCG has become a common term in game design and even in game mar-
keting, with games advertising the immense possibility spaces they offer.
But it absolutely cannot be a crutch the game solely relies on. Sheer vari-
ety, even well-designed variety, is not enough if you do not have the core
mechanics and interactions to let the player have fun with what’ s on offer.
No Man’ s Sky  in particular has received criticism for promising “ Every
Atom Procedural”  but ultimately falling short, causing some to accuse it
of shallow gameplay, despite a massive and beautiful universe. Procedural
content should not be relied on to make your game special, and you should
never lose sight of core gameplay experience.

WHY USE PROCEDURAL GENERATION?

Why should you use PCG in game design?
There are utilitarian, level-headed reasons. Perfectly sensible, logical
points in favor of you using PCG. But there are also fun reasons, motiva-
tions for using PCG that are entirely unique to the process. And while
every developer will tell their boss or producer (or themselves) that they
have those level-headed reasons, in the back of their minds they are
dreaming of playing with infinity.

Utilitarian
A few very practical reasons to generate include the following:

• Time-saving : With PCG, you can produce huge amounts of content

far faster than a manual process. Many geographically extensive
games, such as Skyrim , rely on this to generate their massive worlds,
although they polish details by hand afterwards. The two main pro-
cesses involve either procedural world gen followed by polish, or
manual world gen followed by using procedural tools to generate
varied content (trees, grass, buildings, etc.). Note that time-saving is
 When and Why to Use Procedural Generation    ◾    9

sometimes a myth— one can end up spending far more time balanc-

ing and debugging a procedural system than one would spend on
simply making things by hand.
• Expandable : Generators are usually designed in a modular fashion.
Every time you add a new feature, it gets integrated across the gen-
erator’ s output. Each bit of hand-designed polish you add doesn’ t
affect just one item— it affects a countless number of possible out-
puts. This means that extra work is multiplied immensely in impact
across your game.
• Replayability : One generator can produce many instances of simi-
lar but varied content. Roguelikes and roguelike-inspired games in
particular make use of this to produce highly replayable gameplay. It
would be impractical to hand-produce enough varied content to suit
thousands of instances.
• Reusable code : Generators can get reused between applications. A
desert landscape generator from one game can have a few variables
tweaked to produce icy mountains in another. Modular elements of
generators can be swapped between applications— a visual effect can
be copied from one series of an image processing tool to a tool that
produces an entirely different end effect, but still relies on that same
stage. A fire simulator in one game can make a disease spreader in
another. When doing big projects in succession, this sort of reusable
code is invaluable.
• Rules enforcement : Generators can be built with rules enforced,
ensuring connectivity between elements or certain standards main-
tained across all output. The code, assuming it is bug-free and cor-
rectly designed, will be far more rigorous in applying these rules than
a human designer. Architects rely on this now to ensure that build-
ing standards (and the laws of physics) are enforced in new projects.
Game designers can include restrictions on difficulty balance, con-
nectivity, solvability of puzzles, and many other vital features.
• Modeling reality : Often, simulationist techniques are used in proce-
dural generators, reproducing the wearing of terrain or a naturalis-
tic spread of life. Cell-based generators are inspired by life, and can
be eminently effective at producing details that are similar to the
messiness of reality. Dwarf Fortress  is the undisputed king of this,
10   ◾    Procedural Generation in Game Design

incorporating an insane number of real-world physical laws (or

approximations thereof) into its gameplay and world generation.
• Scales and detail hard to do by hand : Whole galaxies, fractal details,
truly endless content. Games like Elite: Dangerous  and No Man’ s
Sky  are nearly impossible to make by hand. PCG stretches the realm
of possibility to scales beyond the conception of our limited little
minds. This sort of potential can be of primary importance for small
teams with big ambitions.
• Overcome technical limitations : Early in the history of video games,
PCG techniques were used to produce levels of content that would
be impossible to store on the media of the time. Frontier: Elite II 
had a whole galaxy of stars and planets stored on one 720 KB floppy
disk— smaller than a single graphical texture in a modern three-
dimensional (3D) game. These days, there are demo challenges to
produce interesting content, such as visualizations and music from
tiny codebases. This may seem pointless with modern computing
hardware, but could prove to have great potential in the use of wear-
able devices, integrated electronics, and other emerging hardware.

Unique
There are also softer, more subjective reasons to prefer generation:

• Individual experiences : PCG opens up the potential to give every

user a unique experience. A piece of music no one has ever heard
before, or ever will. A gameplay challenge that only they will ever
truly experience, and only in that instance. Slices of infinity that
make the experience stand out in ways static content cannot reach.
Reactive PCG can make it tailored to the user, producing a person-
alized experience. With an interactive medium like gaming, this is
immensely powerful, and makes the gameplay experience both more
memorable to the player and more worth talking about with others.
• New gameplay/interaction modes : Replayable games are just the
beginning of how PCG opens up new gameplay types. Master
Spelunky  players learn how to game the generators, predicting pat-
terns and turning PCG into a new gameplay element to be played
with. Living procedural effects, such as Conway’ s Game of Life ,
demand the user’ s attention in ways repeated patterns do not. PCG
 When and Why to Use Procedural Generation    ◾    11

involves rules, and they can be gameplay rules, as well as simple gen-
erative rules— something the player can interact and play with. This
opens up a gigantic field of new gameplay opportunities.
• Player input: Mushroom-11  uses procedural content in a truly inno-
vative way, having the player control a swarm of cells that regrow
based on simple cellular automata rules. This gives the player a new
form of gameplay through manipulation of PCG techniques. The
player must internalize the behavior of the cellular automata to con-
trol the swarm and overcome the challenges in the game.
• Unpredictable : Even the designer does not know what to expect from
the generators he or she creates. This can be scary, especially from a
QA perspective, but for a creator, it can also be thrilling. It allows you
to experience and enjoy your work in the same way as a normal end
user— something most designers never get to enjoy.
• Living system : Procedural techniques can produce living, reactive
effects. Fire, floods, lava, weather systems, populations, civilizations,
diseases, cultures, life at every level. Hand-designed content will
always appear repetitive when compared with a wild and changing
procedural system. Certain techniques, like neighbor-based decision
making, are tremendous at producing intricate and changing effects
at scale. These feel real to us, because often they are following pat-
terns similar to those that exist in the real world.
• Inhuman creativity : The outputs of PCG can be bizarre. They can
produce things no human might ever think of, from unusual level
designs to unreal animations. YouTube is awash with recordings of
buggy procedural animations in 3D games, producing horrific and
hilarious creations that astound our meat-based brains. Computers
do not think about things in the same ways we do, and sometimes
they come up with logical extremes that are wholly outside the realm
of our imagination.
• Reflections and refractions of humanity : Procedural content is almost
always based on human designed content, even if from small base
blocks. What they produce is a reflection or refraction of our own
creativity. What do we see when we stare into the abyss of Twitterbots
and novel generators? And what do these twisted shadows and
shapes have to tell us about ourselves? Combine this with a system
12   ◾    Procedural Generation in Game Design

that changes based on the player’ s input and you can develop some
thought-provoking content generation. What would a game based
on the player’ s Twitter account and followers look like?
• Inspiration of infinity : Infinity is a hard concept for us to grasp, but
playing with generators of infinite content lets us try to straddle it,
or get delightfully lost in it. A well-constrained generator will make
a dungeon level that fits tidily with your expectations, and can make
you feel the master of an infinite set of dungeons. A wild, uncon-
strained terrain generator will produce forests and fields beyond
your imagination that put you in awe of all that infinity can produce.
• Fun : Ultimately, playing with procedural generators is really quite
fun. To click a button and have a new creation brought into being. To
change a number and click the button again, and the old creation is
replaced by something quite different. To toy, to fiddle, to poke and
prod and experiment, to see what interesting and novel things come
out the other end. Dice and cards have entertained us for millennia,
but now we have computers filled with millions of dice, and we can
make them dance to our own strange tunes. Our living toys, our
endless fun, our design of designers to make more than we could
ever achieve alone. What greater than to create a creator?
 Chapter 2

Managing Output
Boredom versus Chaos

Kenny Backus
Independent 

H umans are natural pattern-finding machines, and a procedural

generation algorithm is often a pattern-creating machine. A central
conflict of procedural generation is to subvert this pattern-matching urge
without merely providing completely random output, output that will
often become its own pattern of “ oh, here’ s another thing that’ s com-
pletely and bizarrely random, lacking any interesting traits.” 
What’ s usually the aim of any generator is to mask the fact that it’ s
a generator; we don’ t think of human authors as automatons who end-
lessly create a bunch of slightly varied versions of the same thing until
they get something they like, despite the fact they sometimes do. We want
generators to make something a human would make, or more precisely,
something we didn’ t expect would be made, a spark of genuine creativ-
ity. We want to be pleasantly surprised. This “ pleasant surprise”  disarms
the pattern-matching apparatus and steers us away from the quagmires of
boredom and chaos, but it also happens to be very hard to accomplish with
any consistency. Your ability to accomplish this with your own generator
is not entirely in your control since the audience gets to decide if it’ s been
pleasantly surprised or not. It will also require you to take a long, hard
look at how your game actually  works and how the content you want to
generate lives inside it.
A common pitfall for beginners to procedural generation, even ones
who do plenty of research in procedural techniques and design, is to
13
14   ◾    Procedural Generation in Game Design

think that procedural content, a replacement for content creation, is also

a replacement for content design. There’ s faith that the system, due to the
fact that even a naive one will create unexpected results, will create pleas-
antly surprising results of high quality. Again, even a naive system will
create these infinitely, so the assumption is that there will be an infinite
procession of pleasantly surprising content.
This pitfall is born from two often unspoken unknowns: an ignorance
of what makes a game’ s design actually compelling and an ignorance of
the traits and requirements of content that will work best with that design.
Simply put, novice developers are tempted to start creating a content gen-
erator before they even know what content they want, or before they even
know what their game really  is. The siren song of infinitely varied levels,
items, and enemies is most enthralling when you don’ t actually have any
idea of what levels, items, and enemies should be in the game and you’ d
rather just have a program do it for you.
An excellent starting point for when you don’ t know what you want to
create is to figure out what you don’ t  want to create. You almost always
will need to introduce constraints  to your generator that provide guaran-
tees  for the sort of content it can or cannot create.
If you’ re randomly choosing enemies for a level at the beginning of
your game, for example, you should probably constrain the enemy set
so it won’ t contain endgame bosses or elite enemies. If you have themes
or biomes to your levels (forests, caves, deserts, etc.), you should also
constrain the enemy set to only those sorts of enemies that fit with the
environment or the intended difficulty level. Constraints are incredibly
powerful, and perhaps the first concept any procedural generation nov-
ice becomes familiar with. You’ ll also quickly become familiar with the
danger of overconstraint: say you’ ve constrained your enemy set so much
that only one type of enemy ever spawns. This is where boredom sets in,
revealing that constraints aren’ t a complete cure and that boredom and
chaos is more of a balancing act to be mediated than a conflict to be won.
The best approach to avoid this pitfall, of not actually understanding
your content before creating a generator, is to actually know in clear terms
what you want to generate before you write your generator. Create, by hand,
a set of levels, items, enemies, or smaller pieces of them, and explore what’ s
best and what’ s worst for that content, and you’ ll actually know when
your not-yet-created generator is actually making good content. In simpler
terms, make what a human content designer would create, and then try to
make your generator create similar content. Life is never black and white,
 Managing Output    ◾    15

but shades of gray, so it’ s actually best to merely be toward the “ prepared” 
end of this “ prepared– unprepared”  spectrum. There are also always excep-
tions to the rule: when you want to make a game so systematically driven
that having a clear idea of what content should live in it actually limits the
system, you should start making a generator sooner rather than later that
will actually inform the game and content design, rather than the other
way around. If you are pursuing this method wholeheartedly, you should
investigate “ novelty search,”  or algorithms that focus on creating things
that haven’ t already been created in the absence of a specific objective.
From a game design standpoint, it’ s very interesting to create a genera-
tor that then actually guides your game design (which then feeds back into
the generator, now that you know what you want), but it often means you’ ll
feel painted into a corner if you end up with a design that’ s just not the
best and leads to generated content that’ s not very pleasantly surprising.
The island generation in Sky Rogue (Figure  2.1), namely, the spawning
and placement of enemies, is an example of this iterative process and grad-
ual discovery of constraints on generated content. Sky Rogue is an action
flight simulator where most of the gameplay occurs on a series of small, pro-
cedurally generated islands or island chains. When starting a new mission,
the player is sent to one of these islands, where a multitude of air enemies are
buzzing around in packs above clusters of lucrative ground targets with sur-
rounding ground defenses. The player must destroy a specifically chosen set
of ground or air targets to proceed to the next island, although they are free
to engage or ignore the other enemy targets spawned on or above the island.

FIGURE  2.1   Image of Sky Rogue .

16   ◾    Procedural Generation in Game Design

This was an attempt to model the level-driven topology of classic rogue-

likes, a “ dungeon in the sky,”  but adapt it to a flight simulator where there
are no corridors, rooms, or walls to provide obstacles to the dungeon’ s
exit. Sky Rogue ’ s design explicitly forbade any sort of floating power-ups
or other in-flight pickups, so as not to veer too far from “ simulation”  ter-
ritory into “ arcade”  territory, and thus the usual discovery and equip-
ping of items in classic roguelikes was relegated to an interisland setup
menu. The lack of these actually made the level design incredibly difficult,
because there was such little material to work with when you couldn’ t
forcibly obstruct a player or provide the item-driven exploration and risk
taking that the genre depends on. It also meant the island terrain that
was actually generated was only of minor consequence: beyond a few hills
that might obscure your approach from ground targets when flying low,
the terrain simply has very little effect on the gameplay, something that
makes perfect sense in a flight simulator but simply comes off as boring in
an action game.
Given these challenges, the island generation relies almost entirely on
the placement of enemies and targets as its means of level design. It began
completely naively: a certain number of ground targets would be generated,
with a certain number of ground defenses a tile away, and a certain number
of flying enemies were spawned in the air (more on what “ a certain num-
ber”  really means later). It very quickly ran into the double-edged sword of
randomness: sometimes enemies were evenly spaced, and sometimes they
were all clumped together. This only created difficulties because enemies
that are spawned close to each other have overlapping weapon range radii,
which means they can attack the player simultaneously and are thus more
difficult than if they were more evenly spaced. A new constraint was cre-
ated: only spawn enemies a certain distance from each other. There were
immediate problems with air targets: fighting one of them at a time, evenly
spaced, wasn’ t as interesting as fighting squadrons, so I now spawned
“ patrol zones,”  where squadrons of airborne enemies would congregate
until the player moved close enough to be detected and attacked.
More problems were found: occasionally, the player would spawn
directly in front of the mission target, making the bare minimum require-
ment of completing an island highly variable. A simple, albeit slightly
nuanced solution: make sure they always spawn on the “ northern”  half of
the island, where the player is spawning at the very edge of the “ southern” 
half. This also happened to “ ground”  the topology of the whole island
in the player’ s mental map. Sometimes, especially in more difficult later
 Managing Output    ◾    17

islands, squadrons of enemies would be spawned practically on top of the

player, such that they’ d be thrown into a defensive battle immediately.
This broke a core rule of randomness: “ random circumstances, not ran-
dom outcomes.”  Again, a simple two-pronged solution: make sure they
don’ t spawn too close, and make them ignore the player for the first 10
seconds of the mission. On later islands, ground defenses were so numer-
ous and evenly placed that there was no “ breathing room”  in the island
design, so the island size in later levels was increased slightly and the
placement of ground targets was changed so that they would form “ sites” 
of clustered, heavily defended ground targets. This meant the objective of
having breathing room was achieved, and with negative space, the exis-
tence of positive space becomes that much more apparent, giving rise to
actual “ points of interest”  on the island. From here, a design subspace was
discovered: how to create a really interesting point of interest or set piece.
Again and again, through iterating on the island generator’ s design,
it was revealed that managing distances between enemies or, more accu-
rately, distances between weapon radii was a core pillar of Sky Rogue ’ s level
design. I would have discovered this much earlier in development if I had
just made a few prototype levels to feel out the sort of experience I wanted
the player to have at each stage of difficulty in the game.
Another case of managing boredom and chaos is a melee weapon gen-
erator that I made for an aborted action-RPG project (Figure  2.2). The
weapon generator followed an approach similar to the Borderlands  series:
given a weapon family, type, and quality level (e.g., hafted weapon, axe,
or masterwork), choose from a constrained set of pieces for that weapon
that would both look appropriate and avoid patterns enough that a new

FIGURE  2.2   Grid of swords.

18   ◾    Procedural Generation in Game Design

weapon would be visually enticing in a loot-driven game. Here, combin-

ing premade pieces to create a unique whole is an incredibly powerful way
of providing enough possibilities that the player will never find the same
weapon twice. Given a weapon made of three parts (grip, cross-guard,
and blade) and five handmade instances of each part type, the generator
creates 125 possibilities. Given the quality level, the generator then colors
each part within a constrained set: low-quality weapons look dark and
dull, while high-quality weapons look bright and vibrant. Given three col-
ors for each of three quality levels, per part, this multiplies the possibilities
even more than having a fourth part type (5  ×   9  ×   5  ×   9  ×   5  ×  9  =  91,125,
rather than 5  ×   5  ×   5  ×   5  =  625) because the handmade instances have
been slightly modified by the generator itself. Compare this number with
an average amount of weapons a player might see in a playthrough, and
you can start to reason mathematically about the possibility that the player
would see the same weapon twice.
This sort of run-time modification of premade content (a parametric
method) is also seen in Spelunky , where a handmade level chunk is slightly
modified by a few tiles, giving rise to an incredible amount of diversity, a
powerful assault in the war against pattern matching that benefits from
the control and ease of use of partial human authorship.
It’ s important to remember, though, that even combinatorially driven
procedural content can run out of steam after the player has seen hun-
dreds of combinations and discovered your premade pieces. The pattern-
matching human brain kicks in, and the illusory curtain of uniqueness
disappears: your artist-in-a-box is revealed for the pattern assembler it is
rather than the infinitely surprising creator it strives to be. Not all pos-
sibilities are created equal: in the randomly generated axe example, the
axe head provides a great deal of the weapon’ s visual identity, so having
many variations of it will give you a lot more mileage than having more
variations of handles.
On a similar note, you could have visually surprising content, but in the
context of a numbers-driven, “ crunchy”  loot-driven sort of game where
the gameplay properties of content are perhaps more important than
the visuals, your visual excellence is undermined by your design. In this
case, the player will likely discard each incredibly unique item in disgust,
because it’ s practically the same as what she’ s already found in gameplay
terms. Hopefully, this can easily be avoided in the human-authored con-
tent made while exploring the design space of your generator before its
actual creation. In this case, it’ s best to try and surprise yourself: allow
 Managing Output    ◾    19

the generator to add a diverse set of properties (fire damage, ice damage,
fire– ice damage, bonus vs. birds?) that can then be cleaned up or balanced
by your generator.
This puts a lot of pressure on the generator to actually understand
the game design, though: Is a sword that does 100 base damage really
equal to a sword that does 50 fire damage and 50 ice damage? If not, just
how unequal is it? Is fighting a dire wolf and a death knight at the same
time more difficult than fighting them one at a time, and by how much?
Quantifying properties like power, difficulty, or value in a way the genera-
tor can use accurately becomes a necessity, something that’ s usually left to
trial-and-error-driven iteration in handmade content.
Ultimately, it comes back to promising a pleasant surprise and carefully
avoiding predictable output. Knowing how to create a convincing illusion
of appropriate randomness is as challenging as generating content in the
first place, but simply knowing that before diving in avoids a lot of pain.
Keep in mind how your players see your generator, rather than seeing the
generator for what you know it is, and try to discern repetitive features
of content that you know is technically unique. From the right perspec-
tive, a million unique items can seem like the same item repeated a mil-
lion times. If patterns can form, find ways to break them, keep the player
guessing about what’ s coming next. If the player stops guessing, if they feel
like they can predict the output even if they essentially can’ t, you’ ve lost
the battle of human versus machine.
Beyond high-level advice, there are a lot of practical, reusable ways to
break up patterns and introduce a bit of chaos to your boredom:

• Avoid grid patterns when possible.  This is more difficult than it

sounds, because grids are excellent for easily organizing environ-
ments and necessary for smooth gameplay in turn-based games. For
games where the player has no need to see the underlying grid, try
and move static objects by tiny amounts to break up the pattern.
• When spawning pieces of your environment , randomly rotate static
objects to make them seem less similar to one another. For example, if
you have three types of rock, rotate them randomly, and clump them
together in irregular patterns. The player will have a very hard time
picking out patterns without closely studying the rock formations.
Compare this to having the same three types that are evenly spaced
and always presented at the same angle.
20   ◾    Procedural Generation in Game Design

• When constructing an environment out of tiles , try using nonstan-

dard shapes instead of squares. Herringbone Wang tiles are a great
example that aren’ t too hard to implement; you can find more
information at http://nothings.org/gamedev/herringbone/index.
html.
• When creating heightmaps  or other maps with Perlin or other kinds of
noise, try multiplying each pixel by several layers of other noise (also
multiplied by a scalar to adjust the other noises’  intensity), using either
a different seed or an entirely different algorithm. This is a quick way
of breaking up a lot of the patterns you might be unhappy with in
your current generator.
• Spawn characters with randomized looks.  Use props, clothing, hair,
or palette swaps, and give them slight variations in scale where pos-
sible. These can influence gameplay or be purely visual, but go a long
way to break up patterns when the player sees packs of completely
identical characters onscreen. This might not be a quick fix depend-
ing on how visually complex your characters are.
• Random circumstances, not random outcomes.  Players are generally
comfortable with, or even prefer, being thrown into an unpredict-
able environment, with an unpredictable starting point in terms of
location or equipment. These are circumstances that can lead to new
strategies to either deal with their limitations or exploit their advan-
tages: you know, feeling smart. What they don’ t like are random out-
comes: a die roll they can’ t influence that just didn’ t go in their favor,
or a randomized start point or starting equipment so influential that
it might as well be an outcome. In these cases, they lack choice, by
explicitly being given either no choice (die roll) or a false choice (bad
start point).
• Be careful with “ pure”  randomness.  There’ s not much connec-
tion between the statement “ this enemy should appear 20% of
the time”  and the enemy appearing at what actually feels like
20% of the time. Saying something should happen “ 3 times out
of 10”  brings with it a subconscious expectation that after seven
tries with no luck, it will certainly happen on the eighth try. This
specific technique of loading the dice should be in your toolkit,
but this particular example of randomness actually being random
digs deeper at an expectation of variety or fairness that players
 Managing Output    ◾    21

have. Pure randomness always seems like a good idea when the
dice are in your favor. The statement “ t here’ s a 33% chance a mini-
boss will spawn each level”  in practice is often better implemented
as “ a mini-boss will spawn every 2, 3, or 4 levels, with the highest
probability being 3.” 
 Chapter 3

Aesthetics in Procedural
Generation
Liam Welton
Failbetter Games 

T he key to creating satisfying aesthetics is the creation of a collec-

tion of guides and rules that enable you to maintain a coherent look
and feel across an entire project. This requires the development of a lan-
guage that you will use to communicate the themes, sense of place, and
story of your game to the player.
When working as a solo developer, these rules also serve as a guiding
principle that will help ensure the work you are producing today matches
the work of the day before, or the days to come. The aesthetic guides you
put in place set parameters within which you know you can safely experi-
ment and improvise without creating something completely inappropriate
for your game.
When you are working in a team, it becomes even more important to
have a visual and aural design that everybody knows they are working
toward. This makes sure that everyone’ s individual contribution can add
together to form a coherent whole.
When establishing the style for your game, and translating it into a
set of design rules, it is also important to bear in mind the strengths and
weaknesses of your team. You might find someone’ s technical execution
lends itself to a slightly different aesthetic approach than that originally
envisaged. Factor in that a team member’ s skill set may lead them to
interpret your rules and guides in a way you did not predict, if you are
insufficiently precise. Establishing good aesthetic rules involves making
23
24   ◾    Procedural Generation in Game Design

allowances for, as well as even embracing, the creative differences that are
inevitable when working with a team.
In this respect, it can be useful to imagine your procedural generation
engine as a member of your team that works alongside your other art-
ists and designers, taking into consideration its own strengths and weak-
nesses. The engine is going to be contributing work based on rules that
you have established for it. It’ ll even be granted some creative freedom,
although admittedly this is usually simulated by introducing a degree of
randomness. The procedural generation engine is a team member that can
follow hard-and-fast rules to the letter, but can fall short when it comes to
making judgments in less strictly defined areas.
It is possible to work around this, as long as you identify the strengths
and weaknesses of your procedural generation engine in advance of using
it, and plan your aesthetic design accordingly. Work out where it will need
help and grant more control to your other team members there.

ATMOSPHERE IN SUNLESS SEA 

When we began development on Sunless Sea , it was clear from the outset
that atmosphere and exploration were going to be every bit as important
to the core experience of the game as survival and danger. Sunless Sea 
is a game about setting off from the familiar and traveling out into the
unknown. We wanted the world to provide both a challenge and a sense of
threat, and didn’ t want these to be lost as the player became familiar with
the setting. To accomplish this, we decided the map should vary each time
a new game was started. This would ensure each new journey was different
and the sea it took place on never felt like an entirely safe place.
But it needed to be balanced. A sense of challenge and freshness would
only add to the atmosphere and thrill of exploration up to a point. We
didn’ t want the player to feel they gained no skill or understanding despite
spending significant time in the game, or for journeys to feel entirely arbi-
trary. To create the sense of a long, strange journey through the darkness,
we needed to induce the feeling of gradually passing from the familiar into
the alien. This meant that the world of Sunless Sea  had to be a coherent
one— while every discovery should feel exciting and fresh, it should also
feel natural to the player in hindsight. Discoveries needed to build on the
expectations players would have developed from exploring areas adjacent
to them. Our greatest fear when designing our procedural generation pro-
cess was that we would expose the players to tonally jarring contrasts that
would throw them out of the world, rather than draw them deeper in.
 Aesthetics in Procedural Generation    ◾    25

ESTABLISHING THE RULES

We decided on a grid-and-tile-based approach toward map generation. The
advantage of this approach was that the rules for tile placement could be
simple enough for everyone working on the project to understand, not just
the designers and coders. As originally envisaged, the system would involve
laying out an 18  ×  18 grid of tiles, each of which would contain content (a
creature, an island, a geographical feature, etc.). The design rules would
determine the type of tile that could be placed to the north and east of each
tile. Tiles were organized into categories, and each category had a particular
style and difficulty level. Blue tiles were placed to the north of the map and
featured icebergs and snowstorms, while purple tiles would generally be to
the far east of the map and feature our most alien flora and fauna.
While working within this system, we found occasional aesthetic dis-
sonance between the tiles. To avoid this, we began creating ever more
specific subcategories, but this reduced the sense of variety and random-
ness we were achieving. We spent a long time working out the best way to
make the moment-to-moment play experience feel harmonious, while still
allowing for variation on a larger scale when crossing the map.
We found ourselves naturally designing areas in 3  ×   3 chunks, as this
provided enough space to create a sense of a coherent location. You would
have a few islands, some aquatic features, a sea beast or two, and then
you would move on to the next area. We therefore decided to reduce our
grid size from 18  ×   18 to 6  ×   6 (Figure  3.1). We retained the region-sorted
tiles, ensuring tiles were placed next to those with a complementary theme
and design, but the decrease in the number of tiles meant that we could be
a lot more lenient with their placement. Finally, to make sure that areas of

FIGURE  3.1   Before (18  ×   18) and after (6  ×   6).
26   ◾    Procedural Generation in Game Design

the map that would conflict aesthetically could never be placed next to one
another, we made sure there were neutral, bridging regions between them
so that they would never adjoin.
Eventually, we wound up with a grid of 36 tiles, 24 of which were split
into five regions and 12 of which were located in a fixed position. Each
tile could be placed anywhere within its corresponding region of the
map, allowing for much greater variety than the earlier ruleset allowed.
Restricting tile type to only five regions, rather than subdividing them
further by specific areas, also meant our procedural generation code had a
greater likelihood of creating exciting combinations.
We also established a series of unifying rules around the use of color,
lighting, sound, and music. These are essential for any game, but we found
them particularly helpful. By adhering to stylistic rules, we could avoid
maps being generated with clashes between the atmosphere generated
by the sound and by the visuals. This, combined with the fact that the
game, by design, has long stretches of open water between inhabited areas,
meant that we were able to avoid some of the uglier issues that can arise
when using procedural generation for environment design.

BLURRING THE BOUNDARIES

The tile rules meant that areas wouldn’ t align discordantly. However, tiles
from different region categories would still be placed next to each other.
Easing the transition between tiles helped to smooth over the joins. When
the player travels from tile to tile, we determine their intention— whether
it’ s to continue into the tile or to move back into the previous tile (zigzag-
ging back and forth between tile boundaries). If they do continue into the
tile, we will slowly alter the color of the sea beneath them, allowing us to
change the game’ s atmosphere without introducing a jarring transition.
A similar technique is used with the background music to determine if
it is appropriate to introduce a new musical theme. The absence of music
is an effective way of building and maintaining atmosphere. By only intro-
ducing underscoring when it seemed absolutely appropriate, we avoided
the jarring feeling of tracks shunting into one another as you crossed the
map and changed regions.
Finally, we allowed the procedural generator to create weather effects
that ignored tile boundaries. This helped to prevent each tile feeling like
an isolated little rock pool— it could snow on a tile that wasn’ t strictly a
“ snowbound”  one, although this would occur rarely and only happened
when such a tile was adjacent to one of the northern ones.
 Aesthetics in Procedural Generation    ◾    27

THE RESULT
In the end, the tile-based approach we took to procedural generation in
Sunless Sea  was successful. It enabled enough variation to keep the chal-
lenge fresh for players, while allowing us to retain sufficient aesthetic
control to craft the experience we wanted. However, the approach wasn’ t
without drawbacks. Treating each tile as a puzzle piece that could be slot-
ted into various locations meant that more time had to be taken designing
them to fit in every arrangement. We made this easier to accomplish by
providing our procedural generation engine with pieces that were visually
and aurally coherent. If you were to put each asset in a line, you would see
evolution and subtle variation, but they would all be of a piece. The color
palette is restrained and harmonious, contrast levels are consistent, and
shapes and sizes are fairly uniform.
One decision we regret was the attempt to introduce more variety by
creating handcrafted variants of each of these individual tiles and ran-
domly assigning one per play session. This was time-consuming, prone to
human error, and very difficult to debug, which taught us the valuable les-
son that sometimes it is better to accept the limitations of your approach
to procedural generation than to patch over them with manual solutions.
For Failbetter Games’ s expansion to Sunless Sea , Zubmariner , we used a
solution that attempts to accomplish a similar effect, but through proce-
dural logic rather than manual placement.
In Zubmariner , players are given the opportunity to explore underwa-
ter in a submarine. This involves creating an entire subsurface level to the
play area— and would be a daunting task if attempted by hand. Instead,
we are using the rules already established for our surface tiles and hand-
ing over even more control to our procedural generation engine. It picks
from a predefined list of atmospherically appropriate pieces of terrain and
decorations and places them in unoccupied areas of the map at a density
our artist can adjust.
In this way, we can hand over the task of decorating large areas of the
map to the procedural generation engine. We have taken the time to estab-
lish the rules it applies and ensured that all the assets it is using are well
crafted and visually coherent. Now we get to sit back and enjoy the work
that our silent, stalwart team member produces.
 Chapter 4

Designing for Modularity

Jason Grinblat
Freehold Games 

MODULES AND GESTALTS

Modularity is the use of discrete units, called modules , to assemble larger
structures, which we’ ll call gestalts . Our motivation for investigating
modular design comes from the combinatorial magic that modularity
conjures. We design a few modules, along with an assembly mechanism,
and we inherit a plethora of gestalts for free. And gestalts are what mat-
ter. They’ re the artifacts our players encounter and care about, be they
dynamic puzzles, dungeon levels, or dialog trees. Modularity multiplies
our work, often bearing novel results in the process.
Modular design has an extensive presence across the genealogy of game
systems. Consider one canonical example— equipment in an RPG. Here,
the modules are the individual pieces of armor. The assembly mechanism
is the player’ s choice of equipment, within the constraints prescribed by
each piece’ s coupling to a particular body part. The gestalt is the entire
equipment loadout. When a fortunate player has amassed more gear than
they can wear and must decide how to outfit themselves, it’ s modular-
ity that gives the system its play. I’ m using a particular definition of play 
here, namely, “ the free space of movement within a more rigid structure.” 1 
When we talk about play in this chapter, we mean the range of interesting
possibilities prescribed by our system’ s rules.

29
30   ◾    Procedural Generation in Game Design

In the context of designing systems that leverage procedural genera-

tion, we limit our consideration to modular designs that satisfy these two
qualities:

1. The assembly mechanism includes some degree of randomness.

2. The gestalt space is too large to craft every possible gestalt by hand.

The aim of the modular designer, then, is twofold. Craft modules that
assemble into desirable gestalts, and design an assembly mechanism to
manifest them. The gestalts themselves are out of our direct reach, and
that’ s why it’ s a challenging problem. Of course, what counts as desirable
depends on the aesthetics of the designer. But we assume the designs that
give our system the most play are the favorable ones, especially those that
contribute to our thematic goals for the system.
This chapter sets out to develop a model for the modular paradigm. We
introduce language to help conceptualize the pieces at play, and we famil-
iarize ourselves with the underlying math. Then we suggest an approach
to designing for modularity. We examine material systems as examples,
but we don’ t get into the idiosyncrasies of specific applications, such as
level generation or dynamic puzzles. We leave those challenges to later
chapters.

ASSEMBLY MECHANISMS AND GESTALT SPACES

A gestalt space is the set of all possible gestalts that can be assembled from
a set of modules, by a given assembly mechanism. When designing modu-
lar systems, it’ s useful to think about the sizes and shapes of their gestalt
spaces. In fact, we can restate the aim of the modular designer in terms of
gestalt spaces: we want to craft modules and an assembly mechanism that
yield a sufficiently large gestalt space with a high proportion of desirable
gestalts.
There’ s no generic metric to assess gestalt desirability. We have to
decide what a desirable gestalt looks like for each of our modular sys-
tems. For equipment in an RPG, it might look like a set of equipment
with synergistic bonuses. Desirability is often graded on a spectrum,
with ideal gestalts on one end and unworkable gestalts on the other.
When a gestalt crosses a certain threshold of undesirability, we say it
enters the null zone  of our gestalt space. This is a useful concept that
we’ l l revisit later.
 Designing for Modularity    ◾    31

Modules Assembly Gestalt

mechanism

FIGURE 4.1   Monsters (modules) are chosen by the dungeon populator (assem-
bly mechanism) to populate a dungeon room (gestalt).

Most assembly mechanisms produce gestalt spaces within a certain size

class. Let’ s look at an example to illustrate this point. Say we’ re populat-
ing a dungeon room with three monsters from a set of five monster types.
Each monster type can be chosen multiple times. Our population algo-
rithm is simple: choose a monster at random from the set and add it to the
room, and then do it again, and then again. For simplicity’ s sake, assume
that the monsters’  positions in the room don’ t matter. In this example, the
monster types are the modules, the population algorithm is the assembly
mechanism, and the dungeon room is the gestalt (Figure  4.1).
How many ways can the dungeon populator choose monsters to populate
the room? The answer is 5  ×   5  ×  5  =  125. Note that this is a different question
than how many gestalts there are. That’ s because in this scenario, the order
that the dungeon populator chooses monsters doesn’ t matter. Whether the
algorithm picks an orc or a goblin first, we don’ t care, as long as they both
end up in the room. A gestalt that cares about order is called a permutation ,
and one that doesn’ t care is called a combination .* Out of the 125 ways the
assembly mechanism can populate the room, there are 35 distinct gestalts.†
Without getting into the nuances of the math, we can note how quickly
the size of the gestalt space grows as we increase the number of monster
types or the number of monsters we populate a room with. Table  4.1 shows
some examples.

* Technically, since we’ re allowing each monster type to be chosen multiple times, it’ s permutation
with repetition  and combination with repetition .
† The number of r  combinations of a set of size n  with repetition allowed is [(r   +  n   –   1)!]/[r !(n   –   1)!].
32   ◾    Procedural Generation in Game Design

TABLE  4.1   Populating a Dungeon Room with Monsters Chosen at Random

Number of Monster Types to Number of Monsters Chosen Number of Possible Dungeon
Choose From (Modules)  (Assembly Mechanism)  Room Populations (Gestalts) 
5 3 35
10 3 220
25 3 2,925
25 5 118,755
25 10 131,128,140
Note:  The size of the gestalt space grows very quickly when we increase the values of the
parameters.

This explosion is the combinatorial magic we mentioned earlier. The

precise numbers change as we vary our assembly mechanism, and varia-
tions are limitless. They include algorithms that, for example, care about
the order modules are chosen, choose modules from a number of different
sets, or weight the choice of future modules based on modules chosen pre-
viously. But in nearly all cases, even a modest number of modules yields a
big gestalt space.
Now that we have a feel for the size of this gestalt space, let’ s examine
how certain design decisions can affect swaths of the space. Our goal is to
understand how interactions between our modules propagate through our
gestalt space. In the dungeon room example, let’ s contrive two monsters:
a fire imp and a water elemental. Whenever a fire imp spawns in the same
room as a water elemental, its attack instantly annihilates the elemental.
This is undesirable behavior, so let’ s say any dungeon room containing
both a fire imp and a water elemental is undesirable.
As it turns out, 5 of our 35 dungeon rooms contain both monsters.*
Therefore, our decision to include both the fire imp and water elemental
among our monster types confines 5 of our 35 gestalts to the null zone. At
this point, we have a few choices. We can live with this proportion, or we
can add a constraint that ensures the two monsters don’ t spawn together,
or we can revisit our monster design. Whatever we decide, the mathemati-
cal context we get from examining our gestalt space enables us to make
better design decisions and work toward enabling play.

* Lettering the monsters a , b , c , d , and e , the following combinations comprise all the gestalts that
include two arbitrarily chosen monsters, a  and b : aab , abb , abc , abd , and abe .
 Designing for Modularity    ◾    33

ENABLING PLAY
What makes a set of modules and an assembly mechanism conducive to
producing desirable gestalts? As mentioned earlier, our goal is to give our
system as much play as possible. Again, by giving our system play, we mean
enabling players to explore the greatest amount of the compelling space
created by the rules of our system, within some thematic constraints. Let’ s
examine three principles of modular design that work toward enabling play.
We’ ll use examples from Sproggiwood , a roguelike game I designed and
released in 2014 with Freehold Games, to illustrate the principles in practice.
Mechanics as Shared Substrates
If certain combinations of modules assemble into desirable gestalts, then
it stands to reason that there’ s some definable relationship between mod-
ules that assemble desirably. Consider this straightforward example of
such a relationship: two pieces of equipment that each grant some bonus
when the other is worn. This design pattern occurs regularly in RPGs in
the form of set pieces. From the assembly of these two modules, we get a
desirable gestalt, in this case the equipment loadout and its corresponding
set bonus. But this relationship tightly couples the two modules, and in
doing so, it prevents either of them from assembling meaningfully with
other modules. While we gain a desirable gestalt, we lose a big swath of
our gestalt space to the null zone. For this reason, I contend that direct
relationships like this one are detrimental to the goal of enabling play.
Accordingly, I dislike the design of set pieces.
Instead, we can relate modules indirectly by using the language of our
game’ s mechanics to define them. Each mechanic acts as a substrate for
our modules to act in. For example, consider the following two pieces of
equipment for an RPG. The first piece is a lance that deals bonus damage
equal to your movement speed. The second piece is a pair of boots that
increases your movement speed by 20%. We can see how the assembly
of these two modules produces a desirable gestalt. But they aren’ t tightly
coupled, and so we haven’ t confined other potential gestalts to the null
zone. Each module can be combined with other modules that are defined
in the language of the game’ s movement mechanics to produce more
desirable gestalts. By avoiding the prescriptive design of set pieces, we give
our players more options for meaningful combinations.
The key to this approach is choosing a suitable suite of mechanics to
act as substrates. The more central a mechanic is to our gameplay, the
more play we can enable by designing modules in the language of that
34   ◾    Procedural Generation in Game Design

FIGURE  4.2   Typical dungeon layout in Sproggiwood .

mechanic. For Sproggiwood , turn-based movement is a central mechanic.

Dungeons are laid out on a grid and populated procedurally with mon-
sters (Figure  4.2). On each turn, the player can either move one tile in any
of the four cardinal directions, attack an adjacent monster, use a power, or
use an item. After the player acts, each monster acts.
Since movement is so crucial to Sproggiwood ’ s gameplay, we designed
several traps and monsters whose powers constrain or coerce movement
on the grid. For this system, the traps and monsters are the modules, the
dungeon populator is the assembly mechanism, and dungeon rooms are
the gestalts. A dungeon room that’ s tactically challenging is a desirable
gestalt. Since so many of the modules are designed on top of the shared
substrate of movement, the number of dungeon rooms that present novel
tactical challenges is enormous. Figure  4.3 shows an example.
In this scenario, the player recently killed a black jelly monster, which left
behind a puddle of jelly (1). Jelly puddles must be walked over within seven
turns, or they spawn another jelly monster. The frog (2) uses its tongue
to attack every few turns by pulling the nearest creature to a tile adjacent
to the frog. Next to the player, the lily pad (3) provides an escape route; it
teleports anything that steps on it to another lily pad in the dungeon.
Players have an interesting choice here. They can move toward the jelly
puddle and risk getting pulled to the frog, or they can escape via the lily pad.
Or, if they possess a power that also acts on the movement substrate— say, by
letting them move multiple tiles in one turn— they can get to the jelly puddle
immediately. By designing our modules in the language of grid movement, a
 Designing for Modularity    ◾    35

1 2

FIGURE  4.3   The dungeon populator (assembly mechanism) combined three

traps and monsters (modules) into a tactically challenging dungeon room (gestalt).

core mechanic in Sproggiwood , we enabled the assembly of colorful gestalts

like this one without nullifying a big swath of our gestalt space.

Orthogonality
In the context of a modular design and a set of game mechanics, orthogonal-
ity refers to the partitioning of the design into modules that don’ t overlap and
whose gestalts span the space of gameplay prescribed by the mechanics. It’ s
easy to see why that second part— spanning the space of gameplay— works to
enable play. The first part is a little trickier, though, so let’ s examine it.
When the majority of the designs for two modules overlap, the gestalts
that include either module have the potential to counterfeit each other,
meaning they replicate each other’ s place in the gestalt space and effec-
tively reduce the number of distinct, desirable gestalts. Let’ s revisit the
example dungeon room with the black jelly puddle, frog, and lily pad.
Remember that the jelly puddle spawns another jelly unless the player
walks over it within seven turns. Imagine we had designed another type
of jelly monster that leaves behind a similar puddle— say, a purple one
that spawns two jellies unless walked over within nine turns. The example
dungeon room and the dungeon room that includes the purple jelly pud-
dle counterfeit each other. They present challenges that are virtually iden-
tical. Moreover, these designs compromise each other’ s thematic identity,
muting the resonance each has with the player.
36   ◾    Procedural Generation in Game Design

FIGURE  4.4   Sproggiwood ’ s assortment of jellies in various stages of their bounce

animations.

In Sproggiwood , we designed six different jelly monsters that leave

behind different puddles or otherwise affect the grid when they die
(Figure  4.4). We tried to give each a distinct identity and considered how
the gestalts they assembled into might counterfeit each other.

• Yellow jelly—leaves behind a puddle that causes walkers to slip one tile
• Blue jelly—leaves behind a puddle that causes walkers to slip one tile,
and deals poison damage
• Red jelly—explodes in a cross pattern one turn after it dies and dam-
ages anything it hits
• Black jelly—leaves behind a puddle that causes walkers to slip, and
must be walked over within seven turns or spawns another jelly
• Purple jelly—leaves behind a puddle that causes walkers to slip, and
shoots a lightning bolt that triggers other purple puddles
• Ice jelly—explodes in a cross pattern one turn after it dies and freezes
anything it hits

Note that we didn’ t achieve total orthogonality. The yellow jelly’ s pud-
dle is subsumed by all the other puddles, and the red and ice jellies present
very similar challenges. It’ s difficult to get all the way there, both because
some overlap is inevitable when designing thematic monsters and because
spanning the space of gameplay, or even articulating the whole space, can
be challenging. However, it’ s useful to let orthogonality play a role in your
design decisions.
 Designing for Modularity    ◾    37

Equivalence of Impact
An overshadowing module is one whose impact is so powerful that it dis-
torts the gestalts it gets assembled into. Similar to how overlapping mod-
ule designs can counterfeit gestalts, overshadowing modules can flatten
the dimensionality of gestalts by compelling so much action that other
modules can’ t compete. Think of the negligible effect other massive bodies
on Earth have on the planet’ s gravitational pull. No matter what the bod-
ies are or how they’ re positioned, they’ re dwarfed by Earth’ s monumental
force.
It’ s unrealistic and undesirable to expect all our modules to converge
to a single impact value, but it’ s a reasonable goal for most  of our modules
to fall within a certain impact range. We can adhere to this principle by
designing modules with roughly equivalent impact values and by ensur-
ing our assembly mechanism rarely combines modules whose impact val-
ues are wildly disparate. Note that it’ s not enough for the average  impact
values of our gestalts to be similar if the impact values of our modules are
distributed unevenly. This is because an outlier module can render the
gestalt undesirable even if the gestalt has a suitable average value.
For an example of an overshadowing module, let’ s look at how out-
of-depth encounters work in Sproggiwood . The dungeon populator in
Sproggiwood  has a chance to produce an out-of-depth encounter, where
a difficult monster from a later dungeon appears in an earlier one. These
encounters add spice to the dungeon population mix, but they do so at
the potential cost of overshadowing the other modules in their gestalts.
The easiest way to illustrate this point is to imagine entering a dungeon
room at one-fourth your total hit points and seeing an out-of-depth mon-
ster. No matter how interesting a tactical situation the other challenges
in the room might present, the presence of the big baddie compels us to
flee (Figure  4.5). This is a principle we understand from other spheres of
design insight; an unbalanced encounter can be unfun. But it’ s useful to
consider it in terms of the effect it has on our gestalt space.
In Sproggiwood , far out-of-depth encounters are rare. We engineered
the dungeon populator to produce out-of-depth monsters at a clip that
adds spice without compromising too much gameplay. Another option
is to tweak your assembly mechanism to pair out-of-depth encoun-
ters with out-of-depth rewards. For example, a difficult monster might
guard a powerful sword. This approach still runs the risk of flattening the
38   ◾    Procedural Generation in Game Design

FIGURE  4.5   A yeti looms over the battlefield. No matter how interesting the rest
of the encounter is, we’ re forced to flee.

dimensionality of the gestalt, but it also adds a new dimension that could
be worth the compromise in the right situations.

PLOTTING DESIRABLE GESTALTS

Now we have some principles to guide us in the design of our modules. But
how do we translate an initial gameplay idea into a modular design that
yields the desired gameplay? After all, gameplay in modular, procedural
systems is emergent. It’ s operating on the level of the gestalts, whereas
we—as designers— are operating on the level of modules and assembly
mechanisms. We stated previously that the gestalts are out of our direct
reach, that we can’ t hope to articulate the whole gestalt space of our design.
But we can envision individual gestalts that typify our ideal gameplay and
use them to inspire our design. This powerful approach acts as a bridge
between our vision and the parameters that enable it.
Look to envision gestalts across the spectrum of your rule systems,
from common interactions to rare ones. Think about a story one of your
players might excitedly tell their friends about how they took advantage
of a rare confluence of events to narrowly escape a deadly situation. Now
think about a more common occurrence, say, a particular strategy a savvy
player might repeat over and over again for some minor benefit. Once we
have a number of desirable gestalts, we can interpolate between them to
develop modules and an appropriate assembly mechanism.
We used this approach to design the faction system for our roguelike/RPG
Caves of Qud , released in Steam Early Access in 2015. Caves of Qud is set in
 Designing for Modularity    ◾    39

the far future, long after an apocalypse of unknown origin, where mutations
have uplifted certain animals to human-level intelligence and beyond. In our
original design, players would find animals from several different species in
the same zone. Most of the animals would act aggressively toward the player
and ignore each other. We wanted to add interspecies conflict to give the
world more thematic texture and dynamism. Specifically, I imagined a sce-
nario where a player stumbles on a fight between a troop of baboons and a
herd of goats, two groups that often inhabit the same zones in Caves of Qud .
I also imagined that occasionally these groups would be fighting for a ridicu-
lous reason; I wanted to match the tone of the game’ s frequent forays into the
absurd (Figure  4.6). I thought, “ What if a named baboon leader was hated by
goats for some esoteric reason that no baboon or goat would normally care
about? And what if that dispute triggered a battle?” 
This imagined gestalt led to the design of a faction system where animal
groups’ behavior patterns get modified by the procedural histories of their
faction leaders. When a faction leader is generated, they’ re given a small
backstory with one to three procedurally chosen factions. That backstory
determines whether the factions like or dislike the leader, which in turn gov-
erns an animal’ s behavior when it’ s generated in the same zone as the leader.
The system ultimately expanded on this idea and yielded many more
interactions than I initially conceived. Animal factions find strange allies
in each other, and players can share water with faction leaders to alter their
reputations with the factions that like or dislike the leader. The exercise of
plotting desirable gestalts gave me a kernel to design around and pushed
the design into unexpected territory.

FIGURE  4.6   A baboon leader angers goatfolk with its radical ideas about
mathematics.
40   ◾    Procedural Generation in Game Design

INSERTING MEMORABLE ASYMMETRY

In closing, let’ s examine a pitfall inherent to the paradigm outlined in
this chapter. The approach of neatly partitioning systems into modules
that enable play is a powerful one. Designs that use mechanics as shared
substrates and adhere to the principles of orthogonality and equivalence
of impact tend to form elegant, symmetric systems. But games are more
than elegant systems, and the beauty of the design shouldn’ t be privileged
over the experience of the player. If we lose our way in the weeds of sys-
temization, we can end up with a system that loses touch with its thematic
inspirations. Make sure to leave room for thematic overrides that violate
the principles of enabling play where appropriate. Asymmetry can disrupt
the balance of elegantly designed systems, but in the right doses, it can do
so in memorable and favorable ways.
We ran into this pitfall in the design for Sproggiwood ’ s weapons. I had
decided early on that I wanted a concise suite of weapons for the player to
find as loot. Each weapon would offer a different tactical tool and would
act on the substrates of grid movement and monster health. I created three
classes of weapons that implemented three nearly orthogonal ideas: freez-
ing, flaming, and vampiric. Freezing weapons freeze enemies for a single
turn. Flaming weapons deal bonus fire damage. Vampiric weapons drain
monsters’  life and give it to the player. Within each class, I designed a
few tiers that elaborated on each theme by producing its effect in different
shapes on the grid or dealing various amounts of damage. I intentionally
decoupled the weapon classes from the six player classes so that any player
class could find a freezing, flaming, or vampiric weapon.
From the perspective of enabling play, this precise systemization and
decoupling was very effective. When a warrior equipped a freezing sword,
the combination of the sword’ s attack and the warrior’ s powers produced
a desirable gestalt. The same thing happened when a wizard equipped
a freezing staff. Furthermore, the freezing warrior and freezing wizard
gestalts played very differently and maintained distinct personalities. But
in the roguelike genre, looted items—and weapons in particular—have a
tradition of juicing up gameplay. Our overwrought systemization caused
some of our players to note that our weapons felt “ samey”  and lifeless
(Figure  4.7).
To fix this problem in our 2015 mobile release of Sproggiwood , we
inserted a dash of asymmetry into our weapon design. In addition to the
standard weapon classes, we created unique weapons that were coupled
 Designing for Modularity    ◾    41

Shovel of Sword of Staff of

biting frost biting frost biting frost

FIGURE  4.7   Three identical weapons for three player classes. The repetition
diluted the thematic impact of the weapons and made them feel boring.

to the player classes. For example, the farmer class got a grappling pitch-
fork that could be shot at walls to slide the player along the grid toward
the wall. Even though we lost out on desirable gestalts by confining that
effect to the farmer class, the benefit of an exciting, thematically resonant
weapon made up for it.
Ultimately, your adherence to the principles of enabling play should be
in service to your vision for your game. When the principles run afoul of
that vision, don’ t be afraid to subvert them.

REFERENCE 
Salen, K., and Zimmerman, E. 2004. Rules of Play: Game Design
1.
Fundamentals. Cambridge, MA: MIT Press.
 Chapter 5

Ethical Procedural
Generation
Dr. Michael Cook
Independent 

T here are a whole bunch of different things to worry about when

you make a procedural generator. Will my generator ever make a mis-
take? Does it produce a lot of boring things? How much time does it need
to generate something? Sometimes, it can be so tricky to get your gen-
erator to work that simply getting it to produce anything feels like a huge
accomplishment. Worrying about what it’ s producing, why it’ s producing
it, or what others might think about it can get lost in the huge relief and
satisfaction of finally seeing little dungeons, poems, or stories coming out
of your computer.
As generative software grows in popularity, we’ re seeing its ideas used
by more people in lots of novel ways, and for lots of exciting new uses.
These developments are encouraging and inspiring, but it also makes it
even more important that we take the time to think about the broader
impact of this technology on people and society. In this chapter, we look
at the ethical issues that can arise when making procedural generators,
and give you some things to keep in mind when you’ re making your next
generator. They won’ t all apply to every person and every project, and you
may not agree with every single thing we suggest, but hopefully just read-
ing through this chapter will give you a new perspective on procedural
generation, and help you think about this awesome art form in a different
way. Along the way, we’ re going to travel to some pretty extreme parts of
the generative world, and pose some open-ended questions about what
43
44   ◾    Procedural Generation in Game Design

we should let generators do, how we should talk about them, and how we
should let them talk about the world.

TALKING IN CODE
A procedural generator encodes a lot of different ideas and knowledge
into a tiny package of rules and procedures. Each time we run it, those
rules and procedures unpack bits and pieces of the knowledge we fed in,
and reassemble them into something new. Some of the ideas we feed into
our generator are technical, and a lot of this book is given over to the
broad and complicated technical challenges that surround procedural
generation. We’ ve already seen some of this in earlier chapters about
modular design and balancing chaos with predictability. Other ideas are
aesthetic or artistic in nature. Procedural generators encode our ideas
about how we think games should be designed: they represent what we
think a good dungeon looks like, how a good story should end, or what
makes a beautiful color scheme. They’ re little digital apprentices that we
train and then trust to finish bits of our artwork in the homes of the
people who view them.
Creating a procedural generator is a bit like making a work of art— some
of the things we put on the canvas are intentional, but other things we
paint on might be unconscious acts we don’ t give much thought to. Some
might simply be mistakes. When we tweak the maximum size parameter
of our dungeon generator, we most likely know the effects it will have on
our level difficulty, or how long our game takes to play— that’ s an example
of a conscious decision. But there are lots of ways in which we can intro-
duce accidental features to our generators, sometimes without even think-
ing. The tricky thing about this is that it can be very difficult to realize
this has happened, because generators are often quite complicated, and
are designed to generate huge quantities of content. If there’ s something
wrong in the system, it might be hard to tell without looking at thousands
and thousands of examples.
Let’ s explore an example of a system that ends up with an unexpected
feature in it. Suppose we’ re making an RPG level set in a spooky grave-
yard, and we want the player to be able to read randomized inscriptions
on each tombstone. We decide to do this using a few simple patterns, so
we write a few templates for gravestone patterns, like “ #NAME lies here,
died #YEAR.”  Our templating system knows that when it sees #YEAR, it
replaces it with a random four-digit number, and when it sees #NAME, it
generates a random name by sticking a first name and a last name together.
 Ethical Procedural Generation    ◾    45

We give it a try and it works great— random names appear with random

years attached. We’ re getting bored of the templates we’ ve written, though,
so we add one more, an inscription for couples buried together. It reads:
“ #NAME and #NAME, reunited in love once again, #YEAR.”  As a quick
exercise, you can imagine a few possible outputs for this template in your
head, or maybe write them down.
This template is a bit more interesting than the ones with only one
name. Take a moment and imagine how we might implement the #NAME
generator. You don’ t need to think about it in terms of code— imagine how
you might do it with pen and paper. Depending on how we think about
relationships, gender, and sexuality, our first instinct might be to write two
lists of names: one list of names that sound like they might be associated
with male identity, and another list of names that sound like they might be
associated with female identity. Another approach is to simply have one big
list of names and not worry too much about separating it out into a binary
of any kind. You might not even think about this distinction— maybe you
just used one list because it was quicker and you wanted to save time. But
this decision, whatever you do, is going to have an impact on what this
graveyard says about sexuality and gender in your game.
To see how, let’ s continue through our example. Suppose we use our
two separated name generators for our new couples’  template, and gener-
ate a gravestone or two. The individual examples will make sense. “ Jack
and Jane, reunited in love once again, 2016.”  They’ ll look fine, and you
might even read a dozen or so while walking around. But “ Jack and John” 
or “ Jane and Lucy”  will never appear— name pairings that might suggest
nonheterosexual relationships. Because of the way we’ ve structured the
data in our generator, our graveyard is now presenting our game world in
a way that implies that only heterosexual relationships exist.
Of course, this might be exactly what you want for your game! Perhaps
you want to show the player that only heteronormative couples are afforded
a proper burial in this town, and it becomes an important part of the plot.
What matters here isn’ t the message sent by the generator (although for
what it’ s worth, we highly recommend the single namelist approach).
What matters is whether you as a creator realize what message your gen-
erator is broadcasting, because that means that you’ re in charge and able
to assess whether that message is something you’ re happy with sending
out into the world. Procedural generators can be powerful tools— they can
amplify our ideas thousands of times for every single person who plays
our games. But it’ s important to know what it’ s amplifying, and our little
46   ◾    Procedural Generation in Game Design

gravestone example is a case where we might be sending a message we

didn’ t expect to.
How can we stop these undesired features from appearing in our gen-
erators, or find them when they appear? For this specific kind of problem,
where the structure of our data or code is to blame, a good rule of thumb
is to simply never introduce distinctions that don’ t mean anything in your
game. If gender doesn’ t affect your player or your game mechanics, there’ s
no reason to build it into the systems of your game. The more variables
you introduce, the more likely there will be something you didn’ t plan
for, so if something doesn’ t need to be defined in code, it’ s safer to simply
sidestep it.
In general, though, there isn’ t a hard-and-fast rule from preventing
these things from happening. Video games are big, complicated pieces of
art and technology all mixed together, and part of the fun and excite-
ment of making games is seeing things you didn’ t predict. The best thing
you can do is always be thinking about what you make, always try to be
critical of what you do, and be prepared to fix mistakes when they happen.
Everyone working in procedural generation is learning something new
every day, and no one should feel bad for not seeing something coming.
Each surprise is another learning experience.

THE BIG WIDE WORLD

A lot of the time our generators don’ t need to know much about the real
world in order to do their job. Spelunky ’ s level generator doesn’ t need
to know anything about how caves are formed, who Indiana Jones is, or
where a dog is most likely to hide if it gets stuck underground. All it needs
is a big pile of level chunks, and the rules for how to stick them together
to make levels.
Sometimes we want our generators to have a little bit more information
and knowledge in them. Maybe we want our recipe generator to name itself
after random places from the real world, or maybe we want to look up the
symbolic meanings of colors for our flag generator. Whether it’ s knowl-
edge we simply don’ t have ourselves, or it’ s faster and more flexible to do
it automatically, procedural generators can easily be fed information from
the Internet or other sources to make them more clever and more powerful.
One way of doing this is to use static databases of knowledge, like
Wikipedia, for information. Argument Champion , a 2012 game about
debating, used an online resource called ConceptNet to generate content
for its game. ConceptNet is a database of facts and relationships designed
 Ethical Procedural Generation    ◾    47

for use by software, especially artificial intelligence (AI) programs. Some

of its knowledge is input by hand, while others are automatically scraped
from other sites and formatted for ease of use.
The results can seem fairly impressive, especially at first. Type in “ cat,” 
and ConceptNet can tell you that cats are capable of hunting mice, that
they have four legs and whiskers, and that they like playing and drink-
ing milk. These are represented as a pair of concepts linked by a relation-
ship, so, for example, “ cats”  and “ whiskers”  are connected by the “ has” 
relationship. Argument Champion  uses these to connect topics together,
allowing a debate to shift focus from schools to computers, from comput-
ers to keyboards, and from keyboards to pianos.
You might have noticed in that last connection that the sense of “ key-
board”  changed from something you type on to something you play music
with. ConceptNet isn’ t perfect, and often conflates concepts (a janitor
is a custodian; The Janitors are a noise rock band). If your game made
these connections, the player might find them at worst a little confusing,
but might actually find them entertaining. Unfortunately, that isn’ t all
ConceptNet thinks. If you type “ woman,”  one of the facts you retrieve
is “ women are sluts.”  Argument Champion  might not come up with that
connection, but the next game to use ConceptNet could well make that
mistake, and most players probably won’ t find that quite as entertaining,
to say the least.
In fact, they might even find it as offensive as if you’ d said it yourself.
As humans, we tend to respond to software being intelligent by expecting
it to behave in increasingly intelligent ways. When we see a game making
clever connections between computers and keyboards, or cats and milk,
we often assume that the software understands the meaning of what it’ s
doing, when in practice it likely does not. That means that when it does
something wrong, we don’ t just see it as an accident— we treat it as if it was
a genuine statement, and that the software knew what it was doing. This
is a great reason why, as developers and designers, we need to think very
carefully about the systems we build.
One approach that can solve some of these problems is using a ban list
for language. Darius Kazemi, a well-known Twitterbot creator who has
done a lot of work with generative software, maintains various ban lists
designed to catch words that might lead to offense being caused. These
lists are often designed to be additionally cautious, under the thought that
it’ s better to accidentally filter out too much than to not filter out enough.
One bad incident is generally all you need to ruin someone’ s day.
48   ◾    Procedural Generation in Game Design

Ban lists are great if you’ re working with language because they work
by filtering the output of your generator, which means they can catch mis-
takes in generators that you might not ever have predicted. The original
Elite , released in 1984, had a name generator for star systems that would
stick together small collections of letters to make long words. The develop-
ers spent some time tweaking this system to make sure it never generated
anything rude— a ban list can help by checking the output of even the
most random words and strings of letters. But they don’ t solve everything,
and data from the outside world can be a real source of danger when build-
ing experimental procedural generators. To give you an idea of why, we’ re
going to look at another game that, like Argument Champion , used online
data. This game was called A Rogue Dream , and was originally made in
2013 for the 7-Day Roguelike jam (Figure  5.1).
The principle behind the game was quite simple— the player types a
noun into the game at the beginning, and the game procedurally generates
a theme for itself so that the main character of the game is the noun the
player input. So, for instance, if the player types in “ cat,”  then they control
a little cat sprite, avoiding droplets of water, eating grass, searching for
cardboard boxes, and using powers called “ Scratch”  and “ Sleep.” 
The way the game did this was by using a technique called “ Google
milking,”  coined by researcher Tony Veale. Google milking works by
reverse engineering the language of question— asking to get information
about the world. When someone asks Google, “ Why do doctors wear white
coats?”  a likely reason for them asking the question is that they believe it
to be true. They’ re looking for a reason, but the fact that doctors wear

FIGURE  5.1   Screenshot of A Rogue Dream  where the player is controlling a cat.
 Ethical Procedural Generation    ◾    49

white coats is considered something true— we can extract the factual part
out of the question and use it as knowledge about doctors.
The more people ask a question, the more people presumably believe
the facts of the question to be true— but how do we know which questions
are popular? Google autocomplete can tell us, because it’ s trained to give
us the most popular search queries. So if we type “ why do doctors”  into
Google, the search queries it suggests are the most common things people
ask about doctors. Why do doctors wear white coats? Why do doctors say
stat? Why do doctors prescribe steroids? Our game might not know the
answer to these questions, but it can use these things to understand more
about what a doctor is.
So if you type “ why do cats hate … ”  into Google, you get “ water”  as an
autocompletion; if you type “ why do cats eat … ,”  you get “ grass,”  and so
forth. Each of these specialized queries is used by the game to generate its
enemies, pickups, goals, and ability names. A Rogue Dream  is entertain-
ing to play, because unlike many knowledge databases, like ConceptNet,
Google is packed with popular culture, everyday observations, and slang.
If you choose to be a ninja, you’ ll be fighting pirates. If you choose to be a
musician, you’ ll be fighting music sensation Kenny G. If you choose to be
a nihilist, your health packs are floating clouds of darkness.
But you might be sensing a problem here, and indeed, A Rogue Dream 
did not get very far in development before unfortunate things started hap-
pening. While a lot of the most generic, innocent queries produced fun
game themes, others carried more weight. If you typed in “ priest,”  as the
player character, one of your abilities would be “ abuse child.”  If you typed
in “ man,”  your enemies were women. If you typed in “ Muslim,”  your abil-
ities included “ k ill.”  People don’ t just type into Google things that they’ ve
directly observed— they type in stereotypes, rumors, misconceptions,
hate speech, conspiracy theories, and worse. A Rogue Dream , unable to
tell the difference between people who see cats eating grass and people
who think French people are lazy, ends up collecting all this information
and using it as fact.
It’ s difficult to throw away work, and it’ s hard to admit an idea is bad.
A Rogue Dream  was almost developed into a museum exhibit for children
to interact with and learn about procedural generation. But after weeks
of trying to filter the system, improve its understanding, limit its search
capabilities, or change its purpose, it was abandoned. Now it’ s mostly used
as an example of how procedural generators with the best of intentions can
end up being harmful— which, as it turns out, is quite useful. Hopefully,
50   ◾    Procedural Generation in Game Design

it’ s also a great reminder of why it’ s important to think about the ramifi-
cations of the systems we build from an ethical standpoint.

YOU ARE WHAT YOU EAT

As an addendum to the previous section, we should also talk about
another way that generators can find bad things out there in the world.
Previously, we focused on real-world data that we might find and want
to use in our game— live data like Google results, knowledge bases like
ConceptNet— and the things we might want to be wary of when doing
this. Just as dangerous is opening our generator up to real people and giv-
ing them some control over the content it creates.
Outside of games, Microsoft learned this lesson the hard way in 2016
when it released Tay, a Twitterbot that interacted with users and learned
words and phrases from them. Tay had a lot in common with a procedural
generator in a game: it created things for other people to look at and enjoy,
from a catalog of data it could intelligently chop up and rearrange. But
Microsoft wanted Tay to interact more closely with its users and to appear
as intelligent as possible, so it allowed Tay to learn things from people
who spoke to it. This had the unexpected (or perhaps completely expected,
depending on your perspective) effect of turning Tay into a mimic for any-
thing anyone told it, from hate speech to pure nonsense. Tay was shut
down within hours of launching, and led to a lot of reflection from engi-
neers and artists about how to create systems that learn from other people.
Tay might seem like it doesn’ t have too much for us to learn from, but a
number of trends in game design and development are leading in the direc-
tion of more human involvement in the games that we play. Games like
Stellaris have built-in mod tools that allow players to edit, upload, and share
data lists for the in-game procedural generation, which means that the
game’ s procedural generator is partly designed by someone who may not
be thinking about the game and content generation in the same way that
the game’ s designer was. Of course, in the case of modding, players are able
to choose what they install and what they ignore. But other trends— like
taking game input from the chat messages on streaming services such as
Twitch— open up the game to unpredictable, hard-to-filter input from a
large quantity of people. If anything, these inputs are likely to be even more
problematic for a generative system than Tay’ s Twitter followers were.
Getting input from people can produce much more interesting and var-
ied content; they can add a dose of human creativity into a rigid system,
and they can help make a generator feel personalized and unique for a
 Ethical Procedural Generation    ◾    51

particular player. But we always run the risk of encountering the worst
side of human creativity with this, and that can be something that is hard
to recover from. Just because opening up our generators is potentially
dangerous doesn’ t mean we shouldn’ t try to experiment and see what new
systems and games we can build. But we need to be aware that these prob-
lems are out there, and do what we can to design our systems to limit these
bad outcomes. One way we can experiment in this area while limiting the
bad outcomes is to restrict the ways we take input into our generator. If
our generator can be given strings of words, then instead of letting our
users have totally free input, we can limit them to a set of a few hundred
words to express themselves with. If our generator uses artwork, we can
let our users snap shapes together instead of allowing them completely
freeform drawing. People will always try to circumvent whatever systems
you give them, but a few restrictions here or there will make life easier and
your generator safer.
Before we close out this section, let’ s dwell a moment longer on Tay,
because in many ways the story of Tay is a good explanation of why this
chapter matters, and why it might have an impact on your game. One of
the reasons why the response to Tay was so negative and so critical was
that once a piece of software is given bad patterns of behavior, the people
who see it repeat those patterns don’ t think about who taught them to it.
They only focus on what the software is doing. If our generator starts pro-
ducing offensive or other kinds of bad content, people won’ t blame Twitch
chat— they’ ll blame our generator. It’ s a totally reasonable response, and
something that should motivate us to do our best to be responsible in what
we build, who we let it talk to, and where we let it do its work.

TALKING THE TALK

Most of this chapter was dedicated to all the things that can go wrong with
our software. It can be taught bad things, it can find out bad things all on
its own, it can spread ideas (both good and bad) far and wide. Of course,
as the designer of that software we are responsible for what it does. But
sometimes we’ re more directly responsible for the issues surrounding our
procedural generators, and so before we move on from this chapter back
into technical and aesthetic concerns, let’ s take a minute to think about
how we talk about procedural generation with the wider world.
Promoting games— or anything else for that matter— is really hard.
Even if money isn’ t involved, simply getting people to look at something
you made is really difficult. A lot of advice is offered to developers about
52   ◾    Procedural Generation in Game Design

how to promote games in particular: advice about who you should talk to,
what you should talk about, and how you should phrase it. Even if you’ ve
never read a thing about promotion, you’ ll probably have noticed common
patterns in the way other people talk about their games or describe them
in stores. It’ s important to summarize things, to focus on the most impor-
tant points, to get across the essence of what it is that you’ ve made, and to
emphasize what makes your game special. What sets it apart? What makes
it unique and worth talking about?
Procedural generation is still a powerful selling point for video games.
Steam has a tag dedicated to it, and many curators recommend games
solely for the presence of procedural generation in them, but games with
generators have been popular for a very long time. “ The heart of Diablo  is
the randomly created dungeon,”  states the very first page of the original
Diablo  design document, “ providing a new gaming experience every time
Diablo  is played.”  Even though it can feel like so many games have proce-
dural generation these days, people are still interested in it. So it’ s natural
for it to come up when talking about games.
It’ s also very easy to get carried away when describing something amaz-
ing that you’ ve created and are proud of. So when we talk to people about
procedural generators, often we say things that express how excited we are,
without thinking about what they really mean. Let’ s go back to that state-
ment from the Diablo design document for a second: “ a new gaming expe-
rience every time Diablo is played.” What does that actually mean? Does
it mean the game mechanics change? If you’ ve played Diablo , you’ ll know
that’ s definitely not the case— you’ re clicking on zombies and casting spells
no matter what happens. Does it mean the levels change? Well, sort of. They
look different and have different layouts. You couldn’ t memorize a level and
run it twice, certainly. But you’ re always going to the same places— the same
caves, the same villages. They are shaped differently but painted the same.
That might sound unfairly harsh on Diablo , because we know what
they really  mean by “ a new gaming experience.”  It’ s a euphemism, like
a lot of language about procedural generation is: “ limitless gameplay,” 
“ infinite replayability,”  “ endless variety.”  These are all phrases we see used
to describe algorithms that are usually doing the computer equivalent of
shuffling a deck of cards and dealing you a new hand. Over time, we’ ve
grown so used to these turns of phrase that we probably don’ t think much
about the language any more— we instinctively know what the person
speaking them means, so we don’ t think about how they sound to people
who are less familiar with these ideas.
 Ethical Procedural Generation    ◾    53

Relying on people to “ k now what we mean”  when we talk about our
work isn’ t a very good idea. People might learn over time, but all we’ re
really doing is passing the problem on to a new community of people who
will make the same mistakes. More importantly, as this book shows, new
ideas are coming out of the procedural generation world every day. When
we try to explain or sell these ideas to people, there will be hardly anyone
who knows what you mean. You’ ll be deciding how people understand
your work, and what expectations people have when they buy your game.
A good example of this is the idea of “ uniqueness”  in procedural gen-
eration. A lot of games try to count how many different dungeons or items
their game has, and use it to illustrate how big their procedural generator
is. The 2009 shooter Borderlands  proudly stated the game’ s 17.5 million
guns as a key part of its marketing campaign— it even managed to find
its way into the Guinness Book of World Records  with this award. But if
we made a copy of a gun in the game and added one point to its dam-
age, most people wouldn’ t consider the two very different except in the
most legal, technical sense. Even if we kept adding damage points until it
felt different, we’ d probably end up in a situation a little bit like Diablo ’ s
dungeons— they’ re distinct from one another, but not unique.
What’ s wrong here? We’ re not specifically lying; we’ re just being a little
misleading or unclear with how we’ re talking about the game. It prob-
ably comes from a good place, from a feeling of excitement and positiv-
ity the developers have about their game. When we put a big number in
front of someone, though, we’ re leaving it up to them to guess at what
percentage of that number will be interesting, useful, fun, or relevant. If
they’ re familiar with procedural generation in games and its limitations,
they probably won’ t mind! But if they’ re less experienced, or perhaps
you’ re offering something new, then there’ s more of an opportunity for
misunderstanding.
That doesn’ t mean we can’ t be passionate and excited about our work,
of course! It just means that we should think carefully about what we say
and write about what we make. Sometimes it can be as simple as changing
the kind of language you use. In Spelunky , the game tells you “ the walls
are shifting”  as a new level is generated. It’ s a small and elegant phrase,
but it tells you everything you need to know: the levels are being shuffled
around; they aren’ t entirely new, but they’ re different to what you just saw.
It doesn’ t promise too much, and it slots neatly into the game’ s theme.
Another approach is to think more carefully about what makes your pro-
cedural generator fun. In many ways, advertising millions of guns or billions
54   ◾    Procedural Generation in Game Design

of levels doesn’ t really make much sense— the average player will probably
only see a few thousand. So what is it that makes our generator cool? What
does it do that makes us smile or want to see more of what it produces?
Maybe it actually produces a lot of rubbish, and maybe that rubbish is what
makes the rare discoveries of something good so wonderful and exciting.
Instead of avoiding these ideas, we can play into them and build our game to
be more strongly based around these ideas of rarity and surprise.
Everyone will have their own idea of how to talk about their procedural
generation, what kind of language they feel comfortable using, and what
story they want to tell people about their game. Don’ t feel like you need
to take everything we’ ve discussed here at face value, but do bear some
of these ideas in mind the next time you tell someone about a procedural
generator.

THE FUTURE
One of the great things about working in procedural generation is how
fresh and unexplored most of the field is. Although games have had pro-
cedural generation for many decades now, we’ ve spent a lot of that time
doing the same things over and over. We’ re getting really good at them
now, but there are a lot of things that people have never even thought of
yet, let alone actually attempted. This book will give you a lot of inspir-
ing ideas to consider and experiments to try, and hopefully inspire some
things that we’ ve never seen before!
With that excitement of pioneering comes the responsibility of being
the first people in a strange new land. It means we should tread a little
more carefully in case something unexpected happens, and that we should
be prepared to think in new ways and break old traditions if the situation
demands it. It also means that the things that we do now, the problems we
decide to solve, the issues we consider important— these become examples
that are set for the people who come later. All the ethical topics in this
chapter do matter, even if they seem trivial right now, because by valu-
ing these issues, we make them valuable for the generations of procedural
generation enthusiasts who will follow us.
Each reader will have a different response to this chapter, agreeing with
some things, ignoring others. All that really matters is that you’ ve thought
about these things long enough to decide what they mean to you, and what
you want to do in response. We hope you’ ll bear some of them in mind as
you enjoy the inspiring chapters in the rest of this book, and as you build
generators of your own in the future.
 II
Procedural Content

T he many domains of potential design expertise can be reflected in

algorithms we use to build our games. Where one designer chooses to
generate a mountain range, another experiments with generating a brush-
stroke and a third interrogates the variables in generating a suitably chal-
lenging opponent. The skills of the designer and the artist can be applied
across a stunning variety of procedural generators.
From Rogue  to Minecraft , the most famous examples of “ procedural
games”  apply generation to their explorable physical spaces. We therefore
begin with procedural level design. From there, this section wanders the
many ways that generation may be employed to create content, whether
audio, visual, or abstract. Specific case studies, broader approaches, and
modes of thinking are explored.
This section allows discovery of different methods of implementation,
as well as a breadth of philosophies of what can define an “ asset”  in a
procedurally generated game. Although no encyclopedia of game content
can claim to be all-encompassing, the reader may find the best inspira-
tion comes from cross-disciplinary study. The following chapters provide
diverse examples serving to illuminate the vast field of possibility and
spark the imagination.

55
 Chapter 6

Level Design I
Case Study

Chris Chung
Independent 

C atlateral Damage  is a casual physics-based three-dimensional

(3D) first-person shooter game where, instead of shooting, you make
messes as an ordinary house cat. In it, you play as a cooped-up cat on
a rampage and use your paws to knock as many of your owner’ s pos-
sessions onto the floor as possible. The game features a variety of levels
within which to wreak havoc, including several different house types, a
supermarket, a museum, and a mad scientist’ s laboratory. When you play
the timed “ Objective Mode,”  there is a set goal to complete (e.g., knock
a total of 150 physics objects onto the floor) to progress to new levels,
along with an optional secondary objective (e.g., knock 15 blue books
onto the floor). All these levels and objectives are procedurally gener-
ated. You can also save the unique numerical value used to generate the
level (its seed) to regenerate and replay it in the untimed, objective-free
“Litterbox Mode.” 
Early in development, I had planned to build each level manually. After
several months, it became apparent that, for a solo developer like myself,
procedurally generating the levels would save time in the long run and
would also allow for many more arrangements than I could possibly make
by hand, thus increasing replayability. I had never built a procedural level
generation system before and was still fairly new to game programming,
so this system is pretty basic and has room for improvement. However, I
am quite satisfied with how the system works and the results it produces.
57
58   ◾    Procedural Generation in Game Design

OVERVIEW
What does a suitable level for Catlateral Damage  look like? Since the game
is a cartoony simulation of a realistic scenario (mischievous house cats
making messes), the levels should be simplified versions of somewhat real-
world interior locales. Each locale is a one-floor interior space divided
into rooms that serve specific purposes. For example, a standard house
should have a bedroom, bathroom, kitchen, and living room, and each
room should contain certain objects that are appropriate for the room
(Figure  6.1). Levels should have lots of different physics objects for the
player cat to interact with in places where the player can reach. Therefore,
furniture should be spread out around the rooms at tiered heights (so they
act as platforms for the player) and covered with objects. An easy way to
think about levels is that each one is a building divided into rooms that
are filled with furniture with physics objects on top. Objectives should
obviously be achievable within a reasonable amount of time for all
generated levels, too.
Generating levels happens in three steps. First, a big rectangular area is
cut up into smaller rectangles for rooms with open doors between them.
Then, each room has appropriate furniture placed within it according to
the type of room (e.g., the refrigerator is only placed in the kitchen). Finally,
the physics objects that the player will interact with are placed on top of
the furniture. Once the level generation is complete, the level’ s objective
and time limit is generated based on the makeup of physics objects that
were placed. When everything is ready, the player is placed in one of the
rooms and can start making a mess!

FIGURE  6.1   Standard bedroom from the first level of the game.
 Level Design I    ◾    59

THE RULES
The rules of the procedural level generation system were designed around
the core mechanics of the game (enhanced, cat-like, first-person shooter
movement mechanics and knocking around physics objects) and what
ideal locations for a destructive cat would be. A couple of limitations were
decided on early on: levels will only be interior spaces (buildings), and
each one will be a single floor. Constraining levels to interior spaces sim-
plified the level creation process and made levels feel nicely contained.
Outside gameplay was not the intended direction for this game; indoor
spaces seemed like they would be more fun. There are also the additional
considerations of what the boundaries of an exterior space would look
like, how to limit it without making it feel too constrained, and if there
would actually be enough objects to knock down. Focusing on a single
floor was entirely to keep level creation simple, specifically to not have to
handle generating additional floors and the placement of stairs.
With the building limitations decided, the rooms that fill the building were
considered next. There should be at least one room in a building, they should
take up space based on their intended function, they should be able to limit
how many doors are connected to them, and each one should know what fur-
niture is allowed inside of it. For the sake of simplicity, only functional rooms
(e.g., bedrooms, kitchens, and bathrooms) are put in buildings; spaces like
entryways, hallways, and closets are not included due to their specific place-
ment requirements and typically small sizes. The room rules all exist to make
the level feel more believable rather than completely random. Being able to
specify how much space each room should take up allows the system to cre-
ate more suitable rooms; living rooms, kitchens, and dining rooms should be
fairly large, while bathrooms should be quite small. Limiting the number of
doors connected to a room restricts private rooms (bathrooms and bedrooms)
to only one door while allowing public rooms (living rooms and kitchens) to
have multiple. The list of furniture allowed in each room is vital for conveying
the purpose of each room to the players while they are playing in the level.
While the building and room rules exist mostly for aesthetic reasons,
furniture rules exist primarily for gameplay reasons. Furniture should be
placed in a visually interesting way, but the main concern is to not hinder
the players’  movement or block their completion of the objective. Typically,
a real room layout addresses both of these since furniture is usually spread
out around the room and does not block other functional aspects of the
room or other furniture (e.g., room doorways and wardrobe doors). The
60   ◾    Procedural Generation in Game Design

procedural furniture rules try to follow these standard real-world rules.

Furniture objects should be placed either along the room’ s walls or some-
where in the middle of the room, furniture should be distributed within
the room in such a way to not block doorways or other furniture, and there
should be plenty of floor space visible so the player has room to knock
physics objects onto it (i.e., rooms shouldn’ t be crammed with furniture).
In addition to these rules, there are also self-imposed guidelines when cre-
ating the 3D models for furniture objects to ensure that they are at appro-
priate heights that the player can reach when jumping to and from them.
Just like how each room should have a list of furniture that can be placed
within it, each piece of furniture should know what physics objects can be
placed on top of it. While some objects, like books, can be placed in mul-
tiple places, certain others should only be placed on certain furniture, like
pillows on a bed. Other information, like where on the furniture objects
should be placed, the probability of each object appearing, and whether
the objects should be rotated, is included with these rules. The rotation
rule is very basic; it just specifies if a physics object should maintain the
same rotation as the furniture (so objects like the TV face away from the
wall) or should be given a random rotation to provide some visual variety.
The level’ s objective also has some rules. Of course, each level’ s objec-
tive should be attainable so the player can progress to new levels. The sec-
ondary objective, which requires a specific type of object to be knocked
down, should only use physics objects that are present in the current level.
The time limit should be enough for players to complete the objective, but
not so long as to make the game too easy.

HOW IT WORKS
The rules of the procedural generation system are made up of several pieces of
hard-coded data. The game takes this data, adds some randomness, builds a
two-dimensional (2D) floor plan (Figure 6.2), and creates the 3D space. There
is one data file for each level type that specifies how large the level is, what
rooms are in the level, what textures to use for the walls, floor, and ceiling, and
more. The generation code uses a squarified treemaps algorithm to take the
big rectangle space and divide it up into smaller rectangular spaces (as close to
squares as possible) with the values specified in the level’ s data file.
Each room in each level also has an associated data file. These files con-
tain the list of furniture that can be placed within the room, with some
information, like the maximum number of each type of furniture to be
placed, the probability of the furniture being chosen, if the furniture
 Level Design I    ◾    61

FIGURE  6.2   Top-down view of a level.

should be placed along a wall, and if the player will spawn there at the
beginning. The furniture objects themselves consist of a model, physics
colliders, an invisible rectangular bounding box (to prevent overlapping
with or being placed too close to other furniture), and some rectangular
surface area objects where physics objects can be placed.
In order for the system to know what physics objects to place, there
is a script attached to the surface objects on each piece of furniture that
contains the list of physics objects allowed on it, along with corresponding
probabilities and rotation information. These surface objects are just rect-
angles on different areas of the furniture (e.g., the shelves on a bookcase)
within which physics objects are placed based on the information included
in the script. The generation code uses a rectangle packing algorithm to
place the rectangular boundaries objects within the surface areas.
Once all physics objects are placed, the procedural system records
and tallies all the physics objects in a dictionary and decides what the
primary objective, secondary objective, and a bonus object that is worth
more points will be. These are always a percentage of the total number of
the objects in the level so the objectives are all attainable. The time limit
for the level is simply calculated by multiplying the primary objective by
a time estimate of how long each object takes an average player to knock
down. Through testing, it was found that it took about 1.4 seconds for
players to knock over 1 object on average, so an objective of 150 objects is
given a time limit of 210 seconds, or 3 minutes, 30 seconds.
62   ◾    Procedural Generation in Game Design

CONCLUSION
Overall, the procedural level generation system built for Catlateral Damage 
creates varied levels that both look natural and are enjoyable by players of
all skill levels. The system not only creates desirable outcomes, but also
has some flexibility for creating handmade content within the ruleset. For
example, it’ s possible to group related furniture (e.g., a dining room table
with chairs) or an arrangement of physics objects (e.g., a row of books) to
be placed all at once.
Feedback from players indicates that most people don’ t notice that
the levels are procedurally generated. This could be both good and bad,
because although the system creates levels that are good enough for the
player to assume they are handmade, they may also assume there is a lack
of content (nine static levels rather than nine types of levels that are gener-
ated procedurally).
The system also isn’ t perfect. There are rare instances where furniture
placement in some rooms is off, like when washers and dryers and the sink
block the toilet in small, thin bathrooms. Minor polish items, like adding
a permanently closed front door, and bigger functions, like multiple floors,
would have been nice to have given more time and/or programming expe-
rience. Even so, this system works great for a silly and casual game like
Catlateral Damage !
 Chapter 7

Level Design II
Handcrafted Integration

Jim Shepard
Independent 

B olstering a procedurally generated play environment with

handcrafted content can add variety and depth to your players’  experi-
ence while at the same time guiding them to play and explore in a method
that uses the best parts of your game system.
In Dungeonmans , I used this technique to create multiple different and
varied adventure zones, such as graveyards, fallen castles, crypts, friendly
towns, mighty towers, and the good old-fashioned dungeon. Here are a
pair of examples that explain how I built these areas, and the goal behind
the design.

STANDARD DUNGEONS
Goal : Expand standard “ room and hallway”  gameplay to include areas
that make the best use of Dungeonmans ’  position and motion-based player
powers and help the game’ s dynamic enemies shine.
Room data : The data is stored simply as text files, with ASCII charac-
ters representing tiles one to one (Figure  7.1). While the blocks of data are
x   ×   y  bound rectangles, the actual room data may contain an irregular
shape. This is important, as the process of placing and layering the rooms
is much more interesting with nonrectangular shapes.

63
64   ◾    Procedural Generation in Game Design

FIGURE  7.1   How ASCII can be used to define playable worldspace.

Two things to note are

1. External room walls are not included, as the game adds those as part
of the area gen code.
2. Data does not include enemy spawns or treasure locations. I didn’ t
feel those were necessary for the game since the area gen code does
a pretty good job of packing the rooms, and I didn’ t want treasure
locations to feel stale. The standard dungeon is the bread-and-butter
adventure area, and these locations would be seen often, so treasure
locations shouldn’ t be static.

Standard dungeon area gen : The area gen code builds out a dungeon
using a time-tested algorithm that extends new dungeon content from
existing areas. This is an iterative loop that continues until you reach a
desired room count or simply run out of dungeon space.
Starting environment : Define the size of your playfield, and define each
tile as unused space . This is not the same as an empty floor tile; this is a
block of undefined protodungeon, a blank slate ready for expansion.
Then, place down a starting room somewhere in that playfield. Generally,
this room includes the stairs back up, unless you’ re feeling cruel. With a
single room full of walls and floor tiles, you’ re ready to begin.

1. Take all existing wall tiles that

• Have two neighboring walls along the same axis
• Are exposed to unused space
and make them into candidates . Take all these and throw them into
a bag so you can pull them out at random, like delicious jelly beans.
2. Select a candidate from the bag, and decide if you want to add a
room or a hallway. We’ ll build a room here. Select a random room
 Level Design II    ◾    65

from the list of handcrafted areas, and then place the room against
the candidate along any given alignment.
3a. If the room fits, excellent. Add a door at the candidate’ s old position, and
recalculate your new candidates given the new shape of the dungeon.
3b. If it does not  fit, that’ s ok. Don’ t place anything; just go back to Step
1 with your existing collection of candidates and rooms. It is possible
to dynamically carve up rooms to fit, but there are two drawbacks.
The first is that you naturally end up with more simple rectangular
areas, and the second is that you lose whatever design value went
into building the rooms in the first place.
4. Once you’ ve placed a room, convert any unused space tile that is
adjacent to a floor tile into a wall (Figure  7.2).
Take note of the second room fitting into the empty space around the
first. Even though the data is stored in a rectangular format, building

Step 1 Step 2

Step 3a Step 3b

Step 4

FIGURE  7.2   Potential generated floor map.

66   ◾    Procedural Generation in Game Design

empty space into the data allows us to fit together rooms in a more
interesting fashion.
Once this is done, you can optionally go back and spice up the
environment a bit, such as
• Find walls that are between two rooms that don’ t have doors
nearby, and add one. This creates the possibility of circular
routes, which are excellent at reducing backtracking.
• Find doors along long empty walls and add a second (or third)
door adjacent, to create a wider opening and prevent easy funnel-
ing of enemies into narrow areas.
• Secret doors! A staple.
The handcrafted rooms for standard dungeons exist to provide
shapes that are fun places to fight in. A mixture of large areas and
small, twisting paths and grand hallways, tight clusters of rooms and
open arenas with cover. You will want to make lots of them, as the
player will see them often, and you want your combinations to have
great variety (Figure  7.3).

CRYPT GENERATION
Goal : Similar to the standard dungeon generation, we want to arrange
handcrafted areas procedurally to build the level. However, the areas are
very specific in shape and size, with the ultimate goal of creating tight
catacomb-style areas interspersed with chambers for breathing room.
Room data : Each room is a 7  ×  7 block, with indicators for exits: north,
east, south, and west. The rooms themselves have additional data for treasure
locations, as well as statues, to help sell the creepy crypt-like environment.
These rooms will be placed out on a grid where exits match. They won’ t all
be solid rooms— as you can see (Figure  7.4), some of the edges have multiple
loose openings, which when aligned together make for areas that appear to be
larger rooms, endcaps, or even the occasional open atrium. In Dungeonmans ,
these rooms are sorted into four lists, one for each cardinal direction. This
means some rooms appear in more than one list, and that’ s fine.

Step 1: Place a starter room. This is a random room tile with at least two
exits. From here, we can run the code shown at the end of the chapter
until we’ re satisfied (or dissatisfied!) with the results.
 Level Design II    ◾    67

FIGURE  7.3   Screenshot of a generated map from Dungeonmans .

FIGURE  7.4   ASCII examples of crypt room designs.

Step 2: Pick a room placed in the level that has an open exit. Note the
direction of that exit (i.e., north) and then draw from a list that has
available exists in the south.
Step 3: Place that new room perfectly adjacent to the first. Recalculate
rooms that have exits, and return to Step 2.
68   ◾    Procedural Generation in Game Design

FIGURE  7.5   Screenshot of a generated Dungeonmans  map, with crypts integrated.

Step 4: Once you have no rooms with available exits, you can evaluate the
dungeon and see if you like the layout. For example, in Dungeonmans 
a crypt is invalid if it doesn’ t contain enough total room placements.
In this case, it’ s tossed out and we start over from Step 1.

Crypt areas feel markedly different from standard dungeons, and this
feeling is complemented by both the art in the crypt and the enemy styles:
hordes that pressure down a hallway, and lots of ice enemies with control-
ling powers (Figures  7.5 and 7.6).

BEST PRACTICES
Here are some pertinent points gleaned over the development process of
Dungeonmans .
Keep the data simple : Fast iteration was super important for these
areas, as was the ability to create a multitude of them quickly. Simple
text files managed in a simple text editor, loaded with a simple process.
 Level Design II    ◾    69

FIGURE  7.6   Screenshot of a hero in a crypt area.

Dungeonmans  is a tile-based game, and text files fit perfectly. Use the sim-
plest possible data format for your project.
Keep the data lean and purposeful : Don’ t feel obliged to fill out every
detail of your handcrafted areas, especially if they are being used as ran-
domized pieces of area generation. Rely on your procedural systems to
populate them with creatures, treasures, and traps. Too many specifics
in your encounters will allow players to recognize every threat in an area
upon entry, which is probably not desirable. If you do place specifics, make
sure you’ re doing it for the right reasons.
Don’ t write an editor : Stopping your development flow to build a new editor
by hand is the surest way to get lost in the weeds. Text files are easy to manipu-
late, and you can do so in Notepad, which may be nature’ s perfect fruit. If your
game requires more complexity, first double-check that it actually does, and if
so, look into one of the many already existing tools for data manipulation that
are quick, clean, and well tested. Don’ t pour time and energy into a compli-
cated system that will cause more frustration than it is worth.

DUNGEONMANS DUNGEON GENERATION PSEUDOCODE 

{
DetermineDungeonWidthAndHeight();
PlaceStarterRoom();
RoomCount = 0;
GoalCount = 20;
NumberTries = 0;
70   ◾    Procedural Generation in Game Design

While(RoomCount <  GoalCount && NumberTries <  1000)

{
NumberTries++;

//every wall tile with two adjacent wall tiles on

the same axis
//and one side exposed to unused space
RecalculateCandidates();

Candidate c = GetRandomCandidateFromJellybeanBag();
TestRoom room = GetCopyOfRandomRoomFromData();

//don’ 
t actually place the room here, just
position it
//and check for overlap with the existing rooms
if( !TryPlaceRoomAdjacentToCandidateInOpenSpace(c,r
oom, out location) )
{
//oops we didn’ 
t succeed
//delete the room if you aren’ 
t in a GC
language ^^
continue;
}

ActuallyPlaceRoom(room, location);
RoomCount++;

}
//now you can fiddle with the layout by introducing doors
//laying out traps and secrets too!
//then fill it with monsters
}

DUNGEONMANS CRYPT GENERATION PSEUDOCODE 

{
//During loadtime sort the room data into four lists
based on N, E, S and W exits.

DetermineDungeonWidthAndHeight();

//Pick a room with multiple exit directions for best

results.
PlaceStarterRoom();

RoomCount = 0;
 Level Design II    ◾    71

GoalCount = 40;
NumberTries = 0;

While(NumberTries <  1000 && HasExits)

{
NumberTries++;

//A room placed in the level with an unmatched exit

Room sourceRoom = GetRoomWithUnusedExits();
if( sourceRoom == null )
{
HasExits = false;
Continue;
}
Room nextRoom = GetRoomMatchingExit(sourceRoom, out
location);

//Make sure your new room doesn’ 

t step outside the
bounds of the level
//or any other consideration
if( !CheckBounds(nextRoom))
{
Continue;
}

PlaceRoom(nextRoom, location);
RoomCount++;
}

//If we didn’ 
t make enough rooms, start over
if( RoomCount <  GoalCount )
{
Cleanup();
Return E_DungeonInvalidTryAgain;
}
//Salt dungeon to taste
}
 Chapter 8

Level Design III

Architecture and Destruction

Evan Hahn
Snowed In Studios 

W indforge  is a two-dimensional (2D) action building block RPG,

set in a floating island dieselpunk world (Figure  8.1). There is a
huge emphasis on the player’ s freedom to explore and create within the
game. Nearly everything in the world is destructible, and players are able
to build new structures anywhere they choose. Functional flying airships
can also be created, which typically serve as the primary way players will
explore the world over long distances.
The world is almost entirely composed of blocks. These give substance
to the world geometry, so that meaningful interiors can be revealed to the
player as it is destroyed. These blocks can also be used as materials by play-
ers to build and craft things.
There are a lot of pros and cons to consider when it comes to deciding
to put procedural content creation in your game, but this wasn’ t a dif-
ficult choice for Windforge . It wouldn’ t be practical to make a game like
this without a least some amount of procedural content creation to make
the world. It would be too time-consuming and tedious to get someone to
build a world of any notable size by hand.
A significant portion of the world in Windforge is generated procedurally. In
top-down order, these are the major aspects of the world that are procedural:

• Area definition and landmark placement

• Playable world areas
73
74   ◾    Procedural Generation in Game Design

FIGURE  8.1   Screenshot from Windforge .

• Architecture
• Object placement
• Enemy placement and spawning

This chapter focuses on the architecture generation techniques we used in

Windforge . These techniques were used to generate the explorable temples,
dungeons, mines, and abandoned buildings that can be found in the world.

ARCHITECTURE GENERATION
We used a divide and conquer approach to generate our architecture. A
complete building or dungeon will typically be complex as a whole, but
fairly simple when decomposed into small individual sections.
The approach starts by determining the bounding area of the structure,
which will then be divided into smaller regions using a recursive process.
These regions provide a starting point where an algorithm can decide the
high-level layout of the individual rooms and how they should be con-
nected. Finally, the individual blocks and objects that will produce the end
result will be placed using a number of techniques.
In total, there are seven main steps to this process:

1. Calculate the bounding box.

2. Split the box into regions.
3. Skim perimeter regions.
 Level Design III     ◾    75

4. Place connections.
5. Assign region types.
6. Make adjustments.
7. Generate the regions.

Step 1: Calculate the Bounding Box

The bounding box defines where the architecture should be placed, and what
its maximum dimensions should be (Figure  8.2). Most of the time, we wanted
our architecture to be at least partially overlapping one of the floating islands in
a world area. Our first step to figure this out was to choose one of these islands.
The selection process was based on what we wished to generate, but for
essential landmarks, we typically chose the largest island in the given area.
Once the island was chosen, we generated a random bounding box
based on the size of this island. Depending on the type of architecture
we wanted to generate, we set up parameters to constrain the size and
aspect ratio of the box. We also had parameters to control the placement
and amount of overlap that should be used in relation to the island.

Step 2: Split the Box into Regions

We used axis-aligned binary splits to divide the space defined by the
bounding box into smaller regions (Figure  8.3). This was a recursive
process. Each time a region was divided, two smaller regions would
be produced, which would be further divided as needed. We typically
stopped dividing regions when one of their dimensions became too small.

FIGURE  8.2   Bounding box of a small dungeon.

76   ◾    Procedural Generation in Game Design

FIGURE  8.3   Recursive splitting at depths 0, 1, and 2, and the final result.

Occasionally, we would also stop dividing early in order to get a bit more
variety and larger regions.
The approach used to choose the splitting axis will affect the result.
Some simple approaches to dividing the region include:

• Split along a random axis : Generates the most variety, but it is the
most unpredictable and sometimes produces undesirable results.
• Always split the shorter axis : Generates long narrow regions. This
provides boring results if used alone, but it works well with hybrid
solutions to provide more variety in region shapes.
• Always split the longer axis : The regions generated with this approach
tend to be similar in size with somewhat square aspect ratios.

In Windforge , we created a hybrid of the approaches above using weighted

random selection. We also introduced a chance of terminating the splitting
process early in order to generate larger than normal regions (Figure  8.4).
Another important part of this step is maintaining the connection
information between the regions with each split. This will be explained in
more detail in Step 4.
Step 3: Skim Perimeter Regions
It would be a little boring if every structure found in Windforge  were a
perfect rectangle. We were able to achieve a variety of shapes by remov-
ing perimeter regions. Different results can be achieved by changing how
these regions are removed.
For dungeons embedded into islands, we typically removed a random per-
centage of the regions along the perimeter of the bounding box (Figure  8.5).
To produce more building-like formations, we used an iterative process
of removing regions from the top (Figure  8.6).
Finally, we were able to produce formations resembling mines and
mazes by traversing over the region connections in a way that avoided
 Level Design III     ◾    77

FIGURE  8.4   Hybrid splitting approach using weighted random selection and
early termination.

FIGURE  8.5   Example of skimming regions from the perimeter before later gen-
eration steps.

FIGURE  8.6   Final result of a building generated using top-down region removal.
78   ◾    Procedural Generation in Game Design

adjacent visited regions if possible, and then removed the regions that
weren’ t visited (Figure  8.7).
The process of removing regions in this step can occasionally cause
regions to float and become disjointed. These disjointed regions will be
detected and removed when we process the region connections in Step 4.
Step 4: Place Connections
This step is responsible for connecting the regions we created in Step 2. To
enable this process, we tracked and maintained all the possible connecting
edges that regions shared after a split. This produced a graph where each
region was a node that could be connected to zero or more other nodes
(Figure  8.8). Each connection also carried information about its bound-
ing region to make block and door generation easier later on (Figure  8.9).
This graph provided a straightforward way to ensure that players
would be able to visit every region we generated without needing to break
through walls. To place the connections, we did a random traversal over

FIGURE  8.7   Final result of a generated mine shaft using traversal-based region
removal.

FIGURE  8.8   Visualization of the region nodes and their connecting edges.
 Level Design III     ◾    79

FIGURE  8.9   Graph formed using the region nodes and their connecting edges.

FIGURE  8.10   Region graph formed after random traversal. The traversal in this
example was biased to favor horizontal movement.

the region graph, and assigned a weighted random connection type over
each edge that was traversed (Figure  8.10).
This traversal can be configured to achieve different results. For
instance, by changing the probability of choosing vertical versus horizon-
tal connections, we were able to control the predominant axis a player
would be traveling on to get through the structure.
After the traversal is complete, it will be possible to detect any
disconnected regions that may have been produced by the region removal
in Step 3. Any region that was not visited by the connection traversal will
not be connected, and can be safely removed during this step if desired.
The connection types that are assigned will guide the generation pro-
cess in Step 7. Some of the types we used in Windforge  include

• Regular door
• Small opening
• Empty connection
• Trap door
80   ◾    Procedural Generation in Game Design

• Boss door
• Treasure door

Step 5: Assign Region Types

This step is responsible for assigning each region a type that will specify
the intended purpose of the region and how it should be generated.
These are some of the region types we assigned in Windforge  (Figure  8.11)

• Treasure room
• Boss room
• Entrance room
• Hub region
• Trap region

We assigned these types in order of importance using a variety of rules,

depending on the region type.
Example rules used to place types are

• Place the boss treasure room in a terminal region at a distance from

the entrance.

• Place the boss room near the treasure room, so that you must travel
through it to get to the treasure.

• Assign trap rooms to terminal regions with ceiling connections.

• Assign treasure and storage rooms to other terminal regions.

Entrance

Storage
room Treasure
room

Boss Boss
treasure room
room Trap
room

FIGURE  8.11   Region assignments.

 Level Design III     ◾    81

Step 6: Make Adjustments

At this point, the high-level features of the architecture will be roughly in
place. Further improvements can also be made by iterating over the region
graph and making adjustments.
One of the most important adjustments we made here was to change
the connection types to better match the connecting regions. Prior to this
step, the connection types were chosen randomly. This was acceptable in
many cases, but there were a few cases where we needed more control over
how neighboring regions were connected.
Example connection adjustments made in Windforge  are

• Use empty connections between adjacent boss rooms.

• Use trap doors for the connections with trap regions.
• Use treasure doors for the connections with treasure regions.

Step 7: Generate the Regions

In this step, we use the connection and region information produced by the
previous steps, and generate blocks and objects for them. This is the first
step that will produce results in the world that players will be able to see.
This step has two main responsibilities:

1. Generate the walls and connections between the regions.

2. Generate the contents of each region.

Generating the region contents is the most involved part of this pro-
cess. This part is responsible for the background pattern generation, object
placement, and platform placement. We managed the complexity of gener-
ating all this by further dividing the region into subregions (Figure  8.12).
This subdivision approach is similar to the region splitting covered in
Step 2. This time around, the subdivision was more flexible and open-
ended. We had a number of subregion types. Some of these types would
divide themselves into other types, which could also be divided. We also
had types that produced custom procedural patterns or structures, such as
background decorations and stairs.
In addition to pattern generation, we used these subregions to guide
platform and object placement. This worked well to ensure that the plat-
forms and objects matched their background decorations, and it helped to
82   ◾    Procedural Generation in Game Design

FIGURE  8.12   Final result of a region divided into subregions (left) and the same
region with outlined subregions (right).
make the placement patterns more regular. Finally, as long as the selected
objects were small enough to fit the dimensions of a subregion, we were
able to guarantee that our objects would not be overlapping each other.

DISCUSSION
The techniques presented here were designed to seamlessly generate archi-
tecture in a larger procedural world. Many aspects of the game design
guided the development priority of these techniques.
For example, Windforge  is a very open-ended game where the player
has the ability to easily create and destroy structures, as well as a variety
of ways to traverse difficult terrain (e.g., by using grappling hooks and
flyable airships). With this in mind, we put less emphasis on puzzle gen-
eration and platform placement than would likely be required in other
games. Instead, our priority was to provide interesting places to explore
with enough variety for a large open world.
The design of this approach made it relatively easy to extend and add
new generation techniques as needed to create landmarks that feel unique
and distinct from each other. We were also able to generate a variety of
structure types just by changing the approaches used for certain steps.
For example, changing how regions were removed in Step 3 allowed us
to generate results suitable for dungeons, mines, small towns, and more.
Most of the architecture was generated on the fly each time a new
area was visited. However, there were a few areas where it was more cost-
effective to manually change the result. We did this with the major towns
of the world and a few areas that required complicated quest-specific
setup. These areas were created by using the usual generation techniques,
and modifying them as needed.
 Chapter 9

Cyclic Generation
Dr. Joris Dormans
Ludomotion 

T here are many different ways of generating dungeons for rogue-

like games. The most popular method seems to be “ drilling out”  the
dungeon from an arbitrary starting point. In this way, tunnels, corridors,
and rooms are added to the expanding level. The big advantage of this
method is that all areas are guaranteed to be accessible from the entry
point. The big disadvantage, however, is that the structure of the generated
dungeon essentially is a branching tree: it will have many dead ends and
force the player to track back frequently.
A good example of this type of level generation can be found in Brogue .
In an interview on Rock, Paper, Shotgun, Brogue  developer Brian Walker
explains that his way of getting around the branching problem is to ran-
domly add doors between rooms in different branches of the tree.* This
strategy seems effective enough; the quality of Brogue ’ s level generator
is testimony to that. Yet, in this chapter I discuss a different approach I
developed while working on Unexplored . This approach does not generate
trees, but incorporates cycles from the get-go, allowing the generator to
incorporate more sophisticated level design patterns.
Unexplored  is a traditional roguelike in the sense that you traverse the
Dungeon of Doom trying to retrieve the Amulet of Yendor and escape the
dungeon with it. The main difference between Unexplored  and games like
Rogue , NetHack , or Brogue  is that it is real time. The gameplay is mod-
eled on top-down Zelda  games, including real-time melee combat and
lock-and-key puzzles. This difference in gameplay also means there are

* https://www.rockpapershotgun.com/2015/07/28/how-do-roguelikes-generate-levels/

83
84   ◾    Procedural Generation in Game Design

different constraints on the levels generated for Unexplored . Levels tend to

be smaller and more focused on a couple of themes (such as the presence
of lava, water, or teleporters), and make more use of lock-and-key puzzles.
An example of one such level can be found in Figure  9.1. It illustrates
how a combination of just a couple of design patterns can create a level
that feels consistent and creates an exploration challenge beyond a series
of individual monsters and hazards. In this case, the patterns are a larger
lock-and-key cycle, with a large chasm in the middle, with an embed-
ded gambit pattern. These patterns are discussed in more detail in the
“ Patterns”  section.

CYCLES
The idea of using cycles for level generation arose during a research work-
shop on computational modeling in games,* and was inspired by previous
research in architecture, urban planning, and hypertext structure. In the
real world, branching trees are rare. In most cities, buildings, and parks,
you can go around in circles. Cycles are also very dominant in handcrafted
levels. Simply google dungeon maps for tabletop role-playing games and
look for cycles in the most interesting ones. In fact, this is one way I har-
vested many of the patterns discussed below.
In Unexplored , cycles are created by connecting a starting point and a
goal by two paths instead of one. Figure  9.2 illustrates how nesting these
cycles often is much more effective in creating dungeon levels than simple
branching.
There are many different ways of using a cycle. Both paths from the
start to goal can offer different challenges to the player. Or maybe if one
path is much shorter, it can be much more dangerous, or simply much
harder to find. One path may actually only be traversable in one direction,
creating a quick route back to the starting point (e.g., after the player has
found a key). So far, I have identified a fair number of these types of pat-
terns, which are discussed in more detail in the “ Patterns”  section.

USING GRAPHS TO EXPRESS CYCLES

Cycles are not easy to work with if your level is represented by a tilemap.
In order to get around this problem, Unexplored  uses graphs to represent
level structure during the initial steps of the generation process. It is only
halfway through that these graphs are converted to tilemaps.

* http://www.birs.ca/events/2016/5-day-workshops/16w5160
 Cyclic Generation    ◾    85

11

2
10 1

6 3
9
12

4
8
7

(1) You enter the level at a platform in the center of a huge chasm. (2) The door on the north is barred. The
only open route is south. (3) Across the lava, you can spot the platform at 6, but you can not really reach it.
(4) You can continue on or check out platform 6, which you might have seen earlier. (5) As you cross the
bridge, it collapses. (6) If you make it to the platform, you are trapped there facing a giant rat. The chest
contains a scroll of teleportation, which allows you to escape. (7) This door is also barred. (8) The narrow
pathway is guarded by a patroling kobold spearman. (9) This section is protected by a sleeping goblin
spearman. You can sneak past easily. But did you spot those triggers that set off an alarm? (10) Similar
section. A trigger might wake up the goblin bowman. (11) Giant ants protect this area, but it also has two
levers operating the doors at 2 and 7. (12) After visiting 11, you can now make your way to this platform
and take the staircase down to the next level.

FIGURE  9.1   Pit-level walkthrough.

This process was inspired by the practice of model-driven engineering.

In short, model-driven engineering advocates that automated processes
use multiple steps to produce the final result. The output of each individ-
ual step is a model, preferably a model that makes some sort of sense on its
own, that gets passed to the next step for further processing.
In general, when one is generating something as complex as a complete
game level, it makes a lot of sense to break down the process into multiple
86   ◾    Procedural Generation in Game Design

Key

start Goal start

Lock
Paths are the base structure of most level generators. Goal
The create treelike levels, often which only one solution.

Key

Lock

Start Goal
Start
Danger

Secret

Goal
In unexplored, cycles are used to create better flow
and to present players with alternative solutions.

FIGURE  9.2   Branching vs. nested cycles.

steps. In fact, I assume that all successful level generators do so. However,
the advantages of multiple steps get hugely enhanced by using different
types of models during the process. Unexplored  initially uses graphs to
represent level structure, as graphs are much better to retain structural
information over distances. For example, Figure  9.3 shows one such
graph showing connected rooms, containing obstacles, locks, and
keys.*
At this stage, all nodes in the graph are deliberately generic. The exact
nature and properties of most obstacles, locks, keys, and rooms are defined
during later stages.
Unexplored  manipulates graphs using transformational graph gram-
mars. These grammars consist of rules that loom for particular pat-
terns in a graph and suggest ways how these graphs can be transformed

* It is important to note that containment in this example is represented as a node embedded in

another node. However, this is not different from a node connected with a special containment
edge to its parent. In fact, that is exactly how it is implemented.
 Cyclic Generation    ◾    87

Bores Bores
Bores

Obx Obx
Obx

Lock

Lock Start

Goal

Key
Key Obx Bores Obx

Obx Bores
Obx
Obx

Open connection Door Window Unlocks

FIGURE  9.3   Graph of a level being generated, with explanations.

by replacing the found pattern with a different pattern. For example,

Figure  9.4 illustrates a graph transformation rule that creates a simple
cycle, while Figures  9.5 through 9.7 illustrate more rules and the graphs
they can generate. It is important to note that the numbers in the rules
are used to match up nodes on the left-hand side of the rule (the pattern
to be replaced) and the right-hand side of the rule (what the pattern is
replaced with).

PATTERNS
Graph transformation rules allow Unexplored  to implement a number
of design patterns. These patterns are common structures found across
games. For example, adding a secret shortcut to a long and dangerous level
is one such pattern. Each pattern has its own positive and negative effects;
88   ◾    Procedural Generation in Game Design

0: start 1: goal 0: start 1: goal

FIGURE  9.4   Graph rule to create a simple graph.

2: goal
1: lock

0: goal 1
1: key

0 1

FIGURE  9.5   Graph rule to add a lock and key.

0: ∗ 1 0: ∗ 2 1

FIGURE  9.6   Graph rule to expand a cycle.

Start Goal
Lock

Key

FIGURE  9.7   Sample of what can be generated from these rules.

 Cyclic Generation    ◾    89

in this case, the shortcut might be rewarding for players that take the time
to explore thoroughly, but it can also bypass a large portion of an other-
wise interesting level. However, the important point is that through trans-
formational graph rules, the level generator is able to manipulate these
patterns directly: it can make deliberate decisions about their use in order
to maximize their positive effects and minimize their potential negative
effects.
For example, most of the time, Unexplored  starts a level by adding a
starting point, a goal, and two paths. The first path, called a , initially
leads from the start to the goal, while the second path, b , leads from the
goal back to the start. Based on the relative and absolute lengths of a  and
b , a number of cycle patterns can be applied. For example, when a  is short
and b  is long, it can place a locked door at the goal, and place the key at
the end of b  just before a spot where the player can cut back to the main
path. This creates a level where players first encounter the locked door,
and when they do find the right key, they do not have to go far to reach
the door it unlocks. Obviously, this requires that the player cannot reach
b  from the cycle’ s entrance straightaway, although it might be interest-
ing if they can already see b  and the key. Figure  9.8 lists a couple more of
these patterns.

IMPLEMENTATION
Of course, the devil is in details or, in this case, the implementation. Graph
transformation and model-driven engineering are powerful methods, but
to make them work in the way we wanted, we still had to add a couple of
things. The most important addition to the graph transformation rules
are attributes attached to nodes and edges in the graph. For example, a
room node can have an attribute that specifies the room type or theme.
The graph transformation rules in Unexplored  can read and write to those
attributes. In this way, the generator can activate a different transforma-
tion rule for rooms that are marked as a “ shrine,”  “ armory,”  “ workshop,” 
and so on.
Attributes turned out to be really important in generating lock-and-
key mechanisms. For Unexplored , simply adding keys and locked doors
was not going to work. There are several game elements that allow players
to push ahead early and often involuntarily: they can jump in chasms to
the level below, use a scroll of descend to similar effects, or use a scroll of
teleportation to teleport to a random location in the level. As the objective
90   ◾    Procedural Generation in Game Design

a Monsters

Start Goal Start Goal

b Traps

Long a, long b Two alternative paths

Start Start

Secret

S
b
Goal Goal

Long a, short b Hidden shortcut

Lock
a
Goal

Start Start
b Valve Key

Short a, long b Lock and key cycle

a Patrolling monster

Start Goal Start Goal

b
Short a, short b Patrolled cycle

Key a
Lock b Lock
Start Start
Lock a Valve Valve Valve
Key b Key
Double key cycle Key behind valve

Some
monsters
Start Start

Visible Many
Goal Goal
monsters!
Dramatic cycle Dangerous route

Collapsing bridge Collapsing bridge

Goal Goal
Visible
reward!
Start Start
Valve Valve Danger?

Unknown return path Lure and setback

Lock
Start Start
Danger?
Visible? Visible!
Key Reward
Simple lock and key Gambit

FIGURE  9.8   Level design patterns.

 Cyclic Generation    ◾    91

of the game is to get to level 20, retrieve the Amulet of Yendor, and make
your way back up, this means that many of the lock-and-key puzzles need
to be solvable in two directions.
A simple solution would have been to make all locked doors operated
by levers and simply make sure there is a lever on both sides of the door.
However, lock-and-key mechanisms can get boring very quickly, espe-
cially when there is little variation to their design. One design objective of
Unexplored  is to use as many game mechanisms as possible as lock-and-
key mechanisms. For example, a potion of resist fire can serve as a key to
a barrier of lava the player needs to cross. Unfortunately, the potion has a
temporary effect, and you are never quite sure when the player will drink
it, and it definitely will no longer work when the player is returning to the
surface.
Using attributes to mark the structural function of different locks and
keys proved to be a very useful way to avoid deadlocks and create variety
in lock and keys at the same time. For example, locks can be marked as
“ conditional,”  indicating a lock for which the player must have a key to
pass. The most typical conditional lock is an indestructible door that can
be opened only with a particular key. An example of a nonconditional
lock would be a room with turrets controlled by a lever. If the player has
not found the lever, he or she can still risk running through the room.
Nonconditional locks have the advantage of being traversable in two
directions automatically, and it is often easier to involve different game
mechanics in their design.

ASIDE: LOCK-AND-KEY ATTRIBUTES

Locks Might Be Conditional, Dangerous, or Uncertain

Typically, we think about locks as binary barriers: if you have the key, you
can cross the barrier; if you don’ t have the key, you simply cannot. Locked
doors, in whatever way they are unlocked or opened, are the typical exam-
ple. Locks as binary barriers are conditional locks. However, locks do not
need to function in that way. Some locks are barriers that might be navi-
gated without a key, but this crossing the barrier might be uncertain or
impose a certain risk. Keys for this type of lock make the crossing less
dangerous, or more certain. A simple example of an uncertain lock would
be a secret door, which is opened if one is able to find it (a certain mat-
ter of uncertainty). A dangerous lock could be a lake of lava that might
be crossed at the expense of hit points (a certain risk). A scroll of magic
92   ◾    Procedural Generation in Game Design

mapping that reveals secret doors or a potion that protects you against fire
damage can act as keys for those doors, making it easier to cross the lock
in both cases.

Locks Are Permanent, Reversible, Temporary, or Collapsing

When you unlock a door, that door might remain unlocked forever (per-
manent), for a short period of time (temporary), or until it is relocked
(reversible). Sometimes, a lock collapses after use, allowing the player
only to pass once. Permanent locks are the safest to use, as once they are
opened, nothing can go wrong. Temporary doors typically produce more
gameplay at a certain risk; they can turn into valves (see the next section).
Reversible locks can create problems, depending on the type of key you are
using for them (see the following sections).

Locks Might Be Valves or Asymmetrical

Certain locks allow you to cross only in one direction (valves), while oth-
ers can only be opened from one direction but traversed in two directions
after they are opened (asymmetrical). Valves and asymmetrical locks tend
to create more interesting conditional locks. They are frequently used in the
patterns above to make sure the player cannot enter a part of the cycle that
is intended as a route back. Valves do not always require a key. For example,
jumping down a high ledge can be considered to be a valve, as it allows the
player only to travel in one direction.

Locks and Keys Can Be Safe or Unsafe

A safe lock is guaranteed to have a solution, while an unsafe lock is not. To
a certain extent, a dangerous or an uncertain lock is always safe. In general,
you want to make sure all conditional locks on the main path through the
game are safe. It is fine to use unsafe, conditional locks on optional paths,
although if that lock is also temporary, collapsing, or a valve, you must
make sure that the path back is safely locked (if it is locked at all).

Keys Can Be Single Purpose or Multipurpose

Single-purpose keys can only be used to open a lock, and for nothing
else, while multipurpose keys can also be used in different ways. Even if a
key can open multiple locks, it can be considered to be multipurpose. In
general, it is better to make use of multipurpose keys, as single-purpose
keys are more boring and frequently end up as dead weight in your inven-
tory. The way Zelda  dungeons frequently use the bow as a key by hitting
 Cyclic Generation    ◾    93

switches from a distance is a good example of a multipurpose key. In fact,

many of the best keys in Zelda  also double as a weapons.

Keys Are Particular or Nonparticular

Particular keys are the only thing that unlocks a particular lock, whereas
several nonparticular keys might unlock a single lock. Just as conditional
locks tend to be more boring than dangerous or uncertain locks, particu-
lar keys tend to be more boring than nonparticular keys, often because
nonparticular keys tend to be multipurpose as well (although they do not
need to be). Particular keys function as do real-life keys. The small keys
of Zelda  are a good example of single-purpose, nonparticular keys. One
effect of using nonparticular keys is that the players, for whatever reason,
might have one when they encounter the lock. When you have not placed
a nonparticular key for each lock you want it to open, this can make con-
ditional locks feel less rigid.

Keys Might Be Consumed or Persistent

Keys that are destroyed somehow in the process of unlocking a door are
consumable, while keys that are not are persistent. Keys that are con-
sumed tend to be less safe than persistent keys, especially if those keys are
also multipurpose and can be consumed to achieve other goals. Small keys
in Zelda  are consumed, as is a potion of resist fire that is placed as a key
to pass a certain fire barrier. Like nonparticular keys, consumable keys
might make a conditional lock less rigid, but also a lot less safe.

Keys Might Be Fixed in Place

Levers and switches are the best example of keys that are fixed in place (and
typically single purpose and particular as well). One problem of fixed-in-
place keys is that if there is only one key on one side of the lock, it makes the
lock asymmetrical. Often, it is best to make the locks triggered by keys that
are fixed in place open permanently, especially when the player can reach the
key only once, which effectively would make the key consumable and prob-
ably unsafe as well. When using keys that are fixed in place, it is often impor-
tant to give some sort of feedback on the status of the door, and to make sure
the lock is permanent and not reversible (as often is the case with levers).

TILEMAPS
Another problem we ran into during the implementation of these techniques
in Unexplored , but also in other games using similar ideas, is how to transform
94   ◾    Procedural Generation in Game Design

graphs into tilemaps. The problem is a generic one: during the initial stages of
the generation process, graphs are used to represent the abstract structure of
the level. Graphs are good for this, as they can include edges to represent the
structural relationships between nodes. These relationships include adjacen-
cies, but also which key unlocks what lock, what enemy patrols which rooms,
and so on. However, at one point the generator needs to build a topological
representation detailing the exact locations of wall, doors, keys, and enemies.
In Unexplored , levels are represented as tilemaps.
Unfortunately, there is no standard solution to this problem. For
Unexplored , we solved the issue by first creating a graph that is laid out
as a grid and stored the topological information in the nodes’  attributes
(Figure  9.9). All graph transformations are then performed on this graph.
This way, when the generator needs to make the transition from graph
to tilemap, there are no issues in mapping the graph onto a two-dimen-
sional tilemap. Initially, individual tiles correspond with nodes and edges
directly, but the resolution of the tilemap is then increased to allow for

Enemy Key
Key Enemy

Cave Room Room

Trigger
Enemy Enemy
Goal

nem
Enemy Enemy

Room Roo Enemy Room

Enemy
Trigger

Enemy

Cav Enemy Cave

Enemy Start

Roo Enemy Room

Initial graph Dungeon structure Low-resolution tilemap

(a) (b) (c)

High-resolution tilemap Populated dungeon

(d) (e)

FIGURE  9.9   Graph grid layout that any dungeon level starts from.
 Cyclic Generation    ◾    95

more details. During subsequent steps in the generation process, transfor-

mation rules, much like the graph transformation rules discussed in an
earlier section, are used to expand rooms, add features as dictated by the
graph, and decorate dungeons.

DISCUSSION
This chapter does not include every detail of implementation. For exam-
ple, before even one dungeon level is generated, a plan for the dungeon is
generated that specifies what themes, monsters, locks, and keys to use on
each level. However, the general idea behind cyclic dungeon generation
should be clear. The implementation in Unexplored  relies heavily on the
use of transformational grammars operating on graphs and tilemaps, but
it is not hard to imagine that other generation techniques could also be
used to generate dungeons based on cycles instead of branching paths.
For Unexplored , the advantages of cyclic dungeon generation are many.
We have found that levels based on a couple of cycles are the most effec-
tive. Two to five cycles give each level a distinct and recognizable shape
and character. This presents the player with a nice, but manageable navi-
gational challenge. With just a handful of cyclic patterns, the number of
possible combinations is still huge. Adding more cycles just seems to clut-
ter the level.
The reason cyclic dungeon generation works well is that as a structure,
cycles are able to express so many more interesting patterns than paths
do. There is no reason to assume that this advantage is unique to cycles
alone. Structures like T-joins or parallel, symmetrical branches might be
able to achieve similar results in addition to, or independent of, cycles, but
so far cycles are the most powerful and expressive structure we have come
across.
In addition, cyclic generation is not restricted to the generation of
sprawling dungeons alone. In fact, the research workshop from which the
ideas behind cyclic dungeon generation emerged tried to come up with a
good way to generate parks or meditative gardens. In such environments,
cycles are also very useful: most people prefer a walk or hike that does
not involve any backtracking. Similar ideas can be applied to open game
spaces and world maps. Cycles can be drawn around obstacles such as for-
ests, lakes, swamps, and mountains, with dangerous shortcuts connecting
two sides of the cycle and prominent sites on locations where two cycles
intersect. There are many opportunities that cyclic map generation offers
that remain unexplored— for now.
 Chapter 10

Worlds
Dr. Mark R. Johnson
Independent 

T he procedural generation of entire worlds is surely one of the

most compelling possibilities of algorithmic content. We are so used
to the standard terraqueous world map we see on the news screen that the
creation of entirely new planetary terrain— and, in most such games, the
gradual discovery of that terrain through the virtual reenactment of explo-
ration and voyage— brings with it a moment of excitement that is never
quite lost, no matter how many games of Civilization , Alpha Centauri ,
Dwarf Fortress , or Ultima Ratio Regum  one plays. Such games bring with
them new and uncharted deserts, vast expanses of tundra and taiga, dense
jungles, open plains, and mountains and ridges and valleys, and even, in
some cases, procedural climates, ranging from heaviest of rainfall to the
driest of lands, and even procedural flora and fauna.
In procedurally generating a world, world elements at both the macro-
and the microscales require modeling (depending on the intended sim-
ulation granularity of the game itself), necessitating either a top-down
approach that builds continents and mountain ranges, and then fills in
the intervening land appropriately, or a system that models natural phe-
nomena such as rainfall, tectonic plates, and weather patterns, and allows
a world to emerge from these processes. If a mere world seems insufficient,
one can go significantly grander in scope and generate solar systems, or
even entire galaxies, if so inclined. At time of writing, I know of no game
that has attempted to procedurally generate a universe, complete with its
own set of algorithmically determined natural laws that differ from play-
through to playthrough, and percolate down into the creation (or not) of
galaxies, stars, and planets, but I throw this out there as a challenge.
97
98   ◾    Procedural Generation in Game Design

Alternatively, for games that don’ t need entire planets and certainly
don’ t need entire solar systems, galaxies, or branes, world generation can
readily be separated into components that function perfectly well in their
own right— one might choose to generate a nation, or a smaller-than-a-
world area of terrain, or cities, towns, fortresses, districts, individual
buildings, the floors within those buildings, and so forth. Although these
smaller examples are arguably not quite the creation of worlds (unless
the building were some truly towering megastructure of the sort found
in Halo  or Dark Souls ), all these are still concerned with the creation of
unique spaces for the player to experience, and it is the grander ends of
this spectrum of practice we’ ll look at here.
This chapter closely examines three elements of procedural world gen-
eration. First, it offers a brief history of world generation with a range of
examples, and we’ ll seek to identify a timeline of sorts and some of the
commonalities of world generation up to the present day. Second, it delves
deeper into the design rationales behind the development of world gen-
eration and the desired gameplay such algorithmic realities are intended
to foster. Lastly, in looking to the future, the chapter explores in depth
my fairly unusual work in this area, in which I have been broadly uncon-
cerned by the creation of physical and spatial “ world”  characteristics of
the sort outlined above, but instead with the development of what I have
termed “ qualitative procedural generation”  and its implementation and
deployment in a world-building context. By this, I mean the procedural
creation of cultural practices, social norms, religious beliefs, ways of
speaking and behaving, and the like, and their physical instantiations and
reproductions within a game world, intended here as a contrast to previ-
ous world generation focuses upon the genesis, variation, and delineation
of procedural in-game spaces and places. In doing so, the chapter gives
an outline of the past, present, and future of procedural world generation;
brings our attention to some of the core philosophies behind these prac-
tices; and examines what exactly it is that makes exploring a procedural
world such a distinctive pleasure.

BRIEF HISTORY OF WORLD GENERATION

It is difficult to establish the moment of the first generated “ world,”  due
primarily to the vagueness of that five-letter word and the attendant dif-
ficulty in drawing the line between a world and a “ very large generated
space.”  However, some obvious early iterations of the idea, or the sorts of
game spaces that would later become indicative of generated worlds, can
 Worlds    ◾    99

easily be identified. Akalabeth  (1980) generated worlds from seed numbers

the player could offer; Elite  (1984) was originally planned to include tril-
lions of galaxies, although its final version contained only a tiny fraction
of that; and then Civilization  (1991) generated Earth-like worlds accord-
ing to assorted variables (land mass, temperature, climate, and age) upon
which real-world civilizations battled for supremacy, becoming perhaps
the first major mainstream game to offer the player an endless volume of
planets to explore. Even the game River Raid  (1982), which offered end-
lessly scrolling procedural terrain, could be understood as the gradual
generation of a world, albeit one that takes some time to appear and can
never be viewed in its entirety. In more recent years, the most noteworthy
examples are perhaps the distant alien planets of Alpha Centauri  (1999),
the intricately modeled high-fantasy worlds of Dwarf Fortress  (2006), the
sprawling cuboid terrain of Minecraft  (2009), the postapocalyptic waste-
lands of Cataclysm: Dark Days Ahead  (2013), and the uncountable planets
and systems of No Man’ s Sky  (2016).
In briefly considering the rise of more-than-world generators, such
as those concerned with creating solar systems or portions of a galaxy,
we actually see something of a shift in the spatial topographies of these
games. Due to the nature of space as being composed of fleeting points
of interest— stars, planets, etc.— in a void of irrelevant nothingness, these
games generally abandon the square or hexagonal grids that are more
traditional in procedural worlds, and instead the systems or galaxies in
other procedurally generated space games like Faster Than Light, Stellaris ,
and Aurora 4X  are node-based networks between which spacecraft travel.
Procedural “ worlds”  are therefore not just limited to standard global geog-
raphy, but can extend into both grander spaces of play and other forms of
spatial representation.
Taking stock of all these examples, we can perceive two different mod-
els of procedural generation. One of these is absolutely exemplified in
Civilization , while the other is exemplified in Dwarf Fortress : these are the
creation of procedural worlds as a space for the player to act upon, and the
creation of procedural worlds as a system for the player to act within. Let
us first consider the generating of worlds as a game space for the players to
affect through their in-game actions.
Games that employ this model create a world that is effectively com-
pletely empty when the game begins, although Civilization  sometimes
places “ Natural Wonders”  or other equivalent points of interest, depend-
ing on the version, and newer versions of Minecraft  can generate worlds
100   ◾    Procedural Generation in Game Design

that contain existing structures of various sorts. In the case of Civilization ,

each civilization starts off with just a single Settler unit about to found a
single city, and the majority of the map will be unexplored and uncon-
trolled by any of the game’ s civilizations for a substantial amount of the
gameplay— in later versions of Civilization , where the player is given more
options, generating a world that is especially large and then populating it
with the smallest possible number of civilizations will lead to a situation
where, even at the end of the game in the present day, parts of the world
will still be untouched by player or artificial intelligence (AI) action. The
map is therefore a blank space upon which the game plays out.
In Minecraft , meanwhile, the player’ s “ goal”  is to construct structures
of whatever sort they desire on this generated world, ranging from func-
tional buildings of the sort seen in the real world to grand artistic proj-
ects. In some extreme cases, the world is truly little more than a backdrop
for action, as it is possible to play Minecraft  with an entirely blank map,
rather than a generated world, and to then enact one’ s creative will on
that world instead. The generation of worlds as a “ space”  therefore holds
to quite a “ sandbox”  conception of the purpose of generating procedural
worlds— these sorts of games leave a tremendous amount to the player,
and are concerned only with making an interesting space for the player to
explore and play with.
The other form of massive generated space we can identify is that of the
procedurally generated system of the world, designed as a complex back-
drop for the player to act within. By this, I mean games that are focused on
giving the player a range of interlocking mechanics dispersed around the
in-game world, such as the thousands of planets and factions in the Elite 
games; the towns, villages, peoples, and still-being-written mythologies of
Dwarf Fortress ; the Iron Age expanses of UnReal World  with its detailed
weather, flora, and fauna simulations; the feuding medieval factions of
Europa Universalis  or Crusader Kings ; and the highly distinctive but glob-
ally intermeshed cultures and societies of my own Ultima Ratio Regum .
These generated worlds are designed less with their terrain and topog-
raphies in mind, although these of course matter and will always affect
other aspects of the game’ s generation, but a world is here instead seen as
something already present into which the player slots, rather than a blank
canvas upon which the player builds. In these “ system”  worlds, we also
see something of a precursor to my own qualitative procedural genera-
tion work to be explored later in the chapter— a concern with a massive
range of interlocking elements, generally nonphysical (for terrains and
 Worlds    ◾    101

biospheres and tectonic plates tend to move a tad slower than social and
political life) but sometimes including shorter-term physical processes,
and on a range of scales from the grand and abstracted worlds of geo-
politics and economies to the microsocial elements of individuals and
conversations, or animal tracking and plant gathering. Players are then
actively invited to engage themselves in these systems and experiment
with the extent to which they can be adjusted, interfered with, turned to
one’ s advantage, and made the center of intriguing emergent phenomena
whereby the player alters one system and sees a knock-on effect in another.
These two rationales therefore produce very different worlds, require very
different forms of procedural generation, and result in games with very
different design objectives.

WHY MAKE WORLDS?

What kinds of feelings and experiences are procedural worlds intended to
foster in the player, and how do they differ from offering large worlds that
have been uniquely designed and packaged with a game?
There are three main reasons to generate rather than handmake a
world: the use of exploration as game mechanics; the use of procedural
generation as a method for creating worlds that are simply too large or too
complex, or both, to ever be handmade; and the use of procedural world
generation as a way to “ enforce”  gameplay variation through providing
different affordances to players each time they start the game.

Exploration
In a nonprocedural game, exploration as game mechanics can only ever be
truly and fully present on the first-ever playthrough of a game when the
player has no prior knowledge of the single fixed game world. Although
games do mitigate this in various ways, such as the difficulty of progres-
sion and the narrative obtuseness of the Souls  series, it is only in proce-
dural worlds where exploration can be consistently fresh game mechanics.
This takes many forms. In strategy games, exploration is essential for the
player who needs to discover the location and disposition of enemy posi-
tions and forces, the distribution of land and ocean and therefore what
units and technologies should be pursued, the location of particularly
valuable resources, and so forth.
In games like Minecraft , and even more so in “ pure exploration”  games
like Proteus , exploration for its own purpose is a central game mechanic
and a focus of the play experience, centered around the discovery of
102   ◾    Procedural Generation in Game Design

interesting sites and varied in-game regions, although in Minecraft con-

cerns of resources and the like are nevertheless also present. In more sys-
temic procedural game worlds, exploration is crucial to uncovering the
best trade routes, the most important clues (as in Ultima Ratio Regum ),
or locations that might merit exploring for loot or story information.
Procedural world generation therefore ensures that exploration remains
a game mechanic that can deliver gameplay experiences on every play-
through, rather than just the first, by keeping every world unknown until
the player actually sets foot within it.

Expansive or Complex Worlds

Procedural world generation can produce in-game worlds of a scale and
complexity that handmade worlds could never hope to match, given the
time and effort required by a developer to place every little part of that
world. This is apparent for some of the most complex generated planets in
games in the present day, such as Dwarf Fortress  and Ultima Ratio Regum ,
but becomes even more apparent when we consider the generation of
entire galaxies, as in No Man’ s Sky — one suspects that there are few game
designers on the planet with the time or motivation to handmake millions
of worlds.
By contrast, a sufficiently robust procedural generation system can cre-
ate worlds that could never be produced by hand, and ensure that all areas
of that world can be potentially extremely complex. The largest handmade
game worlds, such as those of the Elder Scrolls  or Mass Effect  series, for
example, contain only a tiny fraction of in-game areas that can be explored
and interacted with compared with the games listed above.
Of course, not everything generated in a procedural world is of the
same gameplay interest or gameplay value as elements handmade in
another game world, as generating elements with meaning and inter-
est is a challenge in its own right. However, speaking in purely spatial
and volumetric terms, there is no way for procedural generation to be
trumped. This makes such worlds particularly intriguing to explore
for players of a particular mindset who are comfortable with poten-
tially never seeing everything in a game world, and who understand
that some areas will be of less interest than others purely due to the
whims of the underlying algorithm and the lack of a game designer
directly overseeing the process, but instead want the closest possible
experience a game can offer to the true size and complexity of the real
world.
 Worlds    ◾    103

Gameplay Variation
Procedural worlds push increased gameplay variation on the player. This
point is somewhat related to the first in this list, but is distinct— generating
a new world each time a game is played does add replay value, but it also
has the intriguing side effect of forcing gameplay variation and encourag-
ing players toward new forms of play.
If a game of the sort procedural world generation is often used in (strategy,
exploration, roguelike, etc.) was only ever played on a single map, it would
quickly become comparatively simple to divine the strongest strategies and
omit any aspect of the game’ s design that didn’ t fit into that strategy.
By contrast, using procedural world generation forces players to some-
times experiment with parts of a game’ s design they might ordinarily
not use. A player might be planning to execute a particular strategy in
Civilization , for example, but then finds the generated world is covered
with oceans and therefore necessitates the construction of a strong navy
for any chance of victory, encouraging the player toward playing with naval
units and naval strategies he had perhaps never previously employed. A
world might be devoid of a particularly crucial resource in a playthrough
of Dwarf Fortress , which means certain structures or weapons can no lon-
ger be produced (or can no longer be produced in any kind of quantity),
encouraging the players to think outside their traditional in-game rou-
tines and patterns to find new ways to achieve their goals.
Generating in-game worlds of sufficiently high variety is therefore a
way to gently nudge players toward exploring the full scope of a game’ s
possibility space, instead of falling back on well-tried and well-tested tac-
tics and strategies every time the game is played. The combination of these
three elements— the use of exploration as a central game mechanic, the
size and complexity of procedural worlds, and the ability for procedural
worlds to encourage players out of their comfort zones and into the cre-
ation of new choices and decisions— makes playing in procedural worlds
a very distinctive experience, and one that cannot ordinarily be offered
(more than once, anyway) by their handmade cousins.

QUALITATIVE PROCEDURAL GENERATION

With the different kinds of generated worlds established, and some of the
player experiences of actually playing through such worlds, at the end of
this chapter I focus in on a case study of my own work in Ultima Ratio
Regum , specifically in developing what I’ ve termed “ qualitative proce-
dural generation”  as a method for creating complex systemic game worlds.
104   ◾    Procedural Generation in Game Design

Instead of focusing on terrain and environmental modeling (although

these are present and lay the geographical backdrop onto which all the
other elements are added), I’ ve instead chosen to really push the possibility
of generating aspects of “ worlds”  that we experience far more frequently
and far more intimately than any relationship with the physical landscape
upon which we live: that of societies, cultures, politics, religions, institu-
tions, social norms, aesthetics, and the like (Figure  10.1). Although a few
games have touched on generating these sorts of elements, most portrayals
tend to omit the nuance and incredible detail of real-world social interac-
tions, whether in the explicit interactions of conversations and the meet-
ings of cultures, or the microinteractions of how people move around a
space or how they present themselves to others.
I’ m therefore trying to move away from these quite “ mechanistic” models
of culture, and toward a deeper portrayal of individual cultures in the actions
and deeds of their peoples, in shared aesthetic styles and the embodiment of
practices and beliefs, in order to lend a depth, believability, and most impor-
tantly meaning to generated worlds, which are so often lacking.
This task therefore extends from the system rather than the space
approach, as outlined above. Many procedural generation games have com-
plex systems for aspects of social life, like trade or political relationships.
However, instead of abstracting out high-level “ relationships”  between

FIGURE  10.1   Generated religious object.

 Worlds    ◾    105

nations into numerical values of the sort that the Civilization  series does,
perhaps allow the player to walk between cities of two nations. The player
would see how the artifacts of one culture bleed into the other (and vice
versa), or how much linguistic and cultural exchange there is between the
two, how well-patrolled the border is, and how people from each land talk
about those from the other. This is a classic “ show, don’ t tell”  approach,
and one that can only really take place at a particular scale— when deal-
ing with a world-level strategy game, of course individual cities cannot be
explored and depicted, as that would take away from the main focus of
the game.
When controlling a single-player character walking through a gener-
ated world, however, the possibilities open up for portraying a complex
sociological and cultural system in everything the player sees and discov-
ers, rather than just offering a little textual description or numerical value
to this effect instead.
Key to this system is the intersection between all the different elements
the game generates. A particular religious belief about the nature of the
afterlife may affect the colors on book covers; a particularly powerful
house might influence the design of all the weapons and armor manufac-
tured in a nation; the flora and fauna of a particular region might influ-
ence the designs of the sigils of important families; a past war, ancient but
not forgotten, may influence how people from one culture respond to the
player who happens to be from that long-resented foe; the geographical
location of a city might affect the religious beliefs that spring up there,
connected perhaps to the seas or the mountains nearby; and so forth.
Systemic procedural worlds hinge in part upon the interactions between
large numbers of complex systems that have to be uncovered by the player,
but these systems need not be entirely quantitative and reflect only the
flow of trade or battle or some abstracted notion of culture— instead, these
systems can be qualitative in nature and reflect the beliefs and practices of
procedurally generated peoples, and lose none of that desired complexity
in the process. Indeed, if anything, quite the opposite is the case— so many
players of procedural games find their complexity to be one of the most
important selling points (the popularity of Dwarf Fortress  and Aurora 4X 
is a testament to this), and these layers of sociocultural detail only add to
the granularity and detail of the player’ s world-exploring experience.
At this point, one might ask, what is the value of such systems?
Generating spaces offers an area for play, and many games with complex
generated systemic worlds already have a massive wealth of interactions
106   ◾    Procedural Generation in Game Design

for the player to engage in. What value does this kind of qualitative pro-
cedural generation bring to the procedural generation of entire worlds?
As above, part of the enjoyment of exploring procedural worlds lies in
uncovering their complexity, and yet the forms this complexity has taken
to date are only the smallest fraction of forms that real-world complexity
adopts. Incredible detail and nuance can be added to procedural worlds
through these sorts of techniques, especially when the overwhelming
majority of procedurally generated worlds, regardless of their size, are
beset by remarkably large numbers of AI actors speaking the same lan-
guage, building similar buildings, or having few differences beyond the
flags they wave and the procedural histories they share.
The second answer lies in the use of these systems for gameplay— it is
obvious how a complex trade system can be used for gameplay, but what
about a complex cultural system? A complex cultural system allows for
the player to explore the world through means other than walking around
it. It allows for a game to be based around the discovery of in-game cul-
tures and to develop “ cultural knowledge”  as a crucial skill for a player to
be successful; it allows for the player to attempt to fake their cultural or
religious origin as a form of covert or espionage gameplay; it allows for the
player to make informed judgments about their travels in the world, based
on their knowledge of allied and feuding countries. Similarly, it allows
the player to make judgments about what options should be selected in
what conversations, based on their acquired knowledge (acquired through
investigation and learning, not through direct exposition) of the disposi-
tions of the world’ s various peoples.
Qualitative procedural generation is a major future step in the genera-
tion of procedural worlds, both for the sake of the complexity and for the
sake of innovative gameplay.
 Chapter 11

Puzzles
Danny Day
QCF Design 

D esktop Dungeons  is a puzzle roguelike that gives you a disposable

adventurer to guide through randomly generated, single-screen dun-
geons. Your goal is to defeat the dungeon boss(es), leveling up through
combat with less terrifying enemies along the way, and then escape to
pawn boss heads in a kingdom whose economic progress is worryingly
tied to exotic taxidermy. Desktop Dungeons  won an Independent Games
Festival (IGF) Excellence in Design Award in 2011; it was released in 2013
and extensively updated in 2015.
This chapter explains how we built the random puzzle generation sys-
tem that powers Desktop Dungeons ’  replayability and the issues that made
that a unique problem.

PROCEDURALLY GENERATING PUZZLES

For the purposes of discussion, I’ ll assume that puzzles have a deter-
mined starting state and an end or goal state, and that players perform
atomic actions on the puzzle to move it from one to the other (note that
puzzle games may hide or only partially display the game’ s current state
for the player, but the assumption is that the entire state is known inter-
nally). Given that, we can start to consider the kinds of puzzles that can
be usefully procedurally generated and what the outputs of such a system
should be.

107
108   ◾    Procedural Generation in Game Design

Puzzle-Spaces
Algorithms are great at churning out puzzles that are variations on a
theme, puzzles that come from a large and reasonably uniform puzzle-
space. If a puzzle-space is the abstract space populated by every single
puzzle that could emerge out of a set of gameplay rules, then a puzzle-
space can be

• Large  if it contains many puzzles with distinct start and goal states
• Small  if the number of distinct start and goal states is few
• Shallow  if each distinct start state has exactly one corresponding
goal state
• Deep  if a single start state has multiple viable goal states, or vice versa

Puzzles can be defined in terms of the puzzle-spaces they come from.

Human designers are great at producing puzzles at extreme ends of the
puzzle-space spectrum. One-shot puzzles that are enjoyable due to a novel
twist, setting, or presentation require a human in the loop to make them
feel impactful and make poor targets for procedural generation. Puzzles
from a vast or extremely deep puzzle-space generally require a human to
cherry-pick goal states that have some kind of value that makes them stand
out (e.g., humor or carefully selecting for novelty); procedurally generating
within such a puzzle-space produces boring puzzles that feel too similar.

Desired Outputs
Often, the primary role of a procedural puzzle generator is trimming
down a larger puzzle-space into a subset of possible puzzles that differ
along axes that can be easily randomized. A functioning game usually
requires several different types of puzzle to give players a complete experi-
ence. It’ s common to find that you’ ll either have multiple generators that
operate on different “ areas”  within the puzzle-space, or parameters that
move their sampling around within the puzzle-space itself.
One of the most important things a generator needs to do is ensure that
it produces puzzles that are actually solvable! It does nobody any good to
produce a sea of starting states with no guarantee of at least one corre-
sponding goal state for players to reach. Without a goal, you’ re not gener-
ating puzzles and you’ re certainly not respecting the player’ s time. Thus,
our first generation requirement is solvability .
 Puzzles    ◾    109

After that comes the requirement for a continuous solution. A puzzle is

continuous if players can progress from the starting state to a goal state by
following the rules of the game. Applying the atomic actions available to
them should drive the player toward a goal. It’ s surprisingly easy to gener-
ate puzzles that technically do have goal states but don’ t have continuous
paths to those states: if a procedural generation system produced a puzzle
as nonsensical as the infamous Gabriel Knight  cat– hair– mustache puzzle,
we’ d consider it discontinuous. Sudoku-type puzzles are considered dis-
continuous when they have multiple solutions, because at some point the
player has to arbitrarily choose between values for cells instead of follow-
ing the deductive rules of the puzzle.
Continuity  doesn’ t rule out deep puzzle-spaces. It’ s perfectly possible
to arbitrarily remove walls in a single-path maze to create multiple viable
paths; this would still be a continuous maze puzzle. In order to make a
maze discontinuous, you’ d have to remove a reasonable number of adja-
cent paths and walls, rendering maze traversal actions meaningless when
they encountered the lack.
Finally, a deterministic  generator can be very useful. This isn’ t a hard
requirement to produce viable puzzles, but it can speed debugging and
allow players to return to the same puzzle more than once (should you
want to allow them to)— it’ s also a key component of shared “ daily” 
puzzles. Chapter  12 outlines this in detail.

PUZZLE GENERATION APPROACHES

Exactly how you go about structuring a puzzle generator depends massively
on the rules of your game and the kinds of puzzle you’ re trying to produce,
but there are some broad categories of approaches that we can identify.
Examining them each via their puzzle-space and desired output should help
you narrow down good approaches for whatever game you have in mind.

Random Start State

Generating a novel starting state for a puzzle is often a nontrivial prob-
lem, but provided you can create a random initial state, sometimes that’ s
enough. Usually, this is great for puzzles that are about progressively reveal-
ing (or discovering) information, which makes your generation step a sim-
ple rearranging of known data. Of course, this requires that your game’ s
rules guarantee that starting states generated this way are automatically
solvable and continuous. This is easier to achieve with deeper puzzle-spaces
that only have one or two axes of differentiation between puzzles.
110   ◾    Procedural Generation in Game Design

Minesweeper  is a good example of this sort of generation: to generate a

minesweeper puzzle, simply place a number of mines at random positions
in a grid. Players are able to theoretically solve any puzzle generated this
way, as the rules are simply about deducing the placement of the mines.
Any minesweeper puzzle is generally continuous after the first move as
well, even though players may make a mistake and click on a mine, they’ re
only forced to guess when the visible information describes only the cur-
rently revealed tiles. This happens at the beginning of the game when
no squares are revealed, as well as when enclosed “ pockets”  of nonmine
squares are surrounded completely by mines.
It’ s worth noting that random start-state generation can be a great
source of designer inspiration for games with extremely broad puzzle-
spaces. The randomly generated states don’ t even need to be playable,
because all they’ re doing is providing serendipitous new arrangements of
elements that a designer then steals and places into more polished, hand-
tweaked puzzles.

Backward from Goal State

One way to guarantee solvability and continuity is to start from a specific
randomly generated goal state and repeatedly undo atomic player actions
at random until you reach a state of desired complexity. This then becomes
a generated starting state. Backward generation can only produce single-
path solutions; players must exactly replicate the moves the generator
undid in order to reach the goal. This works well for games with shallow
puzzle-spaces and high information transparency in which players won’ t
get lost or discouraged while looking for a single-path solution.
For this approach to work, your player actions need to be bidirectional:
player moves shouldn’ t introduce new information to the game state dur-
ing regular play. It’ s also always a good idea to keep track of previous states
along the generation path so that moves that create identical states (and
thus possible loops or shortcuts) can be avoided or rolled back.
Examples of backward generation include math and word puzzles.
Word search puzzles don’ t start with random jumbles of letters! They start
with the words players are going to find before arranging them in a sea of
randomness.

Heuristics
All the above-mentioned generation paradigms stand to benefit from
specific game-relevant rules: heuristics. We can certainly generate viable
 Puzzles    ◾    111

puzzles purely mathematically, but it’ s possible to generate richer and

more interesting puzzles, faster, with clever heuristic choices. Often,
heuristics take the form of puzzle disqualification states that force the
generator to start again or redo a particular generation step if they’ re
triggered.
It’ s possible to solve the discontinuity problems in Minesweeper  with
relatively simple heuristics based on how players play the game; for
example,

• You could detect a pattern of where players click first when playing
and make sure those squares are guaranteed clear of mines for the
next X puzzles.
• You could offer to suggest the player’ s first move as a guaranteed
safe move, or even perform the least information-revealing move for
them at game start.
• You could simply replace any mines discovered when you detect the
game state to be discontinuous, repositioning the mine in a way that
won’ t invalidate any of the revealed game information.

In Sudoku  puzzle generation, solvers are often used to calculate the dif-
ficulty of the resulting puzzles based on heuristics that provide a difficulty
rating for specific solving techniques. It’ s worth noting that this heuristic
is a postgeneration calculation; if you want a specific difficulty Sudoku 
puzzle, you’ d keep generating puzzles until one was produced (presum-
ably keeping the discarded puzzles in buffers for their relevant difficulties
to prevent needless duplication later).

Extra Bonus: Permutations

Depending on your puzzle rules, it may be possible to transpose elements
of an existing valid puzzle starting state while maintaining solvability
and continuity, essentially generating an entirely new puzzle. While this
approach can broaden the size of your puzzle pool, it cannot help generate
new puzzles from scratch because it always requires a valid puzzle to start
from.
Permutations can be a viable strategy to add variety to your puzzle-
space coverage as a whole, but a reasonable sample size of starting puzzles
is necessary in order to prevent the results from feeling formulaic as play-
ers learn the permutation rules through inspection.
112   ◾    Procedural Generation in Game Design

DESKTOP DUNGEONS, THE PUZZLE ROGUELIKE

Desktop Dungeons  started as a 48-hour prototype when its designer,
Rodain Joubert, wondered how you could make a roguelike that only took
10 minutes to play. The result quickly gained traction, turned into a freely
available “ alpha,”  and was fully rebuilt from the ground up for release
once we understood the game better. That means it underwent open devel-
opment twice, most importantly during a year-long beta that saw us put-
ting out weekly updates. This gave us massive amounts of player feedback
to work with, much of it relating to dungeon generation in some way.

More Puzzle than Roguelike?

Desktop Dungeons  is primarily composed of randomly generated single-
screen dungeons, each 20  ×   20 tiles. Most roguelikes are sprawling explo-
ration affairs with multiple levels that players traverse as they get more
powerful. These small maps are populated by immobile enemies that,
unlike roguelikes, where monsters take turns moving and attacking when
you do, will only attack if you attack then first. The game also has very lit-
tle randomness (outside of the initial dungeon generation); almost every-
thing is deterministic— the game even goes as far as to show you 100%
accurate results of projected attacks so that you can plan better. All the
resources are designed to be granular, and where math became a prob-
lem in player feedback, the game’ s user interface (UI) was progressively
changed to make calculated truths instinctively graspable instead.
The core element that makes Desktop Dungeons  a puzzle experience is
how regeneration works. In most roguelikes, players heal over time (turns)
based on some sort of food or rest value. Desktop Dungeons focuses on
using the condensed nature of its maps and turns unexplored terrain into
a resource: you regenerate health and mana when you uncover new ter-
rain (maps start covered by a fog of war; pushing this back regenerates
any damaged enemies as well). Exploration in roguelikes is traditionally a
boring affair; some even automate exploration until something happens.
Desktop Dungeons  turns this on its head and focuses on making every
player action meaningful, including choosing when, where, and how to
explore.

More Roguelike than Puzzle?

You don’ t get many puzzles with permadeath: the idea of only being able
to try a puzzle only once, and too bad if you mess it up? That’ s usually not
 Puzzles    ◾    113

a great thing for puzzle gameplay, but it’ s great for the tension and desper-
ation that drives players to tell stories of their exploits in roguelikes. The
design goal of Desktop Dungeons  was to condense those moments down
into short, but intense sessions. It achieves this by riding that desperation
line really hard— it constantly feels like you’ re not going to succeed in any
given run. Until suddenly you do, and it feels great!
Desktop Dungeons  is, like most roguelikes, a resource management
sim at its core. The resource model in the game is all about trading off
differently earned currencies against each other to keep ratcheting up
the power level of your character (which is basically shorthand for the
amount of resources you can bring to bear at any one moment). Playing
the game is an exercise in spending health and mana to earn experience,
which increases your health and mana pool. Except you’ re also spending
black space (unexplored regions of the level) to regain health and mana.
But then you need to hunt for enemies in the black space; maybe there’ s
a better target just around the corner? And that’ s completely ignoring
other complexities, like gold, items, the gods, piety, conversion points, and
inventory space.
Everything trades off against everything else, often in ways that aren’ t
atomic actions (and have multiple side effects)— you have to feel the
resource economy on an intuitive level rather than strictly plan your
gameplay responses. It’ s that sort of responsive play that makes Desktop
Dungeons  a roguelike.

PLAYER HOPE AS A RESOURCE

If your players feel like they’ re constantly walking on a knife-edge, one
wrong move away from failing completely, the risk is always there that
they’ ll assume they’ ve made a fatal mistake when they haven’ t. In fact,
players are generally completely wrong when they diagnose an early defeat:
Desktop Dungeons  is a game about gradually uncovering efficiencies in the
resource economy of a dungeon run; thus, players who don’ t know those
efficiencies are going to be wrong when they judge their chances.
The game is balanced to be constantly pushing at the boundaries of
what players are capable of handling. Difficulty is a function of how for-
giving a given dungeon’ s resource-spending values are. Players are also
constantly increasing in skill: a veteran player can absolutely destroy a
dungeon early in the game’ s progression, but it takes desperate situations
to make players see themselves becoming veterans.
114   ◾    Procedural Generation in Game Design

We attempt to train players to keep trying, even if they feel they might
have already lost— in practical terms, the puzzle-space is generally so
deep that there’ s a large buffer for nonoptimal actions in large numbers
of potential solution paths for any given generated dungeon. But this care-
fully herded optimism is completely shut down if the player gets stuck in
a dungeon that’ s actually unsolvable. When a player dies in a run, they
need to feel that they could have done better, not that their attempt was
pointless from the start.

Guaranteeing Solvability
During Desktop Dungeons ’  long alpha, it became clearly important that
any run is solvable. But, given how the game was generating levels, we
couldn’ t guarantee solvability in any of the usual ways:

• The hidden information lurking in a dungeon’ s black space meant

players were constantly making irreversible decisions in discontinu-
ous puzzle states, so the game rules couldn’ t guarantee solvability.
• Generating backward from a goal state was possible but didn’ t lead
to good results; the odds of players finding or deducing the correct
single solution path were far too small, given the massive number of
potential actions available to players at every step.
• Building a solver that played the game to prove a generated dun-
geon had a solution ran into two problems: First, how many solu-
tions were “ good enough”  for the generator, and how would we make
our searches different enough to not simply find a single major solu-
tion path with a couple of actions switched around every time? And
second, it’ s a damn hard game to play already; writing an artificial
intelligence (AI) that can play Desktop Dungeons  was beyond both
our budget and time frame.

We had to come up with a different solution.

Because Desktop Dungeons  emerged as a 48-hour prototype, it used
the quickest and dirtiest generation strategies possible just to ensure that
something possibly playable was ready to go when you entered a new dun-
geon. The game would first generate a map by clearing out lines of walls
until a guaranteed amount of walkable contiguous space was made avail-
able. Then monsters would be randomly placed around the map (up to
40 of them for a single dungeon, including bosses). After that, we’ d add
 Puzzles    ◾    115

pickups and interactive items in empty spaces. Finally, we’ d add the player
starting point and clear walls in a 3  ×   3 space around that. Then it was go
time. Good luck!
That worked surprisingly well, but it could (and did) generate com-
pletely impossible dungeon layouts by “ blocking the player in”  with ene-
mies that were too high level to kill. In the resource parlance of the game,
player progression was blocked by experience having health and mana
costs too high for a low-level player to pay. It didn’ t happen as often as
players assumed it did, but the sheer fact that impossible layouts could
happen lent credence to players’  feelings of unfairness when faced with a
difficult but possible dungeon.

Generating Hope
In the end, we didn’ t change away from random start-state generation.
Adding layers of heuristics to the generation systems produced encourag-
ing results. We also added more terrain generation methods that produced
different overall dungeon layouts for variety.
The first heuristic attempted to keep the available resource costs near
player starting positions within acceptable bounds by introducing a mini-
mum distance threshold to enemy spawns. Higher-level enemies that
spawned too close to the player start location kept rerolling their positions
until they were far enough away. Note that we didn’ t also apply maximum
distances, so lower-level enemies could still spawn far from the player’ s
starting position. Our focus wasn’ t on making sure “ easy”  enemies didn’ t
cost a lot of exploration; rather, we wanted to make sure players were free
to spend exploration if they needed to. This had the predictable effect of
slightly depopulating the area near the player starting location overall, but
this meant that helpful pickups were more likely to spawn in easy-to-reach
locations, so we didn’ t try to correct that.
Minimum spawn distances for enemies definitely helped address the
issue of experience blocking, but we found that player win rates actually
went down during the Desktop Dungeons  beta after the change. It turned
out that the sparse enemy numbers around player starting locations
meant that players were running right past low-level enemies in search
of more pickups, wastefully spending exploration that could have replen-
ished their health and mana after fights. The optimism for exploration
was something we wanted to encourage, but we needed players to at least
have a couple of lower-level fights to encourage good usage of exploration
regeneration.
116   ◾    Procedural Generation in Game Design

Our second successful heuristic made lower-level enemies look near

their random spawn locations for nearby positions that would block off
player movement. This clustering mechanic resulted in layouts similar
to those of the original impossible dungeons, except instead of a level 1
player being blocked by level 5+ enemies, the player would only have to
fight a level 1 or 2 in order to proceed down a passage. Enemy clustering
logic also helped enemies “ cooperatively block”  wider areas.
Finally, we made specific glyphs (spells that players could use to spend
mana to affect the world) that could have smaller but still useful impacts
at lower player levels bias themselves closer to the player start location in
the opposite way of high-level monsters. This helped players have mana
spending options at lower levels and often resulted in providing them with
an extra method to deal with potential blockages.
We tried a number of other generation heuristics during the Desktop
Dungeons  beta, but these were the ones that had the best impacts without
drastically changing the gameplay. That’ s not to say we didn’ t add to the
gameplay as well! The final version of the game had an entire metagame
element that allowed players to bring certain items with them into a dun-
geon run (at a cost). These “ preparations”  would help increase player
resource efficiency, as well as often having side effects that allowed players
to handle potential blockages.
It is unfortunately hard to prove that these changes eliminated impos-
sible dungeon layouts altogether, seeing as we still don’ t do any solvability
checking in Desktop Dungeons . It’ s completely possible that these statisti-
cal approaches only make unfair starting states less likely.
However, there’ s an exception that helps prove their efficacy: one of the
dungeons in the game, Doubledoom, doesn’ t use any of our monster posi-
tioning heuristics. It’ s a callout to the alpha and even warns players that it
might be impossible. Before adding the positioning heuristics, we used to
get regular reports and complaints on our beta forum about unfair dun-
geons. These days, the only complaints we get are about Doubledoom. In
fact, it’ s now standard practice for our veteran players to ask people com-
plaining about unfair starts if they were playing Doubledoom or not. So
far, the answer has always been yes.
We can also look at Desktop Dungeons ’  seeded daily dungeon challenge
as another sampling point. In the 500 daily dungeons shared by players
since the feature launched, every single one has been beaten by at least
32 players. There have been some truly diabolically difficult ones, granted
(hence the 32-player stipulation), but so far the game hasn’ t provided any
 Puzzles    ◾    117

truly impossible dailies. That’ s a success rate I’ m more than happy with!
Players seem more hopeful about the game as well; reviews of the full
game don’ t mention unfair dungeons when that was a regular feature in
writing about the alpha.

CONCLUSION
Procedurally generating puzzles can both produce endless possibilities for
your players and create novel arrangements that inspire hand-designed
puzzles. Generating puzzles successfully requires a pronounced under-
standing of the puzzle-space you’ re exploring.
Defining your generation needs via a deep or shallow puzzle-space, how
many viable solution paths you require, whether solvability needs to be
guaranteed, and whether your puzzles are continuous is extremely helpful
at a conceptual level. The final generation approach you take depends very
much on the interaction between player actions and the information your
game presents.
It’ s usually best to start with a naï ve approach: learning what makes
a bad puzzle is just as important as producing good ones while you’ re
starting to explore a new puzzle-space. You can always layer on heuristic
complexity later as you understand your game’ s requirements and player
limitations better.
A system that covers more of your puzzle-space is always preferable to
one that is incredibly specific: it’ s easier to prune out undesirable puzzles
than it is to coax an overly specific algorithm to produce variety.
 Chapter 12

Procedural Logic
Ben Kane
Independent 

P rocedural generation plays a vital, and largely unseen, role in

Keep Talking and Nobody Explodes . The content we generate is not
the models or the environment or even the game state (though we do that
too!), but rather the game logic itself. Throughout development, we relied
on the generation of data to rigorously enforce the rules of the game while
blurring the line between code and data.
This chapter is presented as a case study for Keep Talking and Nobody
Explodes , so it will be a deep look at the design of the game and what
considerations needed to be made in our approach to procedural genera-
tion. It’ s unlikely that this will be an appropriate solution for your game,
because your game will inherently be facing different design problems.
Instead, read about how our design goals were used to shape our applica-
tion of procedural generation and consider how your own goals should
influence your approach.

BACKGROUND
First, a few words about Keep Talking and Nobody Explodes , because it’ s
a weird game with some weird restrictions. It’ s a cooperative game about
defusing a bomb, but the twist is that only one player, known as the Defuser,
is actually playing a video game. The Defuser sees a randomly generated
bomb, with wires, buttons, symbols, lights, labels, and even a serial num-
ber. To defuse it, he or she must solve a series of puzzles and cut the correct
wire, enter the right sequence, push the button at the right time, and so on.

119
120   ◾    Procedural Generation in Game Design

The Defuser, however, has no idea what the correct solution is. For that, he
or she needs the help of other players, called the Experts.
The Experts have access to the manual, which is a document featur-
ing instructions on how to solve all possible cases for these puzzles. The
manual is a static document, available for free at www.bombmanual.com.
It’ s up to the Experts to obtain a copy of this manual, by either printing it
out or viewing it on something like a computer, phone, or tablet.
The gameplay hinges on player communication, with the Defuser relay-
ing information about what he or she sees, while the Experts sift through
the manual and figure out the solution based on the details that have been
described to them. Once they find the answer, they relay it to the Defuser,
who cuts the wire and saves the day— or explodes.
The communication is where the fun lives, and all of our design deci-
sions are made in favor of fostering interesting communication. Players
should be placed into situations where they need to talk with each other,
trip on their words, get confused, and make mistakes. We use this singu-
lar focus to evaluate design throughout the rest of this chapter.

USUAL APPROACH TO PROCEDURAL GENERATION

Procedural generation is used in several traditional ways. The states of
the puzzles on a bomb are randomized so that there is variance in prop-
erties like the number of wires, their colors, the symbols on buttons, the
state of lights, the presence of batteries, and more. This is what keeps the
game replayable, because each round has a bomb that looks different, and
the solutions to the puzzles, which are dependent on these properties, are
accordingly different too.
At a higher level, the “ missions”  of the game also rely on procedural
generation. A mission is essentially a level or scenario where the puzzles
chosen for the bomb are restricted to an authored subset of the possible
puzzle types. This is because each of the 14 types of puzzles plays differ-
ently, and some will inherently take longer to solve or are more difficult
than others. We therefore create a set of broad parameters according to the
difficulty or style of gameplay we’ re aiming to achieve.
For example, an easy mission might entail a bomb with just three
puzzles from a small subset of the simplest puzzle types. More compli-
cated missions can be authored that result in bombs with puzzle types
that are difficult in combination with one another, such as a puzzle that
requires remembering small bits of information over a long period of time
combined with a puzzle that demands immediate attention, making for
 Procedural Logic    ◾    121

challenging time management. Without tailored missions, a truly random

collection of puzzles on a bomb could vary widely in its difficulty and
players might never experience certain challenging or interesting combi-
nations. Procedural generation allows us to create missions that have some
degree of variance and are therefore replayable, while still satisfying our
need to have a controlled difficulty curve.
These are forms of procedural generation that are very visible to the
player. They will notice that each bomb is different. They can see that
replaying a mission will result in similar selections of puzzles with roughly
consistent difficulty (or if not, they will at least be aware that later missions
are harder than earlier ones, even if they can’ t describe why).
What won’ t be obvious is that the logic of the game is also built on pro-
cedural generation.

A DIFFERENT APPLICATION: PROCEDURAL LOGIC

The rules of Keep Talking and Nobody Explodes  are the meat of the game.
The rules make up the manual that the Experts read, and they are the logic
that the game executes to determine whether the bomb explodes. They
are what force players to communicate back and forth in interesting ways.
The rules are also procedurally generated, which may seem a bit con-
fusing. After all, the manual is available on our website and can be printed
out by players. It has no actual connection to the digital game, yet it abso-
lutely must contain the same logic that the game executes. If the two were
ever to deviate— for instance, if the game was executing a new, different
set of rules— then the whole concept would fall apart. This is a core design
constraint behind our rule generation process.
To make this work, the game shipped with a fixed random seed. That is,
the game procedurally generates its rules, but always does so in the same
way, thus ensuring it matches the manual that players may have already
printed out. This is similar in concept to a game where a large landscape
is generated during development to aid in content creation, but is later
locked down and made permanent before the game’ s release.
Why use procedural generation at all then? The first reason is for future-
proofing. Although our manual is dense with information, it is not impos-
sible for dedicated players to memorize portions of it. We developed the
manual with procedural generation in mind from the beginning so that we
could refresh its contents on a periodic basis if it seemed necessary. Going
forward, we can add additional manual variants for players who wish to
swap out the rules while still supporting previously published rules.
122   ◾    Procedural Generation in Game Design

The main motivation to generate the rules, however, was to ensure our
manual and the game itself would stay in sync. We discovered early on
that creating a manual, well, manually is an extremely error-prone pro-
cess, and the results are difficult to verify. Since a large part of our game
involves players attempting to learn the rules, misinterpreting those rules,
and making mistakes, we had to be absolutely sure that the manual was
correct and accurately reflected the logic of the game’ s code. This was
vital during development when the game was both undergoing frequent
changes and being exhibited at conferences in order to get player feedback.
To enforce the desired synchronization, we took a data-driven approach.
The rules could be expressed and stored as data, which could then be used
to generate a manual from templates and would be executed by the game’ s
logic. The key here is that both the manual and the game would be work-
ing from the same data, and thus we could have some assurances that the
two would be in sync (Figure  12.1).

HOW THE RULE LOGIC IS GENERATED

The puzzles and their rules are quite varied, so the method of generating
rule data differs from puzzle to puzzle. In the simplest case, we need only
scramble a list of possibilities. On the other end of the spectrum, more
complicated puzzles require generating logical queries, as well as corre-
sponding solutions.

Trivial Case
In the Morse code puzzle, the Defuser must observe a blinking light and
describe the dots and dashes to the Experts. The Experts then use actual
Morse code to determine which letters the dots and dashes correspond
to. They then look up the word spelled out in a table and finally relay the
associated frequency back to the Defuser. It’ s a tricky puzzle for players to

Executed by the game

Create rule logic using

Data
procedural generation
Used by templates to
generate the manual

FIGURE  12.1   Diagram of how generated data is used by both the manual and
the game.
 Procedural Logic    ◾    123

If the word is : Respond at frequency:

Shell 3.505 MHz

Halls 3.515 MHz

Slick 3.522 MHz

(...) (...)

FIGURE  12.2   Example of Morse code puzzle rules used by Experts.

execute, but generating the rules is straightforward. The rules are simply
the mapping of code words to frequencies (Figure  12.2).
To generate these rules, we just need to construct this mapping. We
first author a list of possible four- or five-letter words to be used. Then we
randomly associate each word with one of the response frequencies. And
that’ s all there is to it!
It’ s worth noting that we could have overdesigned the solution here and
tried to take into account how similar the Morse code representations of
words would be in order to cause more difficulty. In practice, we found
that players had ample difficulty communicating already, and our design
goals would not be better served by increasing complexity.

Countable Problems
Puzzles have properties, such as color of wires, that vary from bomb to
bomb, which keeps the game replayable. For some puzzles, the possible
varieties are limited, and thus are easily enumerated. The Complicated
Wires puzzle is one such case, but the rules are presented as a set-set
Venn diagram, and the challenge is interpreting the diagram quickly
(Figure  12.3). The four-set Venn diagram means there are four differ-
ent properties that must be examined in order to determine whether to
cut a given wire. This is a small set of possibilities: each property can
be either true or false, so there are 24  =  16 states. Creating the rules is
a matter of randomly assigning a “ cut”  or “ don’ t cut”  solution to each
of the states.
Of course, that would make for a fairly trivial puzzle. Our design goal is
to foster interesting communication between the Expert and the Defuser,
so instead we add a bit more depth by making the solution require some
additional information. For instance, a solution could be “ cut if the bomb
has more than two batteries on it.”  Despite being a fixed instruction (i.e.,
something that can be printed in a manual), it will result in dynamic
124   ◾    Procedural Generation in Game Design

S S

S
C P

C P S D

D
D B
P B
B

Wire has red

Letter Instruction
coloring C Cut the wire
Wire has blue D Do not cut the wire
coloring Cut the wire if the last digit of the
S
serial number is even
Has symbol
P Cut the wire if the bomb has a parallel
port
LED is on Cut the wire if the bomb has two or
B
more batteries

FIGURE  12.3   Four-set Venn diagram.

player behavior based on the particular properties of the randomly gener-

ated bomb.
There are a pool of such solutions, both simple (“ cut” ) and involved
(“ cut if the bomb has a parallel port” ). When selecting a solution, we
weight the possibilities to make repetitions less likely and to encourage
the appearance of the more complicated solutions. By authoring solutions
like this, our procedurally generated rules can help force players to share
more information, supporting our goals.

Not So Trivial: Procedural Logic

Other puzzles have states that don’ t lend themselves to being so easily enu-
merated. Consider our Wires puzzle: it’ s a series of three to six wires, each
with five possible colors. That’ s already tens of thousands of possible com-
binations and would make for a very thick printout if we tried to describe
states exhaustively. It also probably wouldn’ t be much fun if the entire
game was just a lookup table.
 Procedural Logic    ◾    125

• If there is more than one red wire and the last digit of the serial number is odd,
cut the last red wire.
• Otherwise, if the last wire is yellow and there are no red wires, cut the first wire.
• Otherwise, if there is exactly one blue wire, cut the first wire.
• Otherwise, if there is more than one yellow wire, cut the last wire.
• Otherwise, cut the second wire.

FIGURE  12.4   Part of the manual for the Wires puzzle.

Instead, we wanted a more logical approach that used natural language.

We ended up with a series of rules like those shown in Figure  12.4.
We call each of these statements a rule . It’ s a logical predicate (a query 
of the puzzle’ s properties) paired with an actionable solution . It covers a
small part of the possibility space that wire puzzles can occupy, and so
this particular rule will be true for some randomly generated puzzles, but
not all of them. With multiple rules put together, we can cover all possible
puzzle states.

Generating Rules
To generate sets of rules for a puzzle, we author various types of queries
and solutions, and then combine them procedurally. The result is a sur-
prisingly expressive system that can create rules, such as those shown in
Figure  12.5.
Generating rules is more involved than simply randomly drawing from
a pool of possibilities, however. Since the rules make up the logic of the
game, and inherently impact how players will need to communicate, we
needed to keep our design goals in mind when creating our generation
algorithms. Let’ s take a closer look at how a rule is made.

Queries: Asking Questions

A query is designed to ask a yes or no question: Is the wire red? Are there
four wires? Does the serial number start with a letter? A query is based on
a property of either the puzzle or the bomb itself, and evaluates to either

If there is more than one red wire and the serial number is odd, cut the last red wire

Query Query Solution

Rule

FIGURE  12.5   Diagram of how a rule is composed of queries and a solution.

126   ◾    Procedural Generation in Game Design

Query

• Text used in manual: "there (is|are) more than

(batteryCount) (battery|batteries) on the bomb"
• Function used in game: MoreThanXBatteries
• Arguments: batteryCount

FIGURE  12.6   Simplified example of the structure of a query.

true or false for a given puzzle during gameplay. The query has a textual
representation that can be used within the generated manual and is also
tied to logic in the game that evaluates the question.
Figure  12.6 is a simplified example of a query template that asks a ques-
tion about how many batteries are on the bomb itself.
Note that this template can be used with different values for “ battery-
Count,”  so we can easily generate several queries from this single template.
The manual will be populated with the text in the query, substituting in
the appropriate values within a range we define. The game will execute the
function called “ MoreThanXBatteries,”  using the chosen “ batteryCount” 
(say, 1, 2, or 5) to evaluate the answer. And because both the manual and
the game will use the same generated data, we can achieve our goal of
keeping the two in sync.

Solutions: Taking Action

When queries are satisfied, the player must perform an action to solve
the puzzle: cut a wire, push a button, or do some other sort of interaction.
Much like a query, the solution encapsulates this action in a way that can
be used both in the manual and during game execution. Figure  12.7 shows
an example.
Solutions are specific to a particular puzzle type, so a pool of possible
solutions is authored for each of them. As rules are generated, solutions
are drawn from the appropriate pool.

Solution

• Text used in manual: "cut the third wire"

• Function used in game: CutWire3

FIGURE  12.7   Simplified example of the structure of a solution.

 Procedural Logic    ◾    127

IMPROVING THE PROCESS

With the concept of a query and solution, it’ s possible to generate a basic
playable set of rules for the game. Throughout development, we listened
to how players communicated with each other and modified the rule gen-
eration process to create better rules. For our game, “ better”  generation
would mean creating rules that are more varied, require more back and
forth communication, are less unintentionally confusing, and toughest of
all, “ feel”  better.

Better Queries
Queries are the building blocks of the game’ s logic, and they represent
questions that players will need to ask one another during gameplay. Since
fostering interesting communication is the primary design goal, creating
interesting queries is a vital part of the design process. The next sections
are a few additions we made to the query generation process.

Compound Queries
Rules can have several queries associated with them, creating more spe-
cific questions that are less likely to all be true. Complex queries take
longer for players to resolve and are more likely to cause confusion and
mistakes during communication, both of which are desired outcomes for
us as designers. A compound query can look like this:

If there are more than two red wires and the bomb has a parallel
port … 

Compound queries cover a smaller subset of the possibility space and

thus are less likely to evaluate to true. For this reason, we opt to have a
variety of compound and simple queries, so that players aren’ t always
answering no to the majority of questions, as this proved tiresome and
unsatisfying in testing. Our goal is to have interesting communication and
not just more communication, so the rule generation follows accordingly.

Query Contexts
Some queries reference the bomb itself, asking questions about its serial
number or how many batteries are attached to it. These have gameplay
ramifications, as the Defuser must search around the bomb to determine
the answer. However, the answer applies to the bomb itself, and thus any
other rules that make a similar query are going to be less effective, as the
128   ◾    Procedural Generation in Game Design

Defuser will be able to recall the answer rather than searching for it. For
this reason, we make sure that our queries are never based solely on prop-
erties of the bomb itself, but rather involve at least one query that is specific
to a given puzzle module. Otherwise, the solutions to multiple puzzles on
a bomb could degenerate into a single bomb-level query, which is neither
difficult nor satisfying for players.
In practice, this entails adding some information when authoring a
query so that we can classify it as being “ puzzle specific”  or “ bomb spe-
cific.”  The rule generator in turn can take this information into con-
sideration and generate rules that better support our desired gameplay.
Like many improvements, this is not without cost. A more complex rules
generator takes longer to implement and maintain, so any such addition
should be weighed carefully.

Better Solutions
For players, figuring out which solution applies to a puzzle is the chal-
lenging part of the game. Executing the solution is meant to be straight-
forward. After all, this is a game about teamwork and communication, so
that’ s where the challenge should lie. For this reason, a solution directs a
player to take a single action, like cutting a wire. However, in a puzzle with
six wires, that amounts to just six solutions. We needed a way to introduce
more variety.
Our compromise was to introduce a bit of logic into the solutions
themselves. As long as a solution unequivocally corresponds to one and
only one action, it will work within our rule framework. That means that
instead of just saying “ cut the last wire,”  we can do things like “ cut the last
red wire.”  It’ s a subtle difference, but in practice, it successfully results in
more communication and opportunities for errors.
In order to do this, however, we need to make sure that there is a red
wire to cut. This is where the query comes back into play. Recall that a
query is essentially predicate logic, and so when authoring a query, we
can recognize that the query being “ true”  gives us information about the
puzzle itself. For instance, consider this query:

If there are more than two red wires … 

If this query is true, then we can be assured that there is a red wire to
cut. The rule generator is informed of extra bits of information like this
from the query and uses it to expand the pool of possible solutions. This
 Procedural Logic    ◾    129

added complexity to rule generation is worth the extra variety that it pro-
vides, not to mention it makes rules that feel more cohesive as the query
directly relates to the solution.

RIGGING THE DECK: RANDOM THAT FEELS GOOD

In practice, randomly drawing queries and solutions from a pool of pos-
sibilities doesn’ t always result in rules that feel good. Playtest observations
revealed cases where rules sounded repetitive or ended up being easier to
answer than intended. We realized that truly random rules were undesir-
able. Instead, we found we got better results by weighting our selections.
Each time a particular query (or even one from a group of related queries)
is chosen, the rule generator adjusts that query’ s weight to make it less
probable that it is chosen again in subsequent rules. This results in more
variety, and the distribution of types of queries feels more balanced.
In general, we use this weighted selection approach to make less desir-
able selections less likely. We could opt to simply remove the option from
the pool altogether and prevent any sort of repetition, but we found that
having rare exceptions was actually beneficial. It leads to players making
incorrect assumptions and taking shortcuts, only to have their assump-
tions sharply corrected as they realize they have gotten too comfortable
and made a mistake.
Weighting is also used to tune generation toward more desirable
options, which is an approach we employ to emphasize the challenge in
some puzzle types. For instance, a puzzle based on remembering your
input from previous stages is weighted so that rules in the later stages
are more likely to reference actions in the earliest stages. Again, though,
we choose to weight these choices to be more likely rather than outright
enforcing them in order to keep things varied and prevent players from
making too many assumptions.

Dealing with Degenerates

Random rule generation can occasionally result in rules that are degener-
ate cases of what is intended. In one puzzle, players are tasked with find-
ing the first matching word in a list. By simply scrambling the list, the
desired word can end up anywhere— even in the very first position! This
makes for a rather trivial search. We could modify our generation logic to
ensure the word only appears midway down the list, or even right near the
end. But players will catch on to such things and will realize that they can
130   ◾    Procedural Generation in Game Design

ignore the earlier portions of the list. Procedural generation alone won’ t
be enough to solve this design problem.
Instead, the puzzle design makes use of many such lists, and the scenario
where one or two are degenerate search cases is not a problem. Players are
quick to identify these and feel a degree of mastery in their knowledge. If
we were to try to “ fi x”  this problem with a more involved procedural gen-
eration algorithm, we would actually be working against our own design
goals. It’ s important to keep in mind that more sophisticated procedural
generation is not the solution to every problem.

Avoiding the Impossible

As the complexity of rule generation increases, it becomes possible to gen-
erate invalid rules. For instance, consider what may happen if any combi-
nation of queries and solutions were to be permitted:

If there is more than one blue wire and there are no blue wires, cut
the blue wire.

This is not an acceptable rule. Logic purists may argue that it’ s simply
a rule that can never be satisfied, but in practice, most players will be con-
fused in a way that is undesirable. Rule validation is one approach that
could be taken to address this situation: generate a rule, check if it makes
sense, and keep generating rules until you get one that does. Inefficiency
aside, it’ s not a trivial task to verify whether a rule makes sense. Bombs,
and their puzzles, have randomly generated states with billions of possible
combinations, so exhaustive testing is intractable. We would require a bet-
ter understanding of what is valid and what isn’ t in order to prove whether
a rule is acceptable.
Instead, we recognize which combinations have the potential to be
invalid and remove them from the possible query or solution pools. The
general reasoning is that we wish to avoid compound queries that ask dif-
ferent questions about a single aspect. In the example above, there are two
queries that both relate to how many blue wires are present. To prevent the
situation from ever occurring, we simply remove all “ blue wire”  queries
from the query pool as soon as we select one of them.
The keen reader may notice that this approach will also remove some
valid combinations. For instance, you can no longer generate valid com-
pound queries such as

If there is more than one blue wire and less than five blue wires …  
 Procedural Logic    ◾    131

We could devise yet more sophisticated generation algorithms to prune

invalid combinations more closely, but first consider if the results are worth
the investment and complexity. In our case, the design goals of our game
could be met with greater query diversity rather than trying to ensure that
every last valid combination could be included. It was much more practi-
cal, and manageable, to simply create additional types of queries.

PUTTING IT ALL TOGETHER

Rules deliberately cover just a small part of the possibility space for a
puzzle. Players are expected to evaluate several rules by asking questions
back and forth, eventually finding the one that applies to the bomb that
was generated for them. No matter what combination of wires and colors
a bomb is generated with, one of the rules will apply and there will be a
single correct solution.
There are two small considerations that may not be immediately apparent
but are vital to guaranteeing one and only one solution. The first is that the
list has a defined precedence order. There may very well be several rules that
are satisfied within the rules list for a given wire puzzle, but only the first
satisfied rule is deemed to be the correct one. The second consideration is
that we always include an “ otherwise” case with no associated query as our
final rule. This is a simple way of guaranteeing our rules list will contain an
applicable solution regardless of the properties the bomb is generated with.
The otherwise rule is a necessary addition to keep our rules list from
ballooning in size to cover the possibility space. Initially, we were con-
cerned that it would be difficult to ensure that the rules weren’ t too spe-
cific, covering too little of the possibility space and leaving the majority of
playthroughs simply degenerating to the otherwise case. In practice, this
actually proved to have some gameplay benefits when used in modera-
tion. Players would become complacent and make assumptions based on
their experience with the likelihood of particular solutions. It made the
rare exceptions that much more noticeable when their assumptions failed
them and suddenly the tamest puzzle had caused them grief.
As an improvement to gameplay, we also sort the precedence list from
most specific (i.e., the greatest number of queries) to least specific. This
makes it more probable that players will need to check several rules before
finding one that is satisfied. It’ s desirable to have this gradient of rule spec-
ificity because just having a lot of very specific (and thus very rarely satis-
fied) rules ends up feeling repetitive, as the solution will almost always just
be the otherwise catchall case.
132   ◾    Procedural Generation in Game Design

CONCLUSION
Procedural generation can be a valuable tool during development for con-
tent creation, even if that variety isn’ t exposed to the end user. The content
you create is limited only by your ability to model the data, so even things
like logic can be generated if you have reason to do so.
We tried to keep our core design goal of fostering interesting commu-
nication in mind throughout our project. That pushed us to expand our
procedural generation logic in cases where it would support our design,
such as creating more sophisticated queries and solutions. Perhaps more
importantly, it kept us from getting carried away with overengineering the
generation algorithm to squeeze out every last valid combination when a
simpler culling approach worked just fine.
Procedural generation is fun, but shipping a game is more fun. Always
keep sight of your design goals and ensure that your procedural genera-
tion is in service of those goals.
 Chapter 13

Artificial Intelligence
Dr. Mark R. Johnson
Independent 

N ot all procedural generation (PCG) games need artificial

intelligence (AI). Some games have no other actors within the game
world beyond the player— games like Proteus  and In Ruin s are focused on
the feeling of exploration of an unpredictable space, specifically designed
to present enough interest to the player to make this process engaging
in its own right. Such games are concerned primarily by the experience
of discovering the algorithmic world rather than anything else, and such
exploration forms the only form of gameplay they offer. However, when
used, PCG AI brings with it a particular set of design necessities and
possibilities.
AI in procedural games must instead be able to cope intelligently (or
at least with the resemblance of intelligence) with near-infinite varieties
of spaces, other actors, and gameplay scenarios. AI in handmade games
needs to deal with unpredictable situations, but it is only in games with
a strong PCG component that AI might be forced to navigate entirely
unknown and unforeseen terrain, deal with orientations of gameplay ele-
ments that were potentially not even understood by a game’ s designers as
being within that game’ s possibility space, and still present whatever level
of AI challenge and complexity was intended.
This chapter focuses on player experiences of AI and designer inten-
tions of AI, common requirements for AI in games with PCG, and new
and innovative AI development methods that offer potential for the sub-
field’ s future.

133
134   ◾    Procedural Generation in Game Design

UNPREDICTABILITY AND ARTIFICIAL INTELLIGENCE

AI often contains elements of unpredictability, as governed by some form
of random number generation system that is hidden from the player. There
are several reasons for this:

Challenge : By making players respond rapidly, increasing replay-

1.
ability, opening the possibility of emergent phenomena through the
unseen interactions of disparate in-game elements, and so forth.
Unpredictable AI can boost the range of gameplay possibilities over
a more deterministic equivalent.
Exploit resistant : Opponents who behave predictably can often be
2.
exploited. I’ m sure most readers will recall playing older games
whose AI characters followed quite simple processes that could eas-
ily be taken advantage of to subvert design intent. An AI designed to
always dodge a grenade is prone to players using grenades to “ force” 
the AI to dodge in a particular direction; an AI designed to pursue a
player can be lured into traps. Deeply unpredictable AI reduces this
possibility.
Balance : The technical development of an effective AI that succeeds
3.
well at the mathematical and game-theoretic challenges of a game
is vastly simpler than designing an AI in such a way that players
actually want to play against it. A “ fun”  opponent includes elements
of game design, psychology, sociology, and even interface design,
whereas the former is concerned only with computer science and
game theory.

It’ s generally simpler to design a strong or even game-theory-optimal

AI than an AI that can appropriately and competently weigh up metrics
of challenge and fun. The latter requires selecting an action that perfectly
matches up the player’ s current gameplay condition and psychologi-
cal condition. AI systems like the AI Director in the Left 4 Dead  games
attempt to do precisely this, but it is relatively uncommon.
One of the most active current areas in game AI design is the concept of
designing for fun as a counterpoint to designing for challenge, but such a
concept is far harder to write into code than a mathematical decision of the
best in-game move. Otherwise strong AIs can simply be weakened a little
in order to produce an effect that, on the surface, is quite similar. In the
 Artificial Intelligence    ◾    135

roguelike game Dungeon Crawl Stone Soup , for example, many bosses and
high-end enemies are given a selection of “ spells,”  from which (in some
versions of the game) they choose randomly each turn. This is designed
to provide unpredictable challenge, but also to balance the game’ s more
difficult enemies through giving them a selection of stronger and weaker
spells. This allows bosses to deploy deadly attacks without being game-
theory optimal and using their strongest attack on every turn, which
would result in enemies that were far more challenging (although also far
more predictable) and far less interesting.
Players can also assign spurious intention and agency to simple and
semiunpredictable AI patterns. In early roguelike games, monster AI was
rarely more complex than “ wander around, and then when the player is
sighted, pursue them by the most efficient path possible.”  Players, however,
often thought that a far more complex monster AI was at work, and that
monsters would deliberately show up at the worst possible time for the
player. What was happening, of course, was that monsters that appeared
(whether through random movement or active pursuit) at moments where
the player character was safe, well rested, and at full health were rarely
remembered, since those resulted in trivial fights.
By contrast, if the player had just fought a brutal battle and was at
death’ s door, and then two other monsters come around the corner,
that becomes an event the player will remember for a long time and may
subconsciously blame on the monster’ s decision making, instead of the
unlucky vicissitudes of fate or the eventual outcome of a tremendously
simple follow-the-player AI routine.
The unpredictability and predictability of AI can therefore readily
become interwoven and confused even without the role of PCG. PCG’ s
inclusion in the design processes leads beyond these concerns of player-
side understanding of AI actions toward a range of designer-side questions,
focused on how AI should be programmed to respond well in procedur-
ally generated spaces and settings, and how the player will engage with
such AIs.

MOVEMENT AND COMBAT

AI in the overwhelming majority of games tends to be heavily focused on
two related but distinct elements— movement and combat. This is as true
of PCG games as it is of their handmade cousins, but PCG games bring
with them a distinctive set of design challenges.
136   ◾    Procedural Generation in Game Design

In the majority of cases, AI actors will be expected to move over terrain

with likely millions of possible permutations, preventing game designers
from programming in methods for handling specific regions, and instead
resulting in a focus on the handling of more general case and scenarios,
which requires the accurate prediction of all possible categories of move-
ment scenarios.
If an AI actor is expected to move around a city, then code must be
put in place to ensure they can traverse wide alleys and narrow alleys;
get around other individuals; and handle bridges, rivers, other traffic, and
whatever else might be present within that particular world. This uncer-
tainty over the tasks that AI in PCG games will be asked to perform is
further complicated by different kinds of terrain, from standard ground
to ice, water, lava, temporary hazards, oceans, mountains, deserts, alien
terrain, and so forth.
Such extra layers cannot be specifically predicted, and therefore the
ability to navigate them cannot just be coded into a limited number of
creatures who will potentially need to pathfind through such situations,
but instead all creatures need these abilities. This leads to particular kinds
of monster design architectures that are easily modifiable, easily influ-
enced by the particular affordances of a specific creature (cannot traverse
lava, can traverse water, etc.), and applicable to any creature so long as it
has the appropriate information relating to its abilities and restrictions.
All possible permutations must be accounted for— even if a city with
an enclosure of buildings is a one-in-a-million possibility, if an AI starts
within that enclosure, it must be able to figure out how to enter one of
the buildings and then leave on the other side beyond the enclosed space.
If designers do not anticipate the possibility of that enclosure, then it is
unlikely the AI will be able to respond to it, resulting in the game’ s PCG
upsetting the intended behavior of the game’ s AI.
Combat, meanwhile, is a staple of computer games generally, including
those with PCG. The design requirements for combat in PCG games is
broadly comparable to the concerns over movement outlined above, which
is to say that strong PCG AI needs to be programmed not to handle spe-
cific forms of combat in specific known-beforehand contexts, but rather to
respond to any possible scenario. In roguelikes, this normally means intel-
ligent decision making with regard to various weapons and various ranges
of combat, and more advanced strategies, such as retreating, flanking, or
calling for assistance, and an AI needs to be able to assess these without
any prior notion of the combat scenario it will find itself in.
 Artificial Intelligence    ◾    137

Once again, there is the question of the extent to which all combat
scenarios can be predicted, and therefore factored into the AI design pro-
cess. Movement and combat thus both pose comparable design questions
centered around the extent to which designers can predict all possible
PCG scenarios— or rather, the elements of all scenarios, which the AI then
identifies— and the extent to which the game’ s PCG and AI systems can
thereby become closely integrated. If this integration is low, unusual out-
comes will leave the AI unable to make a sensible choice about its actions,
resulting in the AI either freezing, defaulting to a simplistic option, or
selecting a bizarre option for the current scenario; however, if this inte-
gration is high, players will see procedurally placed AI characters behav-
ing well in procedurally generated spaces, irrespective of the nature of the
space and the condition of the character, and thereby behaving compara-
bly to convincing AI in an equivalent handmade game.

AMBIENT BEHAVIOR
Ambient behavior refers to AI that carries out certain actions even when
the player is not nearby or engaged with the AI actor. Perhaps the current
and most fascinating exemplar of ambient behavior is the science fiction
roguelike game Cogmind , in which the player is placed in a world contain-
ing large numbers of robotic “ species,”  many of which ordinarily ignore
the player and go about their business without being concerned by the
player’ s actions, unless the player attacks them, interrupts their routines,
and so forth. This poses a challenge for designers because the AIs must be
able to “ go about their day”  on any possible map, necessitating the devel-
opment of a system that can identify points of interest for each AI actor
when the map is generated, and build up a completely unique spatial and
temporal schedule for that actor each time a level is generated. Equally,
however, it affords opportunities for creating game spaces that feel deeply
alive and lived in, even if they had only actually been generated a split-
second prior, and therefore performing a kind of visual trick on the player
ordinarily only possible in handmade games with AI actors who have been
given meticulously thought-through handmade schedules for their days.

EMERGENT PHENOMENA
Emergent phenomena  refers to situations where sets of simple behaviors
lead to often unanticipated sets of more complex behaviors. The clas-
sic roguelike NetHack  perhaps contains the most illustrative examples
of emergent phenomena. NetHack  contains a range of in-game systems
138   ◾    Procedural Generation in Game Design

and AI behaviors that can readily be overlooked by players— the game is

technically winnable without ever truly engaging with or understanding
some of these systems and AI behaviors.
For example, if reading a scroll of “ genocide” would normally wipe out an
entire species of monster, a cursed version of this item could result instead
in the creation of those monsters around the player. Deliberately reading a
cursed scroll of genocide and attempting to wipe out the “ nurse” creature
will spawn a ring of “ nurses” around the player, whose “ attacks” in fact heal
the player and can even boost the player’ s maximum health. This means
that a desirable item used to destroy monsters was cursed to become “ bad,” 
but some uses may still result in a “ good” outcome, through the interaction
between the behavior of those AI monsters and other in-game elements.
In this and many other emergent phenomena, AI and monster behaviors
are integral to these processes, and although they are almost certainly both
entirely unintended uses of the AI, these kinds of emergent phenomena are
compelling aspects of AI behaviors in procedurally generated games.

CONVERSATIONS
One of the most interesting aspects of AI in PCG games is the generation
of speech and conversations appropriate to a procedural world. Ultima
Ratio Regum  (URR ) development is focused on a complex conversational
AI rather than a complex combat, movement, or ambient AI, as conversa-
tions are fundamental to how players of URR  engage with the game’ s AI.
We can reasonably split generated conversation elements into two cat-
egories— the PCG of the actual words that the AI actors speak, and the
conversation system itself.
The procedural sentences spoken by URR ’ s AI characters consist of a
range of elements, which draw on procedurally generated dialects, and
also the specific opinions and current mood of the person doing the speak-
ing. World generation in URR  includes creating a unique dialect for every
in-game culture, of which there are ordinarily roughly 40, consisting of
large feudal nations, smaller tribal nations, nomadic nations, and perhaps
someday independent city-states (Figure  13.1).

Dialects
Generated dialects consist of several components.
First, the game selects a group of consonants, vowels, and syllables, from
which all “ invented” words for that dialect (such as the names of places and
people) will be constructed— this is a reasonably standard technique for
 Artificial Intelligence    ◾    139

FIGURE  13.1   Conversation in URR .

generators that produce words meant to be indicative of a shared language,

and results in groups of words like Xonatov , Vonnerto , Toravert , or Kaeltram ,
Malatrakel , Altrakem . Different languages and dialects are distinctive and
identifiable by an observant player as belonging to one culture or another.
Second, dialects include generated “ references,”  which are cultural,
historical, political, and religious elements used by all speakers within a
certain culture. These might include nearby terrain; the dominant religion
within a nation; recent history, political ideologies, or diplomatic rela-
tions; or artistic and aesthetic preferences. These are then present in all
conversations from individuals in a particular culture, strengthening the
similarity within one culture, and the differences between many.
At this point, also, we begin to see the development of a conversation AI
that is rather more complex than that in most games, where characters from
different cultures are either handwritten to be different or produce the same
set of identical or almost-identical statements regardless of cultural back-
grounds. The “ rumor” options in conversations with Elder Scrolls  characters
provide a particularly striking example of culture-agnostic character speech.
The third element of generated dialect is the generation of greetings,
insults, farewells, and threats. This does not directly alter AI behavior, but
it does continue to develop nonplayer character (NPC) actors who, when
they speak, sound extremely distinctive as a result of the procedurally
generated origins of their procedurally generated character.
The fifth and final element of generated dialect is what I’ ve termed sen-
tence complexity, which affects how much an individual says. This is the
140   ◾    Procedural Generation in Game Design

difference between introducing yourself in a few words and including a

short list of the great achievements of your nation in your introduction.
More loquacious AI actors therefore offer far more information to the
player than their more taciturn comrades.
The design goal here was to create interesting, procedurally gener-
ated, culturally distinctive AI actors who would speak in unique ways
dependent on their origins. As an element of game design, this is almost
entirely overlooked in even handmade games, let alone procedural games.
Conversant game characters tend to lack any real differences based on
their origins, although a rare few games, such as the Mass Effect  series, are
exceptions. Equally, as above, PCG rarely affects AI speech and conversa-
tions, Dwarf Fortress  being the obvious other current standout project in
this regard.

Conversation System
In addition to the dialect and subsequent sentence generation system
outlined above, URR  also deploys a particular kind of conversation
system. The player chooses from dozens of topics— artworks, history, ide-
ologies, religions, monasteries, travel, animals, and books— and the list
of potential questions in each topic grows as the player learns more about
those topics.
Upon learning about a new artistic movement in poetry, any NPC can
be asked about that movement; upon learning about a particularly bloody
war fought 80 years ago, any NPC can be asked about that war; or upon
learning about a reclusive hermit who might hold a crucial clue to the
player’ s victory, any NPC can be asked about the location and disposition
of that hermit. Crucially, of course, none of these elements were fixed in
world generation— all these conversation options have emerged through
the game’ s PCG.
The conversation system is therefore designed specifically to take
advantage of the fact that these are AI actors within a procedurally gener-
ated game. URR  adapts immediately to whatever new output of the proce-
dural generators has been created and identified by the player, whether it’ s
a painting, a person, or a piece of history, and allows the player to start a
conversation about an effectively infinite number of topics.
Similarly, while sentences are constructed from many elements, an AI
actor is always able to put together an appropriate reply to any topic asked.
This yields a PCG response to a PCG question, even if that response is
only to express a lack of interest because you are asking them about a piece
 Artificial Intelligence    ◾    141

of artwork they’ ve never heard of, a moment of history from a thousand

years ago and half a world away, and so forth.
Conversations with AI actors are core to the gameplay of URR , even
though everything to do with conversations— from sentences and dialects
to questions and replies— is procedurally generated from scratch each
time the game is run. This represents an intriguing fusion of PCG and
AI behavior, and one I plan to develop further in the coming months and
years.

CONCLUSION
Whether your actor navigates terrain, attacks, or sociopolitical debates,
their ability to detect and adapt to changing circumstances is key to their
satisfying performance in a procedurally generated game. The question is
often not how many verbs the AI needs, but rather what types and com-
binations of input it can react to, and what elements of its reaction can
be generated and will yield interesting outcomes. When trying to make a
world that feels alive and unique, PCG developers must consider the val-
ues they want their world to express, and move from this into developing
their characters’  strategies, speech, and desires.
 Chapter 14

Procedural Enemy Waves

Wyatt Cheng
Blizzard Entertainment

A common practice in player versus environment (PvE) games is to

have situations where many enemies are set up to attack the player as
the fight progresses. Sometimes, there are large waves that are distinct and
obvious. Other times, enemies come more in the form of “adds”—indi-
vidual additional enemies who join a battle in progress. Sometimes this
will be a triggered event; sometimes entire levels are created based on the
notion of fighting multiple waves.
Examples of procedural wave spawning in Blizzard games include

• Starcraft 2: Wings of Liberty: In the mission “Mar Sara 3: Zero Hour,”

you must survive for 20 minutes against incoming attacks. Missions
where you have to survive for a minimum amount of time are called
“holdout” missions.
• World of Warcraft: Many bosses will spawn monsters in waves. Gluth,
one of the bosses in Naxxramas, spawns zombies every 10 seconds.
• Diablo III: There are events called “Cursed Chests.” Cursed Chest
events always last exactly one minute. Monsters spawn continuously,
and the faster you kill the monsters, the faster they spawn. If you kill
at least 100 monsters, then you are granted additional reward.

As you can see, spawning waves can be used in a variety of contexts.

Here I’ll cover four different methods that have different advantages and
disadvantages. You can adapt these methods to suit your needs.

143
144   ◾    Procedural Generation in Game Design

METHOD 1: SPAWN BY TIMER

When using Spawn by Timer, the enemies are spawned based on elapsed
time, regardless of the situation and player state. A single enemy can
spawn at a fixed time, or you can spawn many waves (multiple enemies
who all appear at the same time).

Spawn by Timer Pseudocode

// Waves is sets of data pairs that represent the

// time into the mission to spawn a wave
// and the data required to spawn that wave.
Waves = {
{time=15, spawndata=SpawnData1},
{time=30, spawndata=SpawnData2},
{time=45, spawndata=SpawnData3},
{time=60, spawndata=SpawnData4}
}

// A function that is assumed to run every

// game tick (for example, every 60th of a second)
Global currentWave = 0;
Function UpdateTick()
{
// If we are ready to spawn the next wave
If (Waves[currentWave].time <
GetCurrentGameTime())
{
// Spawn the next wave
SpawnWave(Waves[currentWave].spawndata);
currentWave = currentWave + 1;
}
}

Commentary
Spawn by Timer is a good spawn technique when you want to test the
player’s skill, aptitude, or general power level. The enemies appear whether
the player is ready or not, making this a good technique to use when the
approximate power level of the player is fixed and relatively the same for
everyone. Factors such as level-ups, gear, or player skill can cause the
effectiveness of the player to vary considerably, making Spawn by Timer
quickly lose its appeal.
 Procedural Enemy Waves    ◾    145

Levels that use Spawn by Timer are often crafted for a specific diffi-
culty by a designer—if the player is a bit stronger, each wave becomes too
easy and the player can be left bored. If the player is a bit weaker than the
designer anticipated, then the player quickly becomes overwhelmed.
This latter problem is extremely common when using Spawn by Timer.
Suppose you have waves of five monsters. Based on the designer’s playtest-
ing, it takes 15 seconds to defeat a wave, so the waves are set to be 15 sec-
onds apart. If the player is even 20% weaker than the designer anticipated,
then each wave might take 18–19 seconds to defeat. Just 15 seconds into
the fight the player might still have one enemy at full health when wave
two spawns. This combination of enemies might be so challenging to the
player that they have three monsters alive when wave three spawns. This
effect can “snowball,” causing the player to become farther and farther
behind the intended difficulty until completely overwhelmed.
Whether this snowball effect and sensitivity to player power can be
beneficial or detrimental depends on the design goals of the level. If the
strength of the player character is well known and it is the designer’s intent
to test whether the player has reached a certain level of mechanical profi-
ciency with the game’s controls and strategy, then a Spawn by Timer event
would be very appropriate. On the other hand, if the goal is to provide a
cool experience for the player regardless of their gear or skill level, then
the snowball effect is inappropriate and a different spawning technique
should be used instead.
The Starcraft holdout missions are a good example of Spawn by Timer.
On a typical Starcraft level, the power of the player is relatively well known
and the goal is to test the player’s ability to play strategically in order to
win or lose.
Some players may be so quick and adept at killing the enemies of each
wave that there is a possibility they clear a wave before the next wave is
ready to activate. In this situation, the designer has two choices:

1. Delay until the timer elapses, allowing the player a breather. This
approach not only is simpler to implement but also rewards the
player for finishing early. However, this approach also risks disturb-
ing the pacing or boring the player, as they will have no enemies to
fight until the next wave is spawned.
2. Set up your level so that the next wave triggers on a timer, but
should the player clear the wave early, then the next wave triggers
146   ◾    Procedural Generation in Game Design

immediately. This becomes a blend of this method with Spawn on

Completion, which we’ll discuss next. In this latter case, you should
be cognizant of any player advantages to dragging out the fight (such
as health and resource regeneration or cooldown activation).

Key Characteristics
• Good if you want to test whether the player has mastered certain
game mechanics
• Sensitive to the player’s current power level
• Can lead to snowballing
• Allows precisely designed pacing

METHOD 2: SPAWN ON COMPLETION

Spawn on Completion is a spawn technique that involves spawning waves
of enemies where each wave is triggered by the previous wave dying.

Spawn on Completion Pseudocode

// Waves is simple data that contains the

// information required to spawn each wave.
Waves = {
{spawndata=SpawnData1},
{spawndata=SpawnData2},
{spawndata=SpawnData3},
{spawndata=SpawnData4}
}

// A function that is assumed to run every

// game tick (for example, every 60th of a second)
Global currentWave = 0;
Function UpdateTick()
{
// If we are ready to spawn the next wave
If (NoEnemiesAreAlive())
{
// Spawn the next wave
SpawnWave(Waves[currentWave].spawndata);
currentWave = currentWave + 1;
}
}
 Procedural Enemy Waves    ◾    147

Commentary
Spawn on Completion presents some amount of challenge to the player,
but with some flexibility to adapt to how well the player is doing. The dif-
ficulty of each wave can be calibrated by the designer without the difficulty
of one wave affecting the others. Unlike Spawn by Timer, if the player
struggles with one of the waves (but manages to complete it), it won’t have
a disproportionately adverse effect on subsequent waves.
Spawn on Completion is used in the World of Warcraft raid zone “The
Battle for Mount Hyjal,” in which players repeatedly confront 10 waves
followed by a boss fight. The approach allows a mix between testing play-
ers for a minimum gear or skill level and still keeping players engaged the
entire time.
Since the difficulty of each wave does not directly affect the challenge
of other waves, Spawn on Completion can also be a way to test the power
of the player when there are multiple vectors on which to measure player
power. For example, if a spaceship game allows players to excel with mis-
siles, laser beams, and shields, then three different waves might test the
player’s abilities along these three different vectors. If the player is strong
or weak against one vector, then the player will find that wave to be par-
ticularly difficult or easy; however, the player still needs to demonstrate
sufficient proficiency along each vector to complete the level as a whole.
One cautionary note: Waiting for the previous wave to be completed
before the next wave spawns can sometimes lead to a behavior in which
the player deliberately avoids killing the last monster to “drag it out” and
allow health, resources, and cooldowns to recover. This can be mitigated
somewhat (but not entirely) by combining this spawn technique with a
Spawn by Timer and allowing either condition to cause the next wave to
spawn. Doing so introduces the risk of introducing the snowball effect
that Spawn by Timer is susceptible to, but the two techniques are very
often combined successfully.
One often desirable characteristic of Spawn on Completion is that it is
readily apparent to the player that the enemies are coming in waves. This
can be very exciting, as it allows alternating periods of tension and release
as the waves come in.

Key Characteristics
• Good if you want to be somewhat adaptive to how powerful the
player is but still want to have minimum performance requirements
148   ◾    Procedural Generation in Game Design

• Can be exciting to alternate tension and release

• Good for testing along multiple vectors

METHOD 3: CONTINUOUSLY ESCALATING TOTAL

With a Continuously Escalating Total spawning scheme, the number of
enemies active at a time is always increasing. As enemies are killed by
the player, new enemies spawn to maintain a minimum count of active
enemies.

Continuously Escalating Total Pseudocode

// Waves is a set of data pairs that represent the

// time into the mission and the number of enemies
// that should be onscreen at that time
Waves = {
{time=15, minimumEnemyCount=5},
{time=30, minimumEnemyCount=10},
{time=45, minimumEnemyCount=15},
{time=60, minimumEnemyCount=20}
}

// A function that is assumed to run every

// game tick (for example, every 60th of a second)
Global currentWave = 0;
Function UpdateTick()
{
int minEnemyCount;
For each (wave in Waves)
{
If (wave.time > GetCurrentGameTime())
{
minEnemyCount = wave.

minimumEnemyCount;
}
}

While (NumberOfEnemiesAlive() < minEnemyCount)

{
SpawnAnEnemy();
}
}
 Procedural Enemy Waves    ◾    149

Commentary
Continuously Escalating Total is a good technique to use when a dramatic
arc is desired. Because enemies are continuously replacing themselves, the
fight needs to end based on a criterion other than “defeating all the ene-
mies.” The most common is reaching a kill total or a time limit.
With aggressive tuning values, Continuously Escalating Total can be a
fun spawning technique to use for a bonus level or other minievent. Since
the number of active enemies is continuously increasing, the player never
truly “beats” the event. In this context, the event instead becomes a test to
measure how far the player can get.
In many games, gauging the player’s power level is difficult due to
the number of factors that contribute to player success. This makes
Continuously Escalating Total a good spawning technique to use when
the player’s power level may vary widely due to either play skill or charac-
ter progression. Early on, the player might play a Continuously Escalating
Total level and find they could survive 30 seconds, and then later in the
game, when they have grown in skill or character progression, they are
able to play the same level but survive for two minutes. This provides a
tangible benchmark of how far they have come.
A variant of Continuously Escalating Total is used in Diablo III for the
Cursed Chest events. The goal of the Cursed Chest events is not for the
player to either defeat the enemies or die trying. Rather, the design intent
is for the player to always kill enemies, but for the event to be a test of how
many enemies the player can defeat. The player has exactly one minute
to kill as many enemies as possible, and the faster the player can kill the
enemies, the faster they spawn.
A cautionary note: Depending on the exact mechanics of the game and
the presentation of the results, overreliance on Continuously Escalating
Total can become exhausting for the player. Constantly growing in inten-
sity until you are ultimately overwhelmed can be exciting, but as designers,
we should be cautious when we require players to maintain this excite-
ment cycle for extended play periods.
An interesting property of Continuously Escalating Total is the rela-
tionship between player power and how long the event takes based on the
completion criteria:

• If the event is timed, more powerful players cannot complete the

level faster.
150   ◾    Procedural Generation in Game Design

• If the event is based on “kill X enemies to complete,” then the level

will complete faster as the player becomes more powerful.
• If the event is based on “how far can you get,” then the level actually
takes longer to complete as you improve in gear.

Choosing the right completion criteria based on the needs of your game
is an important decision for effective use of this spawning technique.

Key Characteristics
• Adapts to being fun and challenging for both low-end and high-end
players
• Good at allowing players to measure themselves along a continuous
spectrum of progression rather than being a binary pass–fail check
• Can be mentally exhausting to do for a long period of time

METHOD 4: HITPOINT PROGRESSION

With Hitpoint Progression, additional waves are spawned when damage
equal to a specific percentage of maximum health has been done to the
enemies spawned in the previous wave calculated against the sum total of
health. This means that if the player is less powerful, then spawning will
happen slower, and if the player is more powerful, then spawning will hap-
pen faster. The action is continuous, since the next wave is spawned while
the current wave is still active.

Hitpoint Progression Pseudocode

// Waves is sets of data pairs that represent the

// hitpoint percentage of the previous wave that
// should remain before spawning the next wave paired
// with the data required to spawn that wave.
Waves = {
{percent=1.00, spawndata=SpawnData1},
{percent=0.60, spawndata=SpawnData2},
{percent=0.30, spawndata=SpawnData3},
{percent=0.50, spawndata=SpawnData4}
}

// A function that is assumed to run every

// game tick (for example, every 60th of a second)
 Procedural Enemy Waves    ◾    151

Global currentWave = 0;
Function UpdateTick()
{
CurrentHealth = SumOfHealthOnCurrentWave();
MaxHealth = SumOfMaximumHealthOnCurrentWave();
HealthPercentage = CurrentHealth / MaxHealth;

// If we are ready to spawn the next wave

If(Waves[currentWave].percent >
HealthPercentage)
{
// Spawn the next wave
SpawnWave(Waves[currentWave].spawndata);
currentWave = currentWave + 1;
}
}

Commentary
Hitpoint Progression is a good system to use when you want the player
to experience a continuous stream of enemies that is always presenting
a challenge but has some ability to adapt to how well the player is doing.
Similar to Spawn on Completion, Hitpoint Progression gives the designer
the ability to create some waves that are particularly challenging or the-
matic, or stress different vectors of player power while ensuring the player
doesn’t get overwhelmed if they are below the designer’s expected power
level.
If the player is doing well, enemies spawn faster; if the player isn’t doing
well, enemies spawn slower. For this reason, Hitpoint Progression is a
good system to use when a high variance in player power is expected due
to skill or character progression and the designer wishes to create a dra-
matic encounter that is likely to be fun and engaging for the widest pos-
sible audience. By the same token, Hitpoint Progression is a bad spawn
technique to use when you wish to test player performance.

Key Characteristics
• The most adaptive of the four spawning techniques
• Not good for measuring player power or skill
• Good for creating a dramatic combat experience
152   ◾    Procedural Generation in Game Design

CONCLUSION
Changing even what seems like the smallest element of your procedural
algorithm can result in highly varied player experiences. Timers, comple-
tion, continuous escalation, and hitpoint progression are only four of the
many possibilities of spawning enemies. The correct spawning technique
or combination of techniques can be selected on grounds of intensity,
challenge, pacing, exploitation, balance, or other factors, depending on
your gameplay’s unique needs.
 Chapter 15

Generative Artwork
Loren Schmidt
Independent 

I n the early 1950s, Desmond Paul Henry found a World War II bomb
guidance computer at a military surplus store. Fascinated by the move-
ment of the analogue computer’ s mechanical components, he altered it
in such a way that instead of expressing its processes as bomb-targeting
information, it drew pictures. Its inner workings were superficially sim-
ple, on an algorithmic level not dissimilar from pendulum-based drawing
machines. But Desmond Paul Henry’ s drawing machines have made some
of the most deeply beautiful human computer art. And they achieve all
this with great lightness and elegance.
Henry’ s machines draw with energetic, organic lines: there is a slight
flutter to the weight, the edges undulate gently like bone ridges or rippling
skin. They feel deeply appealing on a human level.
Much of today’ s digital art (especially if you work in games) has a fixa-
tion on fidelity and technical showiness. It is hard not to think this way:
beyond the immediate appeal, visual fidelity is often seen as an indicator
of artistic merit. This is a complicated topic, and one that I don’ t have
space here to dissect properly. Let it suffice to say that

1. Today’ s particular brand of “ technical excellence”  is tremendously

dependent on funding, available time, education, and other resources.
2. Cultural emphasis on this type of work is unnecessarily discourag-
ing to people who don’ t have those resources.
3. This is not an interesting or useful metric for evaluating art.

153
154   ◾    Procedural Generation in Game Design

Your art is valuable and valid no matter where you are coming from.
This section aims to arm you not with new technical tools, but with gen-
eral techniques and modes of thinking that are useful in addressing the
human side of generative visual art. It is designed to help you quickly apply
these concepts to any generative work (including games), so please apply
them wherever they are useful in your own life.

TECHNIQUES
Scope is not quality; it is a separate axis : When finishing up or fine-tuning
a project, one might be tempted to try and “ improve”  it by adding fea-
tures (Figure  15.1). The generated tree gets tiny animated flowers; the sky
gets heat haze and atmospheric scattering. This may seem natural, but it is
not: making your project flashier and adding technical tricks is increasing
scope, not quality. Instead, try taking a step back. What does your piece
feel like, and how does that differ from how you want it to feel? Are your
proportions good? Could you tune it better? What is the weakest part of
the piece? Could it be improved by removing anything?
Thinking in terms of processes : We are usually taught to evaluate algo-
rithms in terms of whether they work. But I’ d argue that there is another
more important metric here that largely goes neglected. Are your algo-
rithms emotionally relevant to your project? Code is a rich and complex
language, and there are many different ways to write the same process.
Your individual decisions are a powerful expressive conduit. Also be
aware that sometimes dissonance or distance can subtly insert itself when
the algorithms do not feel the way you thought they would, when the tools
you are using are not in harmony with your goals. If you find this happen-
ing, can you alter your algorithms or your goals so they are better in tune?
Symmetry and structure : Working digitally, one of your great strengths
is the way iterative processes naturally tap into human perception. Humans
have a strong natural response to symmetry, and engage automatically in
pattern-detecting behavior. A few shapes with some symmetry applied
can quickly start to look compelling (Figure  15.2).
Mark making : Consider the ways you are making marks. This might
mean drawing pixels to a screen, rasterizing vectors, or drawing poly-
gons in three-dimensional (3D) space (Figure  15.3). This might mean
adding letters to a string. Certain tools have a very distinct fingerprint:
for instance (as of 2016), much visual work done in JavaScript or Unity
is highly recognizable as such. Perhaps this esthetic cast works in your
favor. But if you find yourself fighting against it or wanting to try out other
 Generative Artwork    ◾    155

FIGURE  15.1   Generated sigils.

FIGURE  15.2   Generated flowers.

156   ◾    Procedural Generation in Game Design

FIGURE  15.3   Generated triangle.

possibilities, consider how your marks are made. Perhaps you can use the
built-in functions in an unconventional way, and come up with something
that works better for you? Similarly, experimenting with writing your own
low-level drawing routines can open up a wide range of mark-making pos-
sibilities not available in tools by default (this can be a technically or com-
putationally intensive thing to do).
Postprocessing : Postprocessing  can mean a number of things, but here
let us consider it to mean altering your output after it has been created.
Conventionally, there is little breadth in these techniques. In the realm
of real-time 3D, for instance, postprocessing is largely limited to a few
standbys: vignettes, motion blur, color correction, and screen space ambi-
ent occlusion. But even built-in tools often have a lot of untapped breadth.
And if you write your own postprocesses, there are many other expressive
options available. For instance, if you modify diffusion and pattern-dith-
ering algorithms such as Atkinson or Floyd– Steinberg by feeding them
unconventional diffusion kernels, you can create a wide range of effects,
such as the one depicted in Figure  15.4. It’ s also worth noting that there
need not be a rigid boundary between creation and post effects. There is a
lot of interesting potential for feedback between the two.

PERCEPTION OF INTENT
Searching for the intent of the creator is a natural human way to interact
with any artifact. One may ask: Who made this? What was their process
like? What were they feeling? Sometimes people deliberately turn this capac-
ity off or deemphasize it when looking at digital art. As mentioned above,
humans are avid pattern detectors. Repetition and pattern can be beautiful.
Sometimes one might make an elegant, crystalline algorithm and wish to
 Generative Artwork    ◾    157

FIGURE  15.4   Generated staircase.

expose its workings and bounds. Sometimes transparency can be desirable

here. In other situations, however, one might find oneself struggling against
this tendency to create a process that feels more varied and intentional.
If the volume of the solid is insufficient, one can linearly increase one
of its dimensions. This happens frequently when working with gram-
mars. Suppose one is populating a town with buildings. The obvious way
to reduce repetition is by linearly increasing the number of buildings in
one’ s set. But this is a linear increase, and does nothing to address struc-
tural blindnesses or faults of pattern. When looking at the town, we see
only randomly placed diverse buildings. It’ s still noise.
Instead of simply pursuing linear growth, one can pursue greater self-
awareness. In the case of our town, we might benefit from algorithms that
158   ◾    Procedural Generation in Game Design

create different types of multibuilding complexes. Like increasing build-

ing variety, this can decrease the repetition of outputs. But more signifi-
cantly, this begins to address the problem of perceived blindness or lack
of intent. Suddenly, when looking at the town one starts to pick out walled
courtyards, parks, and patterns.
On a higher level, one can make an abstract plan that serves as a guide.
For instance, we could make a process that generates plans for towns.
Which feelings do different areas have? What motivates neighborhoods,
the distribution of different varieties of buildings? If we start here with
a high-level concept, then our algorithms for complexes and individual
buildings can be created with awareness of the plan. When viewing the
town, we might find factory complexes along the water, agriculture centers
away from the town’ s center, and markets at neighborhood centers. When
we look at the town, we now have both variation and a sense of intent.

 PROCEDURAL ART CASE STUDY

MARK MAYERS 

Desolus  is a surrealist puzzle game where you explore dreamlike landscapes

and power mysterious technology with solar energy. My background as a
computer scientist allowed me to create Desolus  using procedural art, despite
having no traditional art training (Figure  15.5).
Art assets in Desolus , such as terrain, are first roughly created by hand,
with algorithms like erosion simulation providing refinement and detail. By
manually creating a tree species and then manipulating its generation, unique
trees are able to be grown for the game’ s environments as needed using
SpeedTree. These elements allow for more diversity in assets than could be
developed traditionally, especially by a small team (or, in this case, a solo
developer).
Despite its resplendent appearance, Desolus  contains no rigged charac-
ters or animated models. Instead, the game uses a detailed graphics process-
ing unit (GPU) particle effect system to convey gameplay elements abstractly.
Traditional animations are replaced with particles and shaders, allowing for
more effort to be spent designing compelling worlds. For example, the black
hole in Desolus , which comprises the game’ s core puzzle mechanic, is gen-
erated with a gravitational lens shader and GPU particle system. Tweaking
various parameters of this particle system, such as colors, physics forces,
emission, and shaders, produces a variety of impactful results in various
scenarios. A similar methodology is used for creating the sky of Desolus .
Shaders controlling color and atmospheric properties are manipulated to
 Generative Artwork    ◾    159

FIGURE  15.5   Generated landscape from Desolus .

create a dynamic sky system, which is critical for setting the mood of scenes.
Nonorganic meshes in the scene are generated through fractals and proce-
dural meshes. For instance, the black obelisks present in every scene of the
game are created by recursively cloning a simple cube in a pattern.
These individual systems combine to compose the “ universe”  of Desolus ,
a complex world created from simple algorithmic rules. This process of cre-
ation emulates the natural world, as the universe operates by a set of laws
that humans can ostensibly describe through science. However, it is the art-
ist who interprets what he or she finds most beautiful about these universal
laws. 
 Chapter 16

Generative Art Toys

Kate Compton
Independent 

E veryone likes being creative. And everyone likes discovering that

they’ re more creative than they thought they were. For many years,
people have enjoyed crafts like pottery wheels, Spirographs, Mad Libs,
spin art, paper marbling, and tie-dye. These artistic toys helped everyday
people make interesting artworks (even if those people lacked creative
talent or inspiration) by producing surprising and emergent results from
simple choices.
Now that we have digital systems, we can make art toys with even more
surprising and emergent behavior. No longer bound by the limits of real-
ity, we can build Spirographs in virtual reality (VR), Mad Libs that gener-
ate worlds, tie-dye that evolves over time, and pottery wheels that spin out
curling tendrils of emergent geometry.
I’ ve enjoyed building generative digital art toys, and they’ ve helped me
better understand and use generative methods. I’ ve written this chapter as
a guide to help you to design and build your own. May you find as much
pleasure and fun and strange generative artwork as I have!

BUILDING ART TOYS, FOR EXPERTS AND NOVICES

So why is this chapter on art toys in a book on procedural content for
games? Because generative art toys can use many of the same techniques
as procedural generation (PCG), but don’ t have the same gameplay-based
restrictions as content that has to go into a game. As someone who makes
both generative art toys and PCG for games, I find I have a lot more

161
162   ◾    Procedural Generation in Game Design

freedom making the art toys, and I spend more time enjoying the wildness
of the generativity and less time worried about gameplay issues.
Art toys are also a good way to explore new kinds of generative meth-
ods (algorithms suitable for generativity), or to try out old favorites in a
new context. I’ ve repurposed the particle-based techniques we used to cre-
ate Spore planets to also make interactive drawing tools that come to life.
Lindenmayer systems (L-systems) can create nice trees for game back-
grounds, but they can also be used to make surreal flowers for a flower-
arranging app (Figure  16.1). Context-free grammars can construct dungeon
levels and Mario platforms, or can build music and poetry. Flocking algo-
rithms can move nonplayer characters (NPCs) for combat, or can power a
choreography generator. I’ ve written a popular blog post on different gen-
erative methods you might want to try, “ So You Want to Build a Generator,” 
but I can’ t think of a single algorithm that couldn’ t be turned to making art,
so keep your eyes open for new potential procedural paintbrushes.
One of the reasons I love making art toys with procedural content
is that, in games, the generated content is often invisible (like artificial
intelligence [AI] movement) or in the background (like sky textures and
trees), but it’ s always serving the purposes of the gameplay. The gameplay
is always the star, and the PCG plays a supporting role. What if the user
were interacting directly with the content generator? What if the user were
controlling the generator, giving it feedback, or exploring and curating the

FIGURE  16.1   Art toy by Kate Compton, where the user iteratively breeds new
procedural flowers, generated with an L-system and an evolutionary algorithm.
 Generative Art Toys    ◾    163

things it made? Suddenly, the focus is on the generator itself; now it’ s the
star of the show and the center of attention! As someone who loves build-
ing generators, this is a pretty delightful situation.

EXPERIENCE OF ART TOYS

Digital art toys (and the physical art toys of the past) are beloved because
they help us discover the artists in ourselves, even when we aren’ t particu-
larly good. We might not all secretly be great artists on the inside, but we can
surprise ourselves by creating a neat mandala, abstract artwork, chiptune
melody, generated poem, procedural flower, decorated pot, animated crea-
ture, or any of the many odd and interesting artifacts that you can now make
with easy-to-use digital tools. The tool needs to be smarter than a pencil, but
also let the user feel like they’ re in charge. How can you strike that balance?
Art toys should provide a good experience for users. We’ re familiar
with the kinds of pleasures that games provide (see 8 Kinds of Fun  or
Koster’ s Theory of Fun  for nonexhaustive lists), but creativity apps and art
toys have their own different kinds of pleasures. I’ ve come up with several
modes that an art toy might operate in, depending on how much inde-
pendence it has outside the user’ s interactions: it can feel like a tool, like a
collaborator, or like an artist itself.
If it is a tool, it can magnify or constrain the user’ s input so that the
output becomes bigger, more complex, and better made than the actual
skill and effort they put in. As a collaborator, the art toy can provide inspi-
ration, suggest new directions and modifications, and encourage a user
who might otherwise feel uninspired and undirected. Sometimes the gen-
erator is independent enough to create content on its own, and the user
can become the one providing guidance, steering the generator to explore
different parts of the possibility space, as the user collects and curates and
shares the subset of generated artifacts that they find particularly interest-
ing. Sometimes we even give up understanding how to control the system
at all, in exchange for the fun of chaotically surfing on a generative wave
of art. Try Become a Great Artist in Just 10 Seconds  by Michael Brough and
Andi McClure for a grand example of this.
In any of these cases, the user should feel like they are exploring a
rich and varied space of possible artifacts, quickly and pleasurably. They
should also feel like they, personally, have ownership of the artifacts that
the generator creates, whether they had the skill to build them (with the
generator’ s assistance) or the good taste to pick them out (from the genera-
tor’ s output).
164   ◾    Procedural Generation in Game Design

A user who uses a well-made art tool will often forget about the tool’ s role
in creation altogether, which is not a bad outcome! Saying “ I’ m really good
at this” is a much more powerful feeling than “ this generator is good at this.” 
Conveniently, a user who takes to social media to share their skill at a creativ-
ity is also symbiotically promoting the app, too! Call it the Minecraft  effect:
when the users feel creative and powerful, both the users and the apps win.

TRADING CONTROL FOR POWER

The central idea behind an art toy is that the user is trading away control
and gaining power. What that means (and why it’ s important) becomes
clear if we contrast art toys  with art tools .
There are countless professional digital art tools available today. Very soon
after computers developed, people created drawing software for them (Ivan
Sutherland’ s 1963 Sketchpad ), and this practice has continued ever since.
Many of these tools were created with expert users in mind— professional cre-
atives or skilled artists— and give users total control over the output. This bril-
liant user must have an idea of the final image already in their head; the tool’ s
purpose is to help them render that idea to the screen. Not a simple demand!
But there are more users who want to be creative than just these few creative
geniuses (and professional creatives). Creativity is linked to numerous positive
outcomes in health and psychological wellness. Being creative is a pleasurable
activity: we feel good when being creative. Interestingly, these benefits aren’ t
correlated to how good we are at the creative activity, just whether or not we
feel engaged. This means that this kind of creativity is an autotelic activity;
that is, it’ s practiced as an end in itself, rather than to get some external result.
Professional creativity isn’ t necessarily an autotelic activity. Photoshop,
for example, assumes that its users are professionals trying to get some-
thing specific done. So it is highly flexible and highly complex, and has to
give the user total control. Likewise, three-dimensional (3D) software like
Maya doesn’ t know whether the user is making a 3D printed teapot, an
architectural rendering, or a 3D cut scene, so it has to support all three,
but can’ t provide automated support for any given task. Automated gener-
ators aren’ t customizable enough; professional users require total control.
However, autotelic users are willing to cede control to a generator, as long
as it gives them more power; that is, the generator can magnify or super-
power streamline their creative activity in some way.
There are many digital examples of this trade-off (e.g., Spore vs. Maya
and Instagram vs. Photoshop). But even older childhood toys like the
Spirograph show this trade-off between power and control.
 Generative Art Toys    ◾    165

An unconstrained drawing tool like a pen gives an artist complete con-

trol, able to draw in any direction. “ Line drawing”  encompasses a huge
possibility space, containing some Rembrandts, sure, but mostly ugly and
incoherent scribbles. A novice user will more likely end up with a scribble
than the Rembrandt; they can draw anything, but aren’ t likely to find their
way to the good stuff.
With Spirograph, a gear-based drawing toy from the 1950s, the pen
is stuck in a socket in one of the gears, and the gear forces its motion
into spiraling loops as the gear spins around its track. The user has only
a few discrete decisions to make (color of pen and size of gears), and one
dimension of motion (forward or backward). This creates much smaller
possibility space, but that space is full of universally interesting stuff! No
Rembrandts, but a lot more satisfyingly symmetrical geometry mandalas.
Pottery wheels make a similar trade-off in one more dimension. From
the huge three dimensions of freedom (and accompanying possibility
space) of a hand-sculpted shape, one dimension is automated by the wheel,
leaving the two dimensions of freedom (in– out and up– down) to work
the pot. The automation has taken away a bit of freedom, but has given us
a faster, more reliable way to get to satisfying outcomes. We’ ve given up
control and gotten superpowers in exchange.

KASPAROV’ S CENTAUR 
Isaac Karth 

Today, new AI and generative tools are creating art. A neural network can
transfer the style of one image onto another or change the composition. A
photograph can be made to look like it was painted by Van Gogh or Renoir.
This isn’ t the first time artists have dealt with new technology. Generative
tools have strong parallels with early photography. As photography strug-
gled to establish itself as a medium in its own right, some painters reacted
by embracing Impressionism. Impressionism took inspiration from the more
candid subject matter that photography allowed, but it also embraced the
subjective visual aspects that photos had trouble reproducing.
Likewise, in the wake of his historic loss to IBM’ s Deep Blue, chess grand-
master Garry Kasparov became interested in a new way to play chess. In
1998, he played the first game of “ advanced chess”  against Veselin Topalov:
each player used a computer as a partner.
This cyborg play style, sometimes called “ centaur chess,”  created bet-
ter players than individual humans or machines. Like the centaur, the
human– machine collaboration lets each half of the combination overcome
166   ◾    Procedural Generation in Game Design

the weaknesses of the other. The human is still playing the game, but the rote
work has been automated, freeing the player to express themselves.
An artistic metaphor for this is the generative artist Strangethink’ s Joy
Exhibition . Joy Exhibition  is a little VR art gallery, where the only painting
tool is a collection of billions of procedurally generated paint guns. Each gun
has a distinct and weird pattern. These paint guns are a microcosm of PCG
and centaur art.
Interestingly, impressionist paintings seem to work particularly well for
style transfer. Something about the loose brushstrokes appears to make it
easier for the machine to generate a cohesive result. The future of game and
art generation likely has similar discoveries lurking, waiting for the right artist
to put the pieces together. 

DESIGN AND CONSTRUCTION OF ART TOYS

Hopefully I’ ve convinced you to build an art toy, if you weren’ t already
planning to do so.
If you were designing a game, you’ d probably start with a type of game-
play (e.g., tower defense or match 3). But there aren’ t established genres yet
in art toys, so each one seems to be a unique design solution. So how do you
begin designing one from scratch? Figure  16.2 shows the common design
pattern that I’ ve seen from many of the casual creators and art toys I’ ve stud-
ied. Each art toy has three kinds of elements (inputs, transformations, and
rendering), with an optional fourth element of social or sharing mechanics.
You can think of these as modular Lego bricks that can be snapped together,
creating a different art toy in each configuration. I’ ve made this into a set of
physical cards that you can print out and experiment with at home.*

Inputs
Inputs can be anything from simple keyboard input or scroll wheels, to
touchscreen and mouse-dragging or tablet pens, to full body tracking
with a Kinect or 3D tracked VR hand controllers.
One of the important considerations here is to consider the “ width”  of
the data pipeline. A keyboard produces very sparse data: binary keypress
events, spaced out over several seconds. A mouse provides a continuous
feed of X , Y  coordinates, two dimensions of data recorded continuously
over time. A hand tracker like the Leap Motion records the joint positions
and rotations of wrists, hands, and fingers. If wrists move freely in 3D

* http://www.galaxykate.com/arttoys/arttoy-cards.pdf
 Generative Art Toys    ◾    167

input
input drawings on input output
global wind speed persistent kinect projection on/in
surface living being

outputModifier
inputModifier output input
interactors and
viewers aren’t the on a dress at a projection on face tracking
same species fashion show very small
screen

transformation output transformation inputModifier

simulate braitenberg LED brightness Stipplegen embedded in a tree
vehicles point distribution

FIGURE  16.2   Example card from the printable set.

space, and each finger joint has about one-and-a-half degrees of freedom
(rotation plus a bit of side-to-side motion), a hand produces 50– 100 data
points per sample, a huge pipeline of information to process!
Very high degrees of freedom (like skeleton tracking) require con-
tinuous feedback to the user to help them understand the system. The
Treachery of Sanctuary , a triptych art installation by Chris Milk that
uses the user’ s shadow as input, does a spectacular job at this. As the
user raises their arms, they see their fingers’  shadows extend into large
feathered wings, and lowering the arms retracts the wings. As they flex
their fingers in and out, they see the feathers fan in and out at the same
rate. This continuity (temporal and physical) of the generated shadow
gives a great stream of feedback to the user, and the gallery patrons, even
children, instantly understand and begin playing with their shadow.
Contrast this to many other body-tracking apps: while Treachery of
Sanctuary  uses the entire continuous pipeline of input to generate the
shadows, other apps try to detect “ pose”  occurrences, compressing the
rich pipeline of data into a series of discrete binary events spaced out over
time: the same spacing and discreteness as keyboard input. Good user
interface (UI) feedback is often an issue with these apps, perhaps because
they’ ve discarded 99.99% of the user’ s input before they start. But genera-
tivity like that shown in Treachery of Sanctuary  suggests that this need not
be the case.
168   ◾    Procedural Generation in Game Design

Data and Transformations

So you have some input. What now? That real input  (the motion of a hand,
a body, a crowd, or stock market data) has been turned into some data
format that represents it: a series of points, a set of curves, a grayscale
image, or a number. An art toy takes that input data and can either render
it directly to the screen or transform it to different kinds of data through
many different algorithms.
For a sense of how many different abstract transformations you can
build with a single kind of starting input, let’ s consider hand tracking. As
noted above, raw input from hand tracking returns dozens of data points
per frame, hundreds per second. If we wanted to build an art toy with
hand data, how should we interpret it? The next sections are a sample of
many ways that this data could be interpreted and modeled without dis-
carding information (like traditional gesture recognition does).

Points and Rotations

Leap Motion returns the hands as several arrays of positions (for each fin-
ger, the wrist, etc.). Freeze the hand for a moment in time and look at the
positions of the joints. Each one has a location in space, and each location
might be in front or behind some other geometry (an imaginary plane of
canvas). There may be moments (and positions) where the points cross a
boundary (when the fingers “ pass through”  the canvas). Each finger joint
also has a 3D vector for the direction that it is pointing. With these two
arrays, we could build geometry aligned along the bones, but extending
further outward, as Treachery of Sanctuary  does.

Connectivity and Meshes

From a set of points, we can build a connectivity graph. Between each edge
of the finger is an edge representing a finger bone. But there may also be
edges between fingertips showing their proximity (with edges appearing
and disappearing as fingertips near and touch). If both hands are visible,
we can build a Voronoi map of connectivity, a procedurally generated cat’ s
cradle of edges.

Gestural Curves
Hands are rarely still. They wave and flex, clench and open. Each joint
moves along an invisible animation curve through space, and we can
record those curves as well. Google’ s Tilt Brush  takes this approach,
recording the curves of motion and then reskinning them with different
 Generative Art Toys    ◾    169

strokes or using them as paths for additional actions (see the “ Rendering” 
section later).

Forces and Acceleration

Each joint is moving. From that, we can calculate the velocity (as a 3D
vector) and thus determine its acceleration and implicit force. If my hand
were constantly shedding paint, which direction would the paint fall?
Should I make a mark where my hand is now, or in the direction it is
moving? If I know where each finger is on a relaxed hand, can I tell how
flexed or stretched a finger is? Is that stress also a force to be recorded? If
I’ m swinging my hand, I often have less sense of where it is than of the
direction of its movement and the forces acting on it. Figure  16.3 shows
the velocity and the forces applied to many points along the tracked hand.
Even just drawing these vectors as arrows creates emergent art from the
hands’  motion. These forces could also be applied to particles “ held”  in
the hands, or to simulated agents, or used to manipulate a pixel buffer of
colors.

Rendering
We have inputs and a way to transform those inputs into progressive layers
of different kinds of data structures.
Wait, wasn’ t there supposed to be some art in this art toy? Certainly the
most memorable part of any art toy is the visual results of the interaction.
Drawing is often defined as “ making a mark on a surface,”  but what
does it mean to make a mark, and what is the surface? For an exhaus-
tive ontology of digital drawing tools, I recommend Alvy Ray Smith’ s

FIGURE  16.3   Hand-tracking image. On the left, hands are tracked as arrays
and acceleration. Even drawing these forces creates interestingly emergent pat-
terns. On the right, velocity and acceleration are drawn additively, with subse-
quent hand forces overdrawing each other to leave a trail of movement over time.
170   ◾    Procedural Generation in Game Design

“ Varieties of Digital Painting,”  but for the purposes of this chapter, here
are the basic concepts that you can use to build a rendering method:
Canvas : This may be a buffer or a series of composited buffers. If we
are playing in 3D, or with vector graphics, it may also be a scene graph or
list of renderable shapes. Regardless, it’ s a way of recording what we will
render to the screen.
Mark : A semipermanent element or action. It might read from the can-
vas, as a “ blur”  brush has to sample the pixels underneath before redraw-
ing them, or a “ fill”  command paints outward until it encounters an edge.
It is constructed, it is added to the canvas, and (because we aren’ t working
in reality), it may morph or flow or change or be deleted. For a traditional
digital paint program like MS Paint, the canvas is a simple pixel buffer, and
most marks just write a color into an area of the buffer, or start a function
that recursively writes color into expanding areas until a stop condition is
reached. The buffer is never erased, only added to. An Illustrator graphic
has no persistent buffer as its canvas, but instead has a scene graph. If the
scene graph is changed (a mark is added or a color is changed), then the
scene graph renders itself onto a temporary buffer and onto the screen.
Additive versus parametric rendering : In additive rendering, there’ s a
buffer that gets added to and evolved each time, based on the previous
state of the buffer. Nothing is ever discarded; it’ s just built over, which
mathematically looks like this for some buffer with five marks or actions
being made on it: final buffer  =  ƒ e  (ƒ d  (ƒ c  (ƒ b  (ƒ a (buffer))))). Parametric
rendering is rendering without a buffer memory. Each time, it renders
the whole finished product from scratch, doing whatever subrendering or
compositing it needs to do along the way: final buffer  =  ƒ (a , b , c , d , e . … ,
initial buffer). But the boundary between these two is surprisingly flexible.
In a buffer-based drawing engine like Processing or HTML canvas, I may
have a scene graph that I want to redraw each frame. I can “ erase”  the
buffer before drawing each time (setting each buffer pixel to a background
color), and thus achieve parametric rendering. Or I can choose not to clear
the buffer, and each parametric rendering will be made on top of the pre-
vious rendering. This technique leaves the past trails of a moving object
visible, which can be good for inexpensive trail rendering, notably used in
the Windows Solitaire win animation and the Microsoft 3D Pipes screen-
saver. There’ s also an interesting middle path, where the previous render-
ing is partially obscured. I often draw a partially transparent rectangle on
top of the previous rendering, leaving the trails visible, but fading out over
time in an attractively veiled way.
 Generative Art Toys    ◾    171

The two competing VR drawing apps, Oculus Medium  and Google Tilt
Brush , seem similar at first, with identical inputs, and both using a ges-
tural curve model to record their marks. But under the surface, one is an
additive and the other is a parametric rendering, and this changes their
interaction. Medium  uses voxels, a 3D version of the pixel buffer. The user
can “ mark”  more voxels into space, or write over previous voxels (or write
over them with blank space to erase). But one can’ t go back and edit an
earlier mark, other than by undoing to the point it was made. Tilt Brush 
is fully parametric, where each stroke is stored as a vector and then “ ren-
dered”  into the scene graph (which is later rendered onto the screen in
3D). This allows modification of previous strokes, and more interesting
ways of rendering strokes as animated curves or particle effects or pat-
terned ribbons.
Advanced and indirect rendering: Tilt Brush  and Medium  are both
intended (if unintentionally) as tools for good artists (note how their pro-
motional materials show famous Disney animators and illustrators using
them). They want to be flexible enough for anyone to build anything, so
they kept their rendering very directly related to the input. If you’ re build-
ing a smaller art tool, for more flexible users, the mapping from input to
output can be much less direct. Flexibility and control can be sacrificed for
generativity and power, and that can be a good thing!
Remember that the rendering being described here doesn’ t mean the
final rendering to the screen, but the way that the abstract model turns
into a mark.
One very common parametric approach to rendering is to have an
array of numbers, a “ string of DNA,”  which is rendered into some physi-
cal form. For an animated flower prototype, I turned an array of 30 float-
ing-point numbers to a set of rules for a flower-generating L-system. This
rendering process had several advantages. The space of the flowers was
interesting and varied, but also continuous. This allowed me to perform
an evolutionary algorithm on the flowers responding to the user input of
selecting their favorite. I could also “ lerp”  between new values over time,
mutating the flowers according to time. I also was able to hook the values
up to a music analyzer, which morphed them in time to music, and then
hooked it to a Leap Motion Controller, so that each hand position would
generate a different flower.
This is the powerful part of this “ input– model– rendering” model of art
toys. Once you have it built, you can begin switching out new inputs or render-
ing methods, generating new, playfully strange art toys with each variation.
172   ◾    Procedural Generation in Game Design

You can also use time as an input parameter for any parametric ren-
dering. Rendering at frame 2 will be slightly different than at frame 1 or
0. If your parametric functions are continuous, rerendering frames (or
rebuilding scenes) will create the effect of a smooth animation. The art
game Panoramical  is a lovely example of this. The user sets nine different
two-dimensional (2D) sliders, creating 18 different parameters. Time is
added as a 19th. Each of the many levels is a different parametric function
to interpret those 19 parameters into music and a 3D scene of geometric
objects. Changed over time, it creates an animated soundscape that is con-
trollable by the user.
Another engaging way to interpret models is to have simulations or
other algorithms using them as inputs. Any graph or set of points can be
turned into a Voronoi diagram or Delaunay triangulation. From there, it
could be rendered as a 3D mesh or as paths of connectivity for simulated
agents to move along. Models and inputs can create forces that act on the
elements in the world, rather than being elements themselves. The classic
art installation Text Rain  (Camille Utterback and Romy Achituv, 1999)
drops text from the sky on a projection, and the users create barriers with
their shadows that it can flow down or puddle upon. Art toy users could
also create wind, gravitational wells, water currents, or any other force to
move renderable elements. Imagination (and simulation power) are the
only limits.

OUTSIDE THE GENERATOR: JUDGMENT,

SHARING, AND CURATION
What happens between the user and the generator is important. Just
building it will create an art toy that is intriguing and generative. But often
some of the most powerful interactions happen outside the generator, in
gameplay built around it or in communities that arise outside. How does
the user learn to use the tool? Are extrinsic forces at play when they are
creating something? All these are other things to consider when building
the tool.
With the Spore Creature Creator , it’ s possible to look at a creature and
instantly judge whether you like it. This is true of most visual art tools, but
for artifacts like music, stories, or games, this “ judgment”  step can be sev-
eral seconds to minutes to hours long! Sometimes art toys can play with
this idea of judgment, creating artificial feedback that may or may not
reflect actual judgment. Become a Great Artist in Just 10 Seconds  pokes at
both the idea of judgment and the idea of art school by asking the user to
 Generative Art Toys    ◾    173

match their drawing to the given still life or landscape and judging them
accordingly. But this art toy doesn’ t allow the user to draw directly, only
by pressing keyboard keys to activate surprising and frustrating glitch-art
filters, making the automated judgment a silly and surreal metric.
Another art toy that gives the player control over curation but no cre-
ation is Picbreeder , an online tool to browser neural network– generated
images and “ breed”  new ones. Although the users’  only interaction is
clicking on the ones they like, users still show pride in finding unusual
ones in the space and showing them to other users.

CONCLUSIONS: CREATIVITY FOR ANYONE

There are very many aesthetics available for art toys. From frustrating and
opaque (Become a Great Artist ) to welcoming (Spore Creature Creator ),
intensive and controllable (Medium  and Tilt Brush ), uncontrolled and
automated (Picbreede r), or meditative (Text Rain  and The Treachery of
Sanctuary ), there can be art toys for every play experience and every arti-
fact type. Procedural content can make them more powerful, surprising,
generative, and welcoming to new users.
 Chapter 17

Audio and Composition

Bronson Zgeb
KO_OP Mode 

S kipping Stones  is an experiential game in which the player throws

found rocks from a beach over the water, creating generated music.
There are two major components to a procedural audio system like the
one found in our game Skipping Stones : producing sound and composing.

PROCEDURAL AUDIO IN SKIPPING STONES

Sampling
In order to produce sound, I used a sample-based synthesis approach.
The basic premise is that you create and record samples from your instru-
ments first to use as a basis for all the audio. Another approach would
be to synthesize all your audio in the engine, which, although possible,
requires a greater knowledge of audio synthesis and audio programming.
Furthermore, synthesizing complex audio can put a big strain on your
CPU. A sample-based approach gives you all the power of your existing
instruments, synthesizers, and sound effects (SFX) libraries, and is sim-
pler to understand and implement.
If you’ re familiar with audio hardware or software, you may be familiar
with the concept of a sampler. A sampler plays back short audio files while
simultaneously modulating them to produce modified versions of the base
sample. For example, the most basic sampler can play a sample at differ-
ent speeds, which also modifies its pitch. Even this basic functionality was
enough to start building my procedural audio system. But first, a little bit
of music theory.

175
176   ◾    Procedural Generation in Game Design

Pitch
Western music contains only 12 notes, which are then repeated over about
11 octaves, beyond which the notes become imperceptible to the human ear.
These 12 notes are C, C#, D, D#, E, F, F#, G, G#, A, A#, and B. In some cases, we
might call C# (pronounced “ C sharp” ) a D flat, F# a G flat, G# an A flat, and
A# a B flat, but for our purposes this doesn’ t matter. In fact, the names of the
notes are completely irrelevant and can sometimes be more confusing than
helpful. For example, octaves start on C rather than A, and there is absolutely
no correlation between any of the sharps (or flats) with their corresponding
nonsharps (or nonflats), that is, A# and A. The names of the notes are just
names, nothing more. Octaves are used to represent how high or low a given
note is. That is, an A4 (that’ s A on the fourth octave) is just a higher-pitched
version of an A3, which is consequently higher than an A2, and so forth.
Now that you understand the selection of notes you have at your dis-
posal when making music, it’ s important to understand what a note actu-
ally is. A note is a wave of sound created by an object that’ s vibrating at a
given frequency. So when a speaker plays an A4, for example, it’ s because
the speaker is vibrating at 440 Hz, which travels through the air until it
reaches your ears. The important thing to note here is that every note in
Western music corresponds to a specific frequency. Table  17.1 is a lookup
table of note frequencies by name and octave.
So what does this mean? First, every note in every octave has a spe-
cific pitch. Second, the ratio between a note and its neighbors is always
the same (as is the ratio between its neighbor’ s neighbors, its neighbor’ s
neighbor’ s neighbors, etc.).
From any note, you can multiply its pitch by the same known value
and land squarely on a neighboring note. This is the key to sample-based
synthesis. For example, if you loaded a recording of a C3 note from a piano
into a sampler, you could play it back as any other note from that same
piano within limits.
What limits? A basic sampler modifies the pitch by playing back sam-
ples at a different speed. As a result, there’ s only so far you can stretch a
given sample before it sounds terrible. How far you can stretch a sample
depends on the sample itself, but you don’ t expect to go beyond a couple
octaves up and down. If you need a wider range, you could use multiple
samples, one for every other octave range, for example. Using this infor-
mation, you can emulate an instrument playing any note in any octave
using only a handful of samples.
 Audio and Composition    ◾    177

TABLE 17.1   Pitch Values by Note and Octave

Note  1  2  3  4  5  6 
A 55.00 110.00 220.00 440.00 880.00 1760.00
A#/B♭  58.27 116.54 233.08 466.16 932.33 1864.66
B  61.74 123.47 246.94 493.88 987.77 1975.53
C 65.41 130.81 261.63 523.25 1046.50 2093.00
C#/D♭  69.30 138.59 277.18 554.37 1108.73 2217.46
D 73.42 146.83 293.66 587.33 1174.66 2349.32
D#/E♭  77.78 155.56 311.13 622.25 1244.51 2489.02
E 82.41 164.81 329.63 659.26 1318.51 2637.02
F 87.31 174.61 349.23 698.46 1396.91 2793.83
F#/G♭  92.50 185.00 369.99 739.99 1479.98 2959.96
G 98.00 196.00 392.00 783.99 1567.99 3135.96
G#/A♭  103.83 207.65 415.30 830.61 1661.22 3322.44

Implementation
Most of the instruments in Skipping Stones  are actually just basic sam-
plers. Each one is loaded with a sample of a C note in some octave (usually
C3), and when it plays, I modify the pitch in order to play the desired note.
Skipping Stones  was built in the Unity engine, which allows you to set the
pitch of an audio source as a floating-point number. So in this case, a pitch
of 1 means play at normal speed, 0.5 is half speed, 2 is double speed, and so
forth. I created a lookup table of pitches by calculating the ratio between a
note and its neighboring note and multiplying forward until I had enough
range.
Therefore, playing an instrument with a pitch of 1 will give me a C
(because that’ s the pitch of my original sample), whereas playing an
instrument with a pitch of 1.059463 will give me a C#, a pitch of 1.122462
will give me a D, and so forth. Additionally, since all my instruments are
loaded with a sample in C, I know that the same pitch across all my instru-
ments will produce the same note.
At this point, you’ re able to start building basic instruments, so it’ s a
good time to start working on your procedural composition engine.

PROCEDURAL COMPOSITION IN SKIPPING STONES

In order to build the Skipping Stones ’  procedural composer, I used the
principles of music theory and turned them into software. From that
point, it was possible to guide the stylistic choices of the procedural com-
poser to mimic those of our actual composer, which in your case could
178   ◾    Procedural Generation in Game Design

just be the style of a composer whose style you admire. I did this by adding
extra rules that could be triggered randomly, or using a statistical model.

Beat
The first thing you’ ll need is a way to keep all your notes on beat. In Skipping
Stones , the player skips stones across a lake to create music. Every bounce
generates a part of the composition. Our bounces were physically driven,
based on the force and angle of the throw, so we couldn’ t be sure that a
stone would bounce on beat to the music. As a result, the music sounded
more like a cat walking across a piano than actual music. So in order to
fix this, I wrote a metronome system to keep time. When a stone hits the
water, it submits an event to the internal metronome, which in turn sub-
mits an event to the audio system to play a given instrument when the
next appropriate beat occurs. If you’ re familiar with metronomes, you’ ll
know that after you set the beats per minute (BPM), it’ ll tick once on each
beat. But in order to make interesting music, you need to play notes in
between beats as well. So in order to stay rhythmically correct in between
beats, you’ ll want your metronome to keep track of half beats, quarter
beats, eighth beats, sixteenth beats, and so forth until you’ re satisfied with
the result. Depending on how you’ ve built your metronome, an easy solu-
tion is to double your BPM for every beat division you’ d like to add. For
example, for our purposes 180 BPM is the same as 90 BPM if you consider
every tick of the metronome as a half beat instead of a whole beat.
My metronome implementation was based around the InvokeRepeating
function, as shown below:

 InvokeRepeating( “  ProcessAudioSources ”  ,(60/
bpm),(60/bpm)); 

This was a quick and accurate way to implement a consistent beat. You
may be tempted to try using multithreading (or in the Unity engine, using
“ Coroutines” ), returning a call to WaitForSeconds(), but in practice, this
is prone to drifting off beat over time. InvokeRepeating is a function that
schedules a call to another function, in this case ProcessAudioSources,
after a given time and then repeatedly calls that function again after
another given time. So in this case, I’  m calling a function called
ProcessAudioSources on every tick, as determined by my BPM. Put sim-
ply, ProcessAudioSources fetches a list of any queued audio and plays it
back. In my particular case, it also fetches the next available audio source
 Audio and Composition    ◾    179

(because we operated with a limited number of audio sources that can play
at once) and applies any volume, pitch, or filter values before playing.
In Skipping Stones , the length of a note was controlled by the length of
the sample in almost all cases. However, it would be fairly straightforward
to add a simple envelope to audio sources. The purpose is to have proce-
dural control over your note lengths; that is, are they whole notes, half
notes, quarter notes, and so forth? This is done by dynamically lowering
the volume of an audio source after a given amount of beats, and turning
it off once it reaches zero.

Scale
Here I cover some basic principles of music theory so that we can encode
them into our composer. My goal is to provide as little information as pos-
sible in order to get started, so I take some shortcuts. It’ s also important to
understand that the “ rules”  of music theory were developed over centuries
of observation and experimentation, which is to say that it is not a science.
Music theory should be thought of more as guiding principles to lead you
toward something that’ s generally accepted as “ sounding good.” 
The purpose of a scale is to provide structure by telling you which notes
you’ re “ allowed”  to use, that is, out of the 12 notes in Western music,
which ones are OK and which are off-limits. For our purposes, you prob-
ably only need to understand the purpose of a scale and how it works, and
then let the computer do the rest.
Most of the scales used in Western music are just a variation on the
major scale, so I’ ll focus on that one. The major scale is seven notes long,
but we can’ t say exactly which seven notes until we’ ve chosen a root note.
For our purposes, the root note is simply the first note of our scale, and
choosing it is entirely up to us, or the algorithm.

Steps
It’ s important to understand the difference between whole steps and half
steps in music. A half step means moving one note forward, whereas a
whole step is two notes. First, I’ ll remind you of the order of the notes used
in Western music:

C C# D D# E F F# G G# A A# B C

So, for example, if we’ re on E and we take a half step, we’ ll be on F. If

we’ re on E and we take a whole step, we’ ll be on F#. So let’ s say we choose
180   ◾    Procedural Generation in Game Design

E as our root note, how do we know the rest of the notes that comprise our
scale? We’ re going to start on E, make seven steps forward, and every note
we land on is part of our scale. In the case of the major scale, the seven
steps are whole, whole, half, whole, whole, whole, half. So, referencing the
note order above, that would give us

E F# G# A B C# D# E

And that’ s our E major scale, which composers refer to as the key of E
major. But this isn’ t a book on music theory. For us as procedural design-
ers, it’ s more important to understand how we got here, so that we can
encode it into our composer. We’ ll want to encode other scales as well, as
they can have a big effect on the perceived mood of the music. It’ s impor-
tant to experiment with different scales and root notes while you’ re trying
to find the right mood, and make sure to take note of anything you find
particularly beautiful.
In Skipping Stones , I usually pick the root note at the start of the game,
and change when the time of day changes. In general, the root isn’ t some-
thing you want to change very often unless you know what you’ re doing.
The major scale is generally considered happy sounding, and our game
aims for a moody atmosphere, so we mostly use the minor scale. When the
game starts, or when we start a new composition, I make an array of notes
by cross-referencing my scale tables (pictured below) with my pitch table
(Table  17.1) and my chosen root note.

  
int[] Major = {2, 2, 1, 2, 2, 2, 1 }; 
  
int[] Minor = {2, 1, 2, 2, 1, 2, 2}; 
  
int[] PentatonicMajor = {2, 2, 3, 2, 3}; 

Once I have this information, I have to decide exactly how to traverse

the scale. Just traveling up or down the scale in order isn’ t very interesting,
as it sounds more like somebody practicing piano than playing a piece. I
use a state machine– like technique that starts on the root note and travels
no more than a few scale steps up or down. The state machine has a bit of
momentum, so if the piece is going up, it’ ll keep going up for a bit before
coming back down again, and vice versa. Be aware that some notes in the
major and minor scales create tension, while others relieve that tension.
In other words, it’ s not necessarily a good idea to jump around the scale
entirely randomly.
 Audio and Composition    ◾    181

The pentatonic major scale is a great example of a scale that doesn’ t have
this property. The pentatonic major scale is just the major scale, but with
the fourth and seventh notes removed. As a result, any sequence of notes
sounds good, but it’ s impossible to build a sense of tension and release.
A quick example rule you could deduce from this is that whenever you
play the fourth or seventh note in the major scale (which produce the most
tension), you’ ll want to follow it up with the root note (which releases
the most tension), but ultimately this is a stylistic choice and you should
experiment for yourself to determine what sounds good to you.
At this point, you can start to experience your procedural composer in
action, which means you can start to experiment with it and build stylistic
rules that appeal to you and your game.

Chords
Chords are simply notes that sound good when played together. Chords usu-
ally consist of three notes, but they don’ t have to. Here’ s my quick and dirty
guide to constructing chords. Once you know your scale, start on any note
in the scale. This is your chord root. From here, take the note two up (in the
scale), and finally the note two up from there. So, for example, the first, third,
and fifth notes of a scale make a chord, as do the second, fourth, and sixth, or
the third, fifth, and seventh. Now you know how to construct chords.
Depending on the scale you’ re using, the interval between those notes
can differ, and this will change the mood of the chord. For example, the
chord made of the first, third, and fifth notes of the major scale is a major
chord, and as a result sounds more uplifting. By contrast, the chord made
of the second, fourth, and sixth notes of the major scale is a minor chord;
therefore, it has a moodier sound.
This is the most basic way to construct chords, but there are ways to
modify a chord to give it different voicings. For example, normally the
root note is the lowest note of a chord, but this doesn’ t have to be the case.
Let’ s say we have a C major chord, played C3-E3-G3, where 3 represents
the octave in which the notes are played. We could play this same chord as
E2-G2-C3 or G2-C3-E3, which are known as the first and second inver-
sions, respectively. Notice we’ re playing the same notes, but the resulting
sound has a subtle difference.

Motif and Repetition

In music, we use the term bar  to describe a section of music that consists of
a certain number of beats. Since the music in Skipping Stones  is controlled
182   ◾    Procedural Generation in Game Design

by the player throwing stones on the water, there’ s no guarantee that a

piece of music will be constantly playing, as opposed to a normal song
where once you hit play, it’ ll play until the end. As a result, Skipping Stones 
doesn’ t use the traditional concept of a bar of music. Instead, I considered
every stone thrown as a single bar of music and I grouped together every
four bars as a section.
In order to make this come through in the music, I used a repeated
motif throughout the section. A motif is a short melody often repeated
throughout a piece of music. When the player throws the first stone, I
record the note sequence and consider this to be my motif for the upcom-
ing section of music. Then, on the second throw I’ ll recall this motif by
randomly inserting notes from the first note sequence among the newly
generated notes for this second bar of music. On the third throw, I com-
pletely disregard this motif, generating a completely new sequence. And
finally, the fourth throw repeats the motif, but with some notes replaced
by chords using that note as a root.
In addition to this, when a stone sinks I play either the second, fifth, or
seventh scale note, except in the case of the final stone, where I end on the
root note. As a result, each bar feels as though it’ s building tension until
the final throw, which resolves the tension and brings the section of music
to a logical conclusion, and every four stone throws generates and com-
pletes a new musical idea.
So the idea of motif and repetition gives procedural composition struc-
ture, but after some time it can start to sound predictable and dull. For
variety, I added in little flourishes here and there to give the music a more
improvisational feel. The first of these was to randomly play a short arpeg-
gio, often on the first skip of a given stone. An arpeggio is simply a chord
where the notes are played in succession, instead of simultaneously. The
arpeggios in Skipping Stones  are played quickly in order to give a little con-
trast to the otherwise slow pace of the music. Aside from the arpeggios, I
also play a few well-placed chords instead of single notes. This is really just
the tip of the iceberg when it comes to adding a bit of life into the music,
but for us, it was enough to fill what was missing, while working within
our time and budget.
Finally, I added accompanying instruments. For every time of day in
Skipping Stones , we built a primary instrument and an accompanying
instrument. Often, the primary instrument would be some kind of “ lead,” 
that is, some kind of piano or “ powerful”  (in sound) synth. The accompa-
nying instruments were more dreamy, with lots of reverb and slow release.
 Audio and Composition    ◾    183

Earlier I said that every four stones made up a single musical section, but
actually that’ s not true. The player had the ability to throw two stones
simultaneously, and would then have to wait for one to sink before he was
able to throw again. The first stone controlled the primary instrument.
This is the stone that generated the musical structure. The second stone
controlled our accompanying instrument.
The accompanying instrument didn’ t have as many rules associated
with it as the primary instrument. Because of its generally softer and
dreamier quality, nearly anything it played sounded good, as long as it
supported the primary instrument. In order to do this, it operated in the
same key as the primary instrument, and would sometimes make vague
mentions of the motif or play short arpeggios, but otherwise, it was free
to explore the music however the random number generator felt it should.

CONCLUSION
I could have continued to improve the composer of Skipping Stones , but we
had to end the project. While working on it, I found that every addition to
the engine made an enormous impact on the quality of the music it pro-
duced. If you’ re willing to learn some music theory and experiment, I’ m
certain you could build a more impressive procedural audio composer.
To this end, I recommend that you seek books that give concrete and
practical ways of applying music theory. I’ ve found Improvise for Real  by
David Reed to be one of these books, but you’ ll need access to some kind
of instrument to get the most out of it. Another excellent read is Making
Music  by Dennis DeSantis, which focuses on giving solutions to actual
problems when making music. Not every problem will apply to procedural
composition; however, many of the ideas are directly applicable.
It’ s important to understand that the beauty and expressiveness in the
Skipping Stones  audio engine didn’ t come from just the procedural com-
poser, but also the instruments designed by our human composer to work
within the engine. As in other aspects of procedural game design, we must
find a balance between authored and algorithmic elements, and the two
must collaborate in order to achieve the best results.
 III
Procedural Narrative

G enerating game content or assets is often the first instinct for

designers, but even the story experienced by the player can itself be
generated using similar methods. For more than 30 years, interactive fic-
tion and roguelike games have relied on a variety of techniques to build
narratives, sometimes using nothing but the most elementary compo-
nents. Other games have attempted to control pacing and tension with
artificial intelligence (AI) directors that can detect properties of the user’ s
ongoing playthrough and react accordingly. The text that the player reads
and the words that the player hears can all be generated. Nonplayer char-
acters can be created with personalities and histories that interlock with
the other elements of the setting. Everything in the author’ s toolbox is at
our disposal.
Perhaps even more so than other methods in this book, techniques of
narrative generation must be applied with care, because the results can
easily fall well short of what can be produced by a human author, while
draining a team’ s resources across disciplines. Be cautious with the scope
of your project, and remember that scripted and generated elements can
be favorably mixed. If you choose your goals well and understand the
interactions of your game’ s systems, you can create entirely novel situ-
ations that will surprise and delight players that come into contact with
your story engine. We can dream of a generated world in which the player
has a joint creative role, taking an active part in writing the story that is
being woven before them.

185
 Chapter 18

Story and Plot

Generation
Ben Kybartas
Independent 

T he role of stories in modern video games can be viewed as a grow-

ing field of experimentation and refinement. Stories in games may
simply be a way to provide causality and motivation to the actions taken by
a given player or players in a game, or even go so far as to use the mechan-
ics and interactions to reframe the player as a storyteller, eschewing tra-
ditional gameplay in favor of narrative experimentation. Stories may be
narrated, discovered in the environment, unfolded through cut scenes,
presented in dialogues, given as quests, embedded in mechanics, read
in wikis, and embedded within artificial intelligence, as well as emerge
through simulation, and much more. The point, in essence, is that what it
means to generate  stories for video games is a heavy and complex question.
Any approach to procedural story generation will need to make assump-
tions about both what a story actually means to the game and how it will
be represented. Faç ade  (Procedural Arts, 2005) and Left 4 Dead  (Valve,
2008) both had systems to measure the supposed level of drama and player
tension in the game, using marital turmoil and zombies, respectively, to
create the feelings of rising and falling tension associated with stories.
Contrarily, games like Dwarf Fortress  (Bay 12 Games, 2006) and Crusader
Kings  (Paradox Interactive, 2012) rely more on complex simulation to
structure the player’ s experience into emerging sequences of causal events
that create the feeling of a grand narrative. Thus, Faç ade  and Left 4 Dead 
focused on representing stories through emotion and dramatic tension,
187
188   ◾    Procedural Generation in Game Design

while Dwarf Fortress  and Crusader Kings 2  utilized complex world build-
ing and simulation to create a player-guided form of storytelling.
This chapter, alternatively, focuses specifically on the generation of
sequential stories that occur successively within a game world, similar to
quests in role-playing games or episodes of a television series. Relationships
and character attributes are used to drive a custom story grammar that
generates the actual story content, which in turn can impact the relation-
ships and characters of the game world. Grammars were chosen as the
representation for story in that they are relatively simple to author, under-
stand, and implement. Some disadvantages to grammars is that the results
tend to be highly structured and, at worst, could be perceived as repeti-
tive. This chapter is divided into several sections, starting first with an
introduction to grammars and story grammars. Second, we provide the
representation used for both the game world and story. Finally, we present
the design of the actual grammar rules, and discuss ways to improve the
generation process.

GRAMMARS AND STORY GRAMMARS

The approach to story generation taken in this chapter makes use of gram-
mars, particularly graph grammars. Probably the most well-known and
common use of grammars is the grammar used in language and sentence
construction. While obviously different, the grammars used in this chapter
follow the same philosophy. Essentially, a language grammar is a vocabu-
lary and a set of rules that can be used to construct “ valid”  sentences.
For our purposes, we will be designing a vocabulary and set of rules to
construct valid stories. In this section, we introduce the basic concept of
grammar using a simple string grammar as our example. Grammars are
built up of symbols  that form the content of the grammar, and rules  that
dictate how these symbols may be constructed. In this example, we will
use the symbols A , B , and C  and the following three rules:

A → C 
1.
A → B 
2.
B → CC 
3.

These are more formally called rewrite rules , where the symbol to the left
of the arrow is the pattern  that, if it exists in our string of symbols, can be
rewritten  to the symbol on the right-hand side of the arrow. For example,
 Story and Plot Generation    ◾    189

if we start with the string A , we can rewrite that same string as either C ,
using rule 1, or B , using rule 2. Furthermore, we can rewrite the string B  as
CC  using the third rule. Therefore, strings C , B , and CC  are all valid strings
that can be made with the starting symbol A . While this may seem limited,
if we add a fourth rule, say, C → A , then CC  can be rewritten as AC , CA , or
even AA , which in turn could become AB , and then ACC , and then ACA ,
and so on. Using only rules 1– 3, we are always guaranteed to reach a point
where the string can no longer be rewritten for any starting symbol; with
the fourth rule, we can now expand the starting string A  infinitely.
While the string grammar above may seem interesting, it doesn’ t have
any explicit story “ content”  yet. However, while the symbols above are only
strings, there is nothing that prevents symbols from being actual story con-
tent, such as story events  or structures. Early story structure research actu-
ally proposed that grammars could be developed to represent every story
structure in existence. Such a magical (likely fictional) grammar has yet
to be found; however, grammars have still found use for modeling struc-
tured types of stories, such as children’ s fairy tales (see Vladimir Propp’ s
“ Morphology of the Folktale”  for such an example). Imagine instead that
we were to create a simple grammar for creating sidequests in role-playing
games, which are also typically highly structured. We might break down
the potential events of a given sidequest into its own set of symbols, say
[Journey ], [Encounter ], and [Discovery ]. Next, we create a set of rules:

1. [Journey ]→ [ Journey ][Encounter ]
2. [Encounter ]→ [Encounter ][Encounter ]
3. [Encounter ]→ [Discovery ][Journey ]

Now, with these three rules, and starting with a [Journey ], we can
get quests like [Journey ][Encounter ][Discovery ][Journey ][Encounter ] or
[Journey ][Encounter ][Encounter ][Encounter ]. The next step could then be
to write to expand our set of rules and symbols, which take the abstract
quest structure we just developed and turn it into a more concrete set of
actions and events. Consider the following new rules and symbols:

1. [Journey ]→ [Go to town ]
2. [Journey ]→ [Go to the forest ]
3. [Encounter ]→ [Fight a goblin ]
190   ◾    Procedural Generation in Game Design

4. [Encounter ]→ [Fight a dragon ]
5. [Discovery ]→ [Find gold ]
6. [Discovery ]→ [Find weapon ]

Applying this set of rules to our previous examples gives us much more
specific quests, such as [Go to town ][Fight a goblin ][Find gold ][Go to the
forest ][Fight a dragon ]. We still don’ t need to stop there, and could create
a new set of rules and symbols to state what possible dragons could be
fought, what weapons could be found, and so forth.
In implementation, grammars are often used as procedural content
generators by taking one or more symbols as a starting set and then ran-
domly applying a number of valid random rules until either a point is
reached where no more rules can be applied or a certain number of rules
have been applied.
One advantage to the two grammars discussed in this section is that
they are visually easy to understand and reasonably straightforward to
implement. One disadvantage to the grammar-based approach is that by
keeping the grammar entirely dependent on itself, we avoid any mean-
ingful connection with the game world. Most stories are structured
around characters, and in particular conflicts between these characters.
Essentially, without tying the grammar more directly to the characters
of the world, we can’ t hope to have stories that meaningfully involve the
game world in the generated story. As such, we will want to create a new
form of grammar that operates simultaneously on both a story grammar
and a representation of the social game world. In this way, each rule con-
tains both a social pattern and a story pattern, followed by a change to
both the story and the social world.

GAME WORLD
One of the driving assumptions of this investigation was the idea that
stories are based around character relations, and specifically about con-
flicts between characters. The role of the story then surrounds the resolu-
tion of conflicts one by one, usually with one major conflict driving the
entire plot. Therefore, interesting stories can only meaningfully occur as
the result of an interesting set of characters and relations. This means we
should first aim to model the game world in such a way that all this infor-
mation is made explicit. Within this particular investigation, we chose to
model the characters and their relations as a social network— essentially,
 Story and Plot Generation    ◾    191

as a directed graph where each character is a node, and the edges represent
the relation of one character toward another. We define each character as
minimally having a unique ID and a set of attributes, which are essentially
a set of key– value pairs for each character, where the keys are the same
for each character, with only the value being different. “ Name”  and “ age” 
might be two examples of keys, where each character is expected to have
a name and age, but their actual names and ages will likely be different.
Likewise, we will define the relationships between characters in a similar
way, with each relationship containing a set of keys and values. “ Trust” 
and “ friendship”  may be two such keys. It is also useful to keep each edge
one directional, meaning that the relationship one character feels toward
another may not be the same as the relationship that character feels for
them. As a disclaimer, this is by no means the only way to have a model
of characters and relationships, but such a structure is intuitive and rela-
tively straightforward to author and implement. Likewise, similar rep-
resentations of characters and relations are relatively common, and seen
frequently in role-playing games and interactive fiction.
We assume a dynamic game world, in the sense that we expect that
the relations and even the attributes will change over time. In fact, given
that the core of the stories we are creating is related to characters and
conflict, it is arguable that a dynamic game world is necessary to repre-
sent the changes needed to construct interesting stories. It is also a lack
of dynamism that many players note (and lament) in a lot of branching
choice-based narratives or story-rich video games.
The very simple social model shown in Figure  18.1 will be used as
the basis for our subsequent examples in this chapter. Here, we have a
Western-inspired social world comprised of three nonplayer characters
(NPC) and the player. Each character has four attributes: a name, number
of bullets, money, and a certain level of gambling skills. There are three
types of relations: the money owed, the level of trust, and the affection
level between characters. For clarity’ s sake, we do not show all relations
between all characters in the world.

STORY MODEL
For our model of story, we will break away from the linear model used in
traditional story grammars and instead adopt the directed graph model
more commonly seen in interactive fiction and hypertext. Essentially,
in this model discrete sections of the story content form the nodes, with
edges indicating different branches of the story that a player can select. In
192   ◾    Procedural Generation in Game Design

001 002
Name: Ennio Name: Miller
Bullets: 0 Bullets: 2
Money: 50 Money: 15
Gambling skills: 1 Gambling skills: 2

Owes: 10 Owes: 0
Trust: 2 Trust: 3
Owes: 3 Affection: 0 Affection: 5
Trust: 5
Affection: 2

000 003
Name: PLAYER Name: Clint
Bullets: 6 Bullets: 2
Money: 10 Money: 1
Gambling skills: 3 Gambling skills: 5

FIGURE  18.1   Diagram of relationships between four characters.

this model, the actual story experienced by the player relates to the path
they take from one of the starting nodes of the narrative to one of its ter-
minal nodes. It should be noted that by doing this, we are now technically
working with a graph grammar  rather than a string grammar. This means
that rather than just rewriting one symbol into one or more different sym-
bols, we are instead rewriting nodes in a graph with new graph pieces.

RULE DESIGN
Now, given both our models of game worlds and story grammars, we can
begin to create a new rule structure that makes use of both. We’ ll start by
dividing the types of rules we want to make into two categories, which
we’ ll call initial rewrite rules  (IRRs) and secondary rewrite rules  (SRRs).
We’ ll define the IRRs as the set of rules that give us a basic story skeleton
with which to work, and the SRRs as rules that rewrite and expand the
story skeleton. Essentially, our process to build a story will be to apply one
IRR to get our initial story, and then apply a number of SRRs to progres-
sively make the story more complex. One advantage to this approach is
that we always ensure that at every step of the generation we have a com-
plete story; that is, even if we only apply one rule, we still have a story.
For the IRRs, we only care about searching for patterns in the game
world; that is, when designing the IRRs, the main question to ask is “ What
 Story and Plot Generation    ◾    193

interesting stories can arise out of the possible patterns in our social
graph?”  Returning to our Western example from before, we can see that it
is quite possible for someone to owe more money than they currently have.
Thus, we can use this need for money as the main conflict for a possible
story. Following the example of many typical Westerns, we can envision a
potential story where a character in debt tries to win back their earnings
in a poker game. Given that in many cases we want to involve the player in
the story, a possible IRR may look like that shown in Figure  18.2.
Put simply, the pattern on the left of the arrow is the game world pat-
tern we’ re searching for; in this case, we are looking for a situation where
there is a character X  who has less than 5 money and owes more than
that to another character, Y . If this pattern exists in the game world, then
we can initialize the story on the right, in which character X  plays poker
against the player and takes whatever winnings they have to character Y .
We use the circle with the dotted line to indicate possible starting points
of the story, whereas the plain circle marks the possible ending points of
the story. In this case, there is only one start and one end to the story in
question. Looking Figure  18.2, we can see such a pattern existing between
Clint and Ennio, so it is valid for us to generate the above story with Clint
taking the role of character X  and Ennio taking the role of character Y .
When designing an IRR, it is usually advised to use relatively simple
game patterns, such as using only two characters and their relations. The
reason for this is twofold. First, it can be quite costly performance-wise
to search a large graph for very complicated patterns, and it is a non-
trivial search to implement. Those familiar with performance issues in

Game world pattern Story results

X challenges
PLAYER
X Y
to poker
Money < 5 Owes > 5

X takes
winnings to Y

FIGURE  18.2   Effect of a rewrite rule.

194   ◾    Procedural Generation in Game Design

computing will recognize that it is actually NP hard to search a graph

for any arbitrary patterns, whereas it is much simpler and faster to search
with only two characters. Second, simpler patterns are far more likely to
exist in the game world than complex ones. While it is easy for a pattern
like the one above to exist, if we had a large pattern with many characters
and rules, the likelihood that that pattern would exist in a given game
world is increasingly unlikely. This is a problem, though, as complicated
relationship patterns and large complex stories are far more interesting
than the simple story we generated above, and it’ s for this reason that we
design the SSRs.

SECONDARY REWRITE RULES

SSRs are similar to IRRs, except that we are not creating a new story.
Instead, we are looking to rewrite specific events from our story, again
using patterns in the game world, to expand and complicate the exist-
ing story. This means that while IRRs are used to create general stories,
the SRRs should be much more malleable, and should be able to apply
to many events in many different stories. Returning to our example, we
can assume that there will probably be more than one story where play-
ing poker occurs as an event. We also know that cheating in poker is yet
another staple of Westerns, so we might want to define an SRR in which
two characters can mutually agree to cheat someone out of all their money
in poker.
Such a rule might look like Figure  18.3. This means that in order to
apply this rewrite rule, there must be a poker event in the story and  there
must be a character Y  who has a high enough level of trust of X  to help
them cheat. One such relation exists between Miller and Clint in our orig-
inal graph. This means that in our sample story, we can rewrite the event
such that Miller helps Clint to cheat at poker, robs the player of all their
money, and then has Clint deliver that to Ennio. Continuing along this
line, we may want to give the player the chance to discover the cheating
in question. Maybe Y  in the above pattern has high enough affection for
the player to let them in on the truth. Or maybe the player simply has high
enough gambling skills to tell something is wrong. These could both be
two separate SRRs that can be used for any gambling event where cheat-
ing is occurring. Then, we can author rules where the player is allowed to
confront characters they catch doing something illegal at gunpoint, given
a certain number of bullets. This could then devolve into duels, or surren-
ders based on how many bullets the criminal has, and so forth. In this way,
 Story and Plot Generation    ◾    195

Game world pattern Story results

X Y X cheats at
Trust > 2 poker with Y

Story pattern

X wins all
X playing poker money from B
with B

FIGURE  18.3   Another rewrite rule in effect.

we end up with a list of SRRs that could end up with our card game ending
in a violent duel, but those same rules could be used to duel wagon rob-
bers in the desert, or to catch people cheating at sharpshooting contests.
Ideally, there should always be more SRRs than IRRs, with many different
possibilities for all potential story events.
In terms of the actual application of SRRs, in practice it was reasonably
effective to first search for all valid SRRs, and then choose one at random
and apply it. This process could be repeated a fixed number of times for
each story. As an alternate approach, one could continuously apply SRRs
until there are no more valid SRRs to apply. However, it is easy to design
SRRs that may be applied infinitely and can break the generation. As such,
even if the designer is very cautious, it is still good practice to cap the gen-
eration process after a certain number of rule applications.

GAME WORLD SIMULATION

At this point, it’ s important to discuss the dynamism of the game world.
Currently, our stories do not have any concrete effect on the game world,
but in reality, we want our game world to be dynamic. In our Western
world, for example, we should have ways to gain trust and affection, as well
as ways for new characters to get into debt, and so forth. In this way, while
we may be generating similar stories, they will be occurring as the result
of an ever-changing, dynamic game world. To achieve this, we extend our
definition of the story model to include preconditions and postconditions.
Essentially, things in the game world must be true for certain events to
occur, whereas each event can in turn make changes to the game world.
196   ◾    Procedural Generation in Game Design

If we go back to the IRR defined above, we may want to have X  take the
winnings of poker to Y , only if X  actually won money in the poker game
(e.g., if the poker game was a minigame). In this case, we may want to have
the story branch at the two possible results, with a new second event that
X  leaves defeated if they didn’ t win at poker. Likewise, we would want to
reduce the trust toward characters who attempt to cheat a specific charac-
ter at poker, and so forth. This adds an extra level of complexity to design,
in that the designer must consider each possible arrangement of precondi-
tions that could occur in the story. In practice, it is often good to run a pass
through a generated story and assert that each event has its precondition
met, taking into account any postconditions that are also applied.

METRIC-GUIDED GENERATION
In this section, we discuss a brief and slightly more technical alternative to
the application of SRRs. Previously, we stated that simply selecting valid
SRRs at random and applying them was a valid way to generate stories,
and this is still true. Interestingly, though, since we are calculating all
valid SRRs, it is quite possible to be more selective about which SRRs we
choose. Essentially, if we define certain features we would like to see in the
resulting story, we can prioritize which valid SRRs we choose based on
how we want the final story to be structured. We may, for example, want
to prioritize player choice, and so we prioritize valid rules that add one or
more branches to the resulting story. Likewise, we may want our Western
to be filled with shoot-outs, and thus we prioritize any SRRs that add gun-
fighting events. Conversely, if we want to keep characters alive longer, then
we may want to avoid actions that involve their death. (It is morbidly easy
to end in Shakespeare-like scenarios where all characters have murdered
each other.)
As a more complex example, imagine we kept track of all the post-
conditions resulting from each individual story. This means we know all
changes to the game world that occurred as the result of a story that the
player directly experienced or influenced. Now imagine we prioritized any
valid SRR whose game world pattern is met due to the changes the player
has made in previous stories. In essence, this means that when we generate
stories, we prioritize stories that occur as a result of something the player
did in the previous story. This adds a great deal of causality between stories
and helps the player to feel that their interactions are valid. For example,
if the player chooses to duel with someone and loses all their bullets, and
then we prioritize the generation of a story where the player is held up and
 Story and Plot Generation    ◾    197

can’ t defend themselves, then the consequences of the previous stories can
affect all the subsequent stories. This, on top of a simulated game world, is
another way to keep the stories interesting, because even though the types
of stories generated may be similar, the context out of which those stories
are being generated is dynamic, ever changing, and sensitive to all the
actions taken by the player.

CONCLUSIONS
Story generation is by no means an easy task, and to date there are relatively
few games that have focused on story generation as one of their main fea-
tures. By consequence, story generation presents an interesting alternative
to the traditional means of authoring stories, by allowing the author to
focus more on the structures and rules that guide the player story, rather
than try to tie them to a particular story created by the author. And this
seems relevant, especially with an increased level of criticism toward the
minimal effects players typically have on the stories and worlds in game.
As the horizons of what a game is expand further and more experimental
works gain serious interest, it seems that games can be used not just as a
way to tell stories, but also to challenge traditional storytelling methods.
By focusing on experimental and generative elements, designing stories
becomes an exploration into how stories may be formed, opening that up
to the player, as opposed to simply telling them a good story through audio
logs and journals.
 Chapter 19

Emergent Narratives
and Story Volumes
Jason Grinblat
Freehold Games

MOTIVATION
In April 2015, I was fortunate enough to join a live stream that featured
Nick Scratch, Austin Walker, and Cameron Kunzelman—members of a
group of streamers called Stream Friends—playing my roguelike game,
Caves of Qud.* Caves of Qud is set in a far-future, postapocalyptic Earth
populated by mutated animals, robots, and sentient plants. The Stream
Friends were cooperatively playing as a single character, and they had
just descended into the depths of the game’s first dungeon, a shale cavern
called Red Rock. As they paraded into the darkness, they stumbled on a
pack of aggressive, humanoid hyena creatures called snapjaws. Carefully,
they assessed the danger of the situation. They scanned their surround-
ings for corridors to use as choke points, boulders to hurl at the snapjaws,
and anything else that might help them survive.
At this point, Nick noticed that one of the snapjaws looked different
than the others. He moved his cursor over it and examined its description,
discovering that it was a named snapjaw faction leader (Figure 19.1). Its
name was Ugo, and it was known as the “calloused tot-eater.” Its descrip-
tion also included details about procedurally generated relationships it had
with some of the game’s other factions. It turned out that Ugo was loved

* https://www.youtube.com/watch?v=4ewMeqsjtTI

199
200   ◾    Procedural Generation in Game Design

FIGURE 19.1  Description of Ugo, the calloused snapjaw tot-eater.

by snapjaws (predictably), hated by mollusks for “casting doubt on their

beliefs,” and admired by antelopes for “sharing freshwater with them.”
Reading this description shot sparks into the streamers’ conversation.
Immediately, they began to ponder, what do mollusks believe? And why
was Ugo so kind to the antelopes? After all, freshwater functions as cur-
rency in Caves of Qud. Sharing some with another clan would serve as an
act of great friendship. They conjectured that the mollusks were feuding
with the antelopes, and Ugo was caught in the middle. At this point, they
accessed their character’s reputation screen. They were dismayed to dis-
cover their dismal reputation with antelopes. The creatures actively dis-
liked them. All of a sudden, it made sense that Ugo and his pack would be
hunting their character.
The battle commenced. The streamers hacked at Ugo with their bronze
battle axe, while Ugo and his compatriots retaliated with crude clubs,
claws, and bites. The battle turned sour for the streamers, and they were
forced to flee. They put distance between themselves and Ugo’s horde
and rounded a nearby corridor, only to find another pack of snapjaws.
They sighed collectively, knowing it was over. Resigned to their fate, they
accessed their character’s inventory and took a swig from a canteen of pre-
cious freshwater. Then, at Austin’s suggestion, they defiantly emptied the
canteen over their character’s head. Ugo would be sharing no freshwater
with antelopes on that day. Charging toward Ugo, they were felled by the
snapjaw horde.
 Emergent Narratives and Story Volumes    ◾    201

When we designed Caves of Qud, we had no idea who Ugo was or

that he would exist at all. We didn’t craft a narrative about a water feud
between mollusks and antelopes, nor did we tell the story of a bold adven-
turer who in his last moments defies his enemy by spilling his precious
cargo. We didn’t tell these stories, but we set the stage for them with pro-
cedural systems.
Emergent narratives like the one that unfolded before Nick, Austin, and
Cameron have a special home in procedural games. As designers who seek
to evoke certain experiences, they offer us a tool beyond the reach of tra-
ditional narratives. They’re personal—no one will repeat Stream Friends’
experience, as no narrative track has been laid out with it in mind. The sto-
ries that emerge from games that dive deep into the realm of procedural
systems are some of the most shareable precisely because the conditions
that produce them are ephemeral. Think of Dwarf Fortress, whose myriad
systems and starting conditions engender all sorts of emergent stories
with only the barest narrative framework.
Emergent narratives are also authentic. In resulting from the collision
of systems rather than the script of an author, they mirror the experi-
ences we have in our everyday lives, the stories we hear from our family
and friends, and the tales we read about in history books. In his book
Persuasive Games: The Expressive Power of Videogames, Ian Bogost argues
that “videogames open a new domain for persuasion, thanks to their core
representational mode, procedurality.”* He calls this domain procedural
rhetoric, “the art of persuasion through rule-based representations and
interactions rather than the spoken word, writing, images, or moving pic-
tures.” Stories that emerge from game systems are manifestations of an
argument those systems make. Part of their power is the authenticity they
possess on account of their scriptlessness.
Games have a unique power to enable player narratives. This chapter
sets out to explore that power and guide designers toward using procedur-
ality to cultivate the kinds of stories they want their players to tell.

THE VOICE IN THE MACHINE

The leading lament of designers considering a foray into the world of
emergent narrative is the loss of authorial voice. The power of a scripted
narrative lies in the author’s use of literary elements like theme, tone,

* Bogost, Ian, Persuasive Games: The Expressive Power of Videogames (Cambridge, MA: MIT Press),
ix.
202   ◾    Procedural Generation in Game Design

voice, and mood. The expressiveness of the narrative form, then, lies in the
author’s faculty to leverage those elements in their stories. On account of
their uncontrollable nature, do emergent narratives occupy a space inac-
cessible to the author or designer who seeks to express not just something,
but something in particular?
I argue that there is room for authorial voice in procedural systems
that enable emergent narratives, although the designer’s role diverges
from that of the traditional author. In procedural design, it’s possible to
insert elements like themes and mood into both the rules of generative
systems and the handcrafted atoms that get procedurally combined. The
result is less precise than the traditional story but richer in other ways.
Whereas the author crafts a storyline that traverses narrative space, the
procedural designer crafts a story volume* that contains numerous adja-
cent stories exploring similar themes or invoking a similar mood. It’s
important to note that a story volume doesn’t enclose an arbitrary lump
of narrative space. Instead, it encloses a family of emergent stories, all
of which are begotten by a set of carefully curated system parameters
(Figure 19.2).
For example, with Caves of Qud, while we didn’t anticipate the pre-
cise story of Ugo, we did want our players to explore certain themes
present in his story. For one, we were interested in how we (as humans)
relate to our morphology. The exploration of what drives a sentient
mollusk’s beliefs taps into that theme. We were also interested in a
world where the social structures are built on the scarcity of resources
we take for granted. The streamers’ defiant act of water wasting only
resonates in such a world. By observing the inhabitants of that world’s
relationship with water, we get some insight into our own relationship
with it.
Consider the story volume constructed by another game that leans
heavily on procedurality yet manages to maintain a clear voice. Red Hook
Studios’s 2016 dungeon crawler Darkest Dungeon explores the psychologi-
cal stresses of dungeoneering. When you lead an adventuring party of
procedurally generated characters from your gothic manor into the rotted
Weald, the outcome is undetermined. The narrative that emerges is your
own. However, the game’s systems—such as the afflictions your characters
accrue from traumatic events—constrain the space of possible narratives

* Project Horseshoe Group Report, “Generative Systems, Meaningful Cores,” “Story Volumes
Rather than Storylines” section, http://www.projecthorseshoe.com/reports/featured/ph15r3.htm
 Emergent Narratives and Story Volumes    ◾    203

Storyline

Story volume

FIGURE 19.2  Single storyline contrasted with several storylines that share
themes and comprise a story volume.

to those that explore the effects of trauma and evoke the mood of hope-
lessness. Even success stories are premised on overcoming your afflictions
rather than avoiding them altogether.

FIASCO: A CASE STUDY

Let’s examine the particulars of a procedural game whose rules create a
well-defined story volume of emergent narratives. Fiasco is a collabora-
tive, tabletop role-playing game published by Bully Pulpit Games in 2009.
It follows in the tradition of other pen-and-paper role-playing games like
Dungeons & Dragons, but it distinguishes itself in a few key dimensions.
It’s designed for short play sessions, it privileges role-playing over rules
complexity, and it eliminates the role of the game master in favor of a col-
laborative approach to story creation.
Fiasco also distinguishes itself with its explicitly articulated themes. As
argued by media researcher Felan Parker, “Fiasco is designed to simulate
204   ◾    Procedural Generation in Game Design

the caper-gone-wrong subgenre of film.”* In particular, Fiasco draws

heavily from the films of Joel and Ethan Coen. These films, although dis-
parate in their plots and settings, all tend to explore the themes of human
stupidity and overreaching ambition. They usually feature an everyman
character—typically male—navigating a social system he doesn’t under-
stand, often at a tragic cost.
Although Fiasco uses pen and paper as its medium, it features proce-
durality at the core of its systems. At the start of the game, players roll a
pool of six-sided dice and consult a series of tables containing plot and
character elements, such as locations, relationships, and needs. Taking
turns, players choose elements by selecting a die from the pool and match-
ing it to the element’s index in the table. Each player must have a need and
a relationship with the players next to them. Then, the players role-play a
series of scenes based on these plot elements. Halfway through the game,
players reroll the dice and choose elements for the Tilt, an unexpected
event that shakes up the story. The players role-play a second series of
scenes informed by the Tilt, and then they resolve the game in an epilogue.
The stories play out to their inevitable tragic ends. The game supplies a
variety of settings—a Western-style saloon, sleepy suburbs, and a remote
research station in Antarctica—for the capers to go wrong. How does Fiasco
constrain narrative space to include only those emergent stories that touch
on its themes? It embeds the restrictions in its game rules. By requiring
each player to select a need and a relationship with their neighbors, and by
constraining player choice through the use of dice assignment, the rules
coerce players into choosing conflicting needs and relationships primed
for souring. Furthermore, the rules ripen the conditions for conflict and
tragedy by introducing an unexpected event during the Tilt—one that’s
usually at odds with the goals of the characters. Players can’t avoid the Tilt;
they must simply adjust to its ramifications.
Fiasco also encodes its themes in the handcrafted units that get proce-
durally combined during the game’s Setup and Tilt phases. These are the
individual entries in the plot and character tables. Let’s examine one of the

* Parker, Felan, “The Set-Up, the Tilt and the Aftermath: Role-Playing the Caper-Gone-Wrong Film
in Fiasco,” http://www.academia.edu/4530327/The_Set-Up_the_Tilt_and_the_Aftermath_Role-
playing_the_Caper-Gone-Wrong_Film_in_Fiasco
 Emergent Narratives and Story Volumes    ◾    205

tables from the Tales from Suburbia setting. These are relationships from
the Work and Crime categories.

Work Crime
1. Business rivals in a dying industry 1. Drug manufacturer and dealer
2. Service worker and client (restaurant, bank, 2. Gambler and bookie
janitorial)
3. Professional supervisor and employee 3. Th
 ieves (shoplifters, burglars, car
thieves)
4. Tradesman/homeowner (lawn care, 4. Small-time vandals, ne ’ er-do-wells
plumbing, HVAC)
5. Salesman/homeowner (siding, drive 5. Embezzler and company accountant
resurfacing)
6. P rofessional/client (pastor, doctor, lawyer, 6. Organized crime figure and wannabe
banker)

Most of these relationships entail a degree of tension. A gambler and

bookie, business rivals in a dying industry, service worker and client—
we can see how they’re primed to erupt when combined with contentious
needs or unexpected obstacles. Introducing a Tilt element like “someone
develops a conscience” into the relationship of an embezzler and company
accountant is likely to bring the powder keg to its inevitable conclusion.
Fiasco’s designers could have chosen to include any number of rela-
tionships from among the myriad ways we relate to each other as human
beings. They chose these relationships because they work to constrain nar-
rative space in a thematically resonant way. While each story plays out dif-
ferently, they’re all bound by the themes Fiasco explores. In its own words,
Fiasco is “a game of powerful ambition and poor impulse control.”*

UNPACKING YOUR THEMES

The first step in approaching a procedural design that yields a particu-
lar story volume is to identify your themes. This process can be trickier
than expected, especially if you don’t have experience engaging critically
with media. Many creators, myself included, are driven by intuitions
around what makes for compelling ideas. I get a vision in my head, such
as a strong-willed and principled mollusk, and I think, “This needs to be
in my game!” The goal is to examine these intuitions and try to unpack
them. Once you’ve teased out the kernel that makes your idea tick, you

* Fiasco back cover.

206   ◾    Procedural Generation in Game Design

can formulate it as a broader theme. Then you can envision narratives that
explore other aspects of that theme. Finally, you can develop systems that
enable player stories within your desired story volume.
With Caves of Qud, we started with fond memories of our Gamma
World campaigns. Gamma World is a tabletop role-playing game origi-
nally published by TSR in 1978. It’s set in a pulpy, postapocalyptic North
America that’s been transformed by nuclear fallout into a bizarro ecosys-
tem of mutant flora and fauna. As teens, we at Freehold Games played
Gamma World extensively. Over the course of our gaming sessions, I role-
played as, and interacted with, teleporting tortoises, psionic pelicans, and
parasitic fungi. These experiences were absurd, hilarious, and wonderful.
They served as our primary inspiration for Caves of Qud.
Guided by our intuitions, we transposed the simplest of these expe-
riences to Caves of Qud by designing mechanics that enacted them. For
example, we wanted to let the player play a character with a carapace,
wings, telepathy, or all three. So we designed a character creation sys-
tem that includes mutations, which combine to produce morphologically
diverse characters. As we started to explore more complex social interac-
tions, and systems that could mimic them, we examined our intuitions
around what made these interactions compelling. For instance, why did I
enjoy the thought of a mollusk bristling at a challenge to its beliefs?
After much meditation, we concluded that the kernel of these interac-
tions is our relationship with our morphology. When I’m imagining how a
mollusk might respond to a social exchange, I first imagine how I might
respond as a human. Then I extrapolate by imagining how the morpho-
logical differences between mollusks and humans might inform their
interactions. I go through a process of anthropomorphizing the mollusk,
and then deanthropomorphizing it. All the while I’m implicitly examin-
ing how my own morphology informs my interactions.
This formulation helped us envision how other creatures’ morphologies
might inform their social exchanges. As we went further, it helped us start
to parameterize these exchanges and design a system to reproduce them.
We followed a similar process for the theme of scarcity. Originally, the
idea to use water as currency came to us as a play on familiar roguelike
mechanics, hunger and thirst. Then, we started to tease out the ramifica-
tions of our idea. In particular, we observed how roguelikes are some of
the few games to accurately model our utter dependency on water. We
realized our extension to the mechanics could be used to explore the social
dynamics that emerge from that dependency.
 Emergent Narratives and Story Volumes    ◾    207

Ultimately, these examinations led to the design that produced the net-
work of relationships between Ugo, mollusks, antelopes, freshwater, and
the player. The narrative that unfolded from these relationships, along
with all the other narratives that unfold from its variations, comprise the
story volume articulated by our design.

GRIST FOR THE NARRATIVE MILL

As a final point, let’s look at how simple touches can help players narra-
tivize the events they experience in your game. I argued for how systems
themselves can work toward shaping your story volumes. For example,
the presence of the Tilt in Fiasco works toward creating narratives where
characters can’t anticipate their outcomes and are often devastated by
them. But small details in your content can facilitate narration and shape
your story volumes as well. Character and place names, item descriptions,
level layouts, and sound effects can all serve as grist for the mill when it
comes to enabling thematic player stories.
Take our decision to give Ugo a name. By virtue of his position as a
snapjaw faction leader, we gave Ugo unique art and a descriptive title. For
the purpose of communicating his mechanical distinctiveness, his name
is superfluous. But a name is thematically resonant. It implies certain
characteristics about Ugo’s culture, and so it acts as a narrative spring-
board for players’ interaction with him. If we were handwriting a story
about a snapjaw faction leader, we’d almost certainly give him a name to
personalize him. By the same principle, emergent narratives emerge more
discernibly when their subjects are positioned as components of a narra-
tive. Stories need characters, and Ugo’s name positions him as a character.
Beyond facilitating narration, you can also shape your story volume
with your choice of details. Ugo’s descriptive title of “calloused tot-eater”
evokes a certain mood. In conjunction with other details, like the aggres-
sive attitudes of snapjaws and the scarcity of water, we know that the nar-
ratives likely to emerge around Ugo involve hostility and conflict. Ugo’s
affable relationship with antelopes adds even more texture, and a portrait
starts to emerge of a complex figure mired in unusual social relationships
premised around scarce resources. With this amount of context, the play-
er’s interactions with Ugo take on new meaning. As the Stream Friends
discovered, Ugo was more than a monster in a dungeon.
 Chapter 20

Poetry Generation
Harry Tuffs
Independent 

1. “ Tales of the Lunatic Horn of Pride” 

Dramatic verse. 88 lines, 22 stanzas. Extract:

As prophets foretold, the ladies will distil

Never march pleasant into rusty mystique
Disdain be not mean, though some call thee still;
So remember this ballad of vanity and technique.

2. “ Hunting with the Magnificent Horse” 

Heroic epic. 642 lines, 107 stanzas. Extract:

The stallion of chaos is burning perfumed

Mythic, tough, lusty, and consumed
Always move gentle into lusty rocks
Such was the pony whose madness was locked.

3. “ The Dancing of the Politician” 

Satire. 56 lines, 2 stanzas. Extract:

Politician! At last I behold thee, dirty and fair!

Bloody, dumb, dirty and fair
With monkeys as cheap as your bloody light,
The sister of wrath and wounded flight.

209
210   ◾    Procedural Generation in Game Design

These are three examples of the procedurally generated poetry in my

RPG, A House of Many Doors . The poems I generate share a few com-
mon qualities: they are all presented in game as four-line “ extracts”  from a
larger piece. They all (attempt to) deal with a single overarching theme— a
horn, a horse, a politician. And most importantly of all, they’ re all pretty
bad. Just not very good at all.
Why are they bad? Well, because procedural generation of text is
difficult, and our brains are such finely tuned word tools that “ lan-
guage”  is basically a synonym of “ t hought.”  When we generate rocks
and trees and continents, they are imperfect, but unless they’ re really
misshapen, we don’ t notice. When we generate poems, their lack of
quality is immediately obvious because we are deeply familiar with
how language works. It’ s an effect of uncanny lexis, similar to the
uncanny valley— we can make beautiful trees in computer-generated
images (CGIs), but we haven’ t mastered the human face. Familiarity
breeds difficulty.
Now some insanely clever people at Google— people who are far better
at this than I am— have also been teaching computers to write poetry. But
instead of procedural generation, they’ ve been using staggeringly complex
neural networks. I include an example below:

i don’ t like it, he said. 

i waited for what had happened.
it was almost thirty years ago.
it was over thirty years ago.
that was six years ago.
he had died two years ago.
ten, thirty years ago.
“ it’ s all right here.
“ everything is all right here.
“ it’ s all right here.
it’ s all right here.
we are all right here.
come here in five minutes. 

Now, this is hugely impressive for many different reasons. This repre-
sents the struggle toward true artificial intelligence (AI). AI that can learn
and adapt, AI that actually understands grammar rather than being told
by a programmer to generate text in a way that obeys grammatical rules.
 Poetry Generation    ◾    211

In almost all respects, this is much more impressive than anything I’ ve
done. One area that it doesn’ t impress, though, is the quality of the poetry.
See, my proc gen poems are bad. But not that bad.
By qualification, I am not a computer scientist. I am an English literature
graduate, giving me an appreciation for— and understanding of— poetry,
and I think this helped when it came to coding a proc gen engine that
created the stuff. I went into the process with a different perspective, I
suspect, than most computer scientists might have had. One of my biggest
priorities was to categorize my word database not only by rhyme, lexical
category, and meaning, but also by number of syllables.
See, meter is important in poetry. Especially in the kind of premodern-
ist poetry I wanted my engine to write. And while I wasn’ t quite up to the
task of categorizing each word by which syllable was stressed, I could at
the very least ensure that the number of syllables was taken into account.
And just like that, I had a way of ensuring my proc gen poems had
appropriate cadence. It made an immediate difference, and I suspect I’ d
never have thought about it if I approached this as a computer science
problem rather than a poetry problem.
So what’ s the takeaway from this? Real-world knowledge of what you’ re
trying to generate can be helpful at worst. At best, it can cause you to
totally reevaluate your approach. If you don’ t have three years of college
study on hand, a week or two of intense research will probably do the trick.
Let me take you through my process of poetry generation, top to bot-
tom, beginning with a terrible admission: I originally didn’ t intend to gen-
erate poetry at all.
Part of the premise of A House of Many Doors  was that the player
character would collect memories and experiences in the course of the
story— A Moment of Melancholy, say, or a Horrifying Ordeal, or a Droll
Recollection— and spend them to write poems. It was essentially just a
way to collect experience points and level up. I created a system to generate
the poems’  titles, and then told the player, ‘ You wrote a poem!”  The poetry
itself was left to the imagination.
Then, in the run-up to the ultimately successful Kickstarter, I carelessly
mentioned that the game allowed the player to write generated poetry.
This caught people’ s imaginations like wildfire. With trepidation and
excitement, I realized I’ d have to actually generate proper poems.
I’ d already been interested in doing this, but I was under the impression
that it would be a waste of time— I didn’ t realize that so many people were
actually interested in the idea.
212   ◾    Procedural Generation in Game Design

Why am I telling you this? To explain one of the central pillars of my

poetry system— it selects a genre first, titles second, and then everything
else after, based only on these first two. This is the opposite of how real
poems are written. It wasn’ t a part of any plan, but it ended up work-
ing rather well! Let’ s take “ Hunting with the Magnificent Horse”  as an
example.
If the player spends Breathtaking Spectacles to write a poem, then
the poem will be a heroic epic. A title will be collated from a database of
appropriate words— since this is an epic, it might be a grandiose adjective
like magnificent  and a martial-associated noun like horse . Then a title will
be created from these words. At this stage, it’ s essentially poem title Mad
Libs.
Along with the genre, the title becomes the basis for everything after.
Synonyms are collected for hunting  and horse — and then when the poem
needs to find a noun or verb, it has a high chance to choose one of these
(e.g., stallion  or pony ). Loosely related words, but not exact synonyms, are
also collected— there is a slightly lower chance of these being picked (e.g.,
lance  is associated with horse , but is not an exact synonym). Then there is a
chance to choose something completely unrelated, but otherwise relevant
to the genre (for a heroic epic, this could be something like sword  or hero ).
Finally, there is also a very small chance of totally random deviation (just
for fun).
This has the effect of creating a (mostly) coherent subject matter and
theme that runs throughout the poem, but which isn’ t totally rigid. And
remember that all these words are categorized by syllable, too, and by
rhyme for the words that end each line.
Now that the important nouns and verbs and adjectives have been
chosen, the system runs through a similar process for lines. Let’ s take a
line from “ The Dancing of the Politician,”  “ With monkeys as cheap as
your bloody light” — before the important words are added, this would be
“ With       as       as your               .”  Which lines are chosen
is weighted by genre and by the grammatical construction of the preced-
ing lines. Again, there’ s a very small chance of utterly random deviation
here, because sometimes it’ s fun to throw a spanner into all my hard work.
The word database in A House of Many Doors , with all its synonyms and
rhymes and syllable counts and genre weighting, consists of almost 10,000
words. It can spit out more poems than there are atoms in the universe.
I tried to give the player as much control over their poems as possible.
In retrospect, though, I wish I’ d gone further.
 Poetry Generation    ◾    213

The player in A House of Many Doors  isn’ t passively clicking a button

labeled “ Generate poem.”  With hard work and in-universe justification,
they acquire a resource, and then choose where to spend it— choosing what
kind of poem they wish to write. An epic saga? A melancholic lament? A
satire, a romance, a dream vision? The player feels involved in shaping the
poem that ultimately results, even if the code does the heavy lifting.
This is part of the appeal of the very best world builders, too. Clicking
a button labeled “ Generate world”  is fun. But choosing world size, popu-
lation demographics, continent layout, or— as will be the case in Dwarf
Fortress — the magic system? That makes it a process that the player can
endlessly delight in. Rather than a novelty act— a “ look what the computer
can pull out of its hat”  one-time trick that gets dull quickly—the process
of procedural generation becomes an experiment.
I encourage other proc gen pioneers to hand over the reins to their play-
ers as much as possible. By allowing your players to experiment like this,
you’ re not just ensuring they have fun— you’ re exposing the complexities
of the system in an intuitive way that would never be apparent otherwise.
Your world builder can model rainfall distribution and no one will notice.
But give the player a slider to alter rainfall distribution, and let them watch
as deserts form— suddenly they will admire the depth of the simulation
that much more.
Like I said, though, I wish I’ d gone further with this in A House of
Many Doors — it wouldn’ t be unfeasible to allow the player to choose to
write about a politician!—  but I’ m too far along in development to change
this now. Perhaps in my next project.
Another regret is the relatively short time span I had. This is an inevi-
table side effect of commercial game development. If I had a few more
months to spare, and the will to remake the system from the bottom up,
there’ s a lot I’ d do differently. I would categorize words by metric feet as
well as syllables. The syllable count system helped created poems with an
approximation of meter, but with (a lot) more time I could add a system
that checked for stressed and unstressed syllables too. Then I could gener-
ate poems with proper, ruthlessly administered iambic pentameter in the
Shakespearean vein— and almost anything in iambic pentameter sounds
nice and literary.
Even with all this theoretical extra work, though, the poems would still
be bad. As discussed already, all computer-generated poems will be bad
and stilted and suffer from uncanny lexis (when this rule stops being true,
we will all have been enslaved by our robot overlords anyway).
214   ◾    Procedural Generation in Game Design

So I suppose if I had any one piece of advice it would be, embrace bad-
ness, as long as it’ s funny or interesting or both. Your proc gen system
might be very complicated and probably took a lot of work, but it will occa-
sionally produce bad results. If you can make sure those are funny– bad,
rather than frustrating– bad, you’ ve already won.
 Chapter 21

Characters and
Personalities
Emily Short
Independent

INTRODUCTION
When we talk about procedurally generated character and personality,
there are several things we might mean by that:

Description: A system that procedurally creates descriptions of

characters or personalities, but which is not primarily concerned
with making those characters interactively meaningful.

This type of system might be used to fill in background detail, for

instance, by offering small profiles of the antagonists one encountered in
a sandbox video game environment, or as part of a generative art project.

Realization: A system that procedurally creates nonplayer char-

acter (NPC) dialogue and actions in response to the player’s
actions, interactions with other NPCs, and/or events within the
world model.

This type of system finds the best way to perform predefined character
traits, allowing a character to demonstrate a depth of responsiveness that
we could not accomplish purely through writing. However, the characters’
essential traits have been chosen by the game writer or by the player dur-
ing a character creation process.

215
216   ◾    Procedural Generation in Game Design

Creation. A system that procedurally assembles NPC character

profiles, compositing personality traits, appearances, loyalties,
and other features that are both fictively meaningful and func-
tional within the game world, so that the player encounters a
perennially changing cast of characters.

This type of system can verge on story generation, since in some cases
narrative events might arise from the dynamic interaction of characters.
Highly simulation-focused systems such as The Sims can produce emer-
gent (if not always consistent or well-formed) stories when characters with
conflicting personalities are put in the same household, for instance.
In this chapter, we focus primarily on realization and creation tech-
niques. Character description has a number of applications in game
development and procedural art—in fact, it is the kind of procedural char-
acter work I have had the most call for over the course of many different
text-based game projects. However, autogenerating a character descrip-
tion often uses many of the techniques already outlined in this book. We
might, for instance, use a context-free grammar to build up a paragraph
about a character, drawing on a preexisting corpus of character traits.
Character realization and character creation, on the other hand, require
some additional techniques.

SOURCE MATERIAL
In discussing these issues, I’ll be referring to examples from several of my
past projects.
In When in Rome 2, the player has to encounter, identify, and subdue an
alien life-form. The game selects the alien type randomly at the beginning
of each play session. Different behavior rules correlate with different alien
traits, so the player can learn about the alien by watching how it interacts
with its environment. When in Rome 2 is available at http://ifdb.tads.org/.
Versu is a game platform that Richard Evans (Black & White and Sims
3) and I developed together. Versu modeled characters as agents with inde-
pendent desires; each agent would choose the action they most wanted to
perform out of a list based on the current social situation. Although this
system is no longer under production and the games written with it are
not available, it has been fairly extensively documented. More information
about Versu is available at versu.com, including papers for both lay and
academic audiences about the underlying system and the toolset that we
used to create new content.
 Characters and Personalities    ◾    217

Blood & Laurels was the final game I wrote for the Versu platform, and
it was released as a stand-alone app for the iPad. (It is no longer avail-
able due to incompatibilities with recent versions of iOS.) Blood & Laurels
was set in a fictional version of ancient Rome, and was a story of intrigue
and conspiracy to overthrow the emperor. The story had a preauthored
branching narrative structure, but within each scene the interactions were
highly procedural, allowing characters to share information and flirt with
and fight, betray, and poison one another before moving on to a fresh
scenario—a sandbox embedded in a preauthored structure.
The Mary Jane of Tomorrow is a game about training a robot to exhibit
some personality traits (enthusiasm, courtesy, and knowledge of partic-
ular subject matter) and suppress others (ennui and knowledge of other
subject matter). Because the player can assemble traits in almost any com-
bination, this project relied heavily on text generation techniques to real-
ize the robot’s evolving speech patterns. The game can be found at http://
ifdb.tads.org/.

REALIZATION
For most of my projects, procedural realization has meant choosing what
to say, and then how to say it.
Choosing what to say might be dictated by the needs of the plot in
a story-based game—for instance, we might know that a particular
NPC needs to set a quest during the course of the next scene. We
might also choose dialogue in response to a question from a player,
or as part of a social interaction that we’re modeling.*
If our character doesn’t speak English or isn’t primarily engaged in
social interactions, this might translate instead into choosing what to do
and how to do it: for instance, in When in Rome 2, a meat-eating alien
from a low-gravity planet might choose to steal and eat a salami, but have
difficulty eating the salami because of its low strength.

Selecting Dialogue
Versu modeled characters who had their own unique dialogue around cer-
tain topics and ideas. Versu’s conversation rules then controlled transitions

* There’s a great deal more to say about dialogue modeling than fits into this chapter, especially as it
isn’t our primary focus here. My blog at emshort.wordpress.com contains a number of posts in the
“conversation modeling” category that deal with particular approaches.
218   ◾    Procedural Generation in Game Design

from one piece of dialogue to another, encouraging characters to stay

more or less on topic.
In one of our unreleased demo projects, we wrote a scene in which any
Versu characters could be brought together to play a game of cards and
chat freely from their corpora of dialogue. One of the characters, Patrick,
was a modern frat bro; another was an early nineteenth-century admiral.
When Patrick told a story about his numerous dates, the admiral followed
up with a story about sexually transmitted disease running rife on his
ships during a Pacific voyage.
Both pieces of dialogue were tagged as pertaining to the topic of sexual
behavior, so the social affordance allowed a transition from one to another
although the characters had never been explicitly written to talk to each
other. This topic structure provided a useful generality that let compo-
nents from different genres interact.
Players who encountered this transition read more into it than was actu-
ally there. The admiral was not intentionally scolding Patrick for sleeping
around or expressing an equivalence between Patrick and his own sailors.
But because the stories were written to be specific to the characters’ expe-
riences and offer some subtextual depth, they seemed to mean more when
juxtaposed.

Layering Dialogue Features

In The Mary Jane of Tomorrow, the procedural system spends little time
working out what to say in a macrosense. There are only a few questions
(on the order of a dozen or so) that the player may ask. Most of the system’s
content focuses on rendering the nuances of the dialogue to demonstrate
the robot’s current settings.
For instance, we can ask the robot to do us a favor, and we get this
response from a robot who is trained in flirtation, enthusiasm, and faux
medieval diction:

The robot crosses her legs. “If thou wishest it, yea!”

The gesture of leg crossing is available because the robot is flirtatious.

She agrees in medieval language because of her diction training, and she
uses an exclamation point because she’s enthusiastic. If we asked when she
was bored instead, we might get

The robot shrugs. “Verily.”

 Characters and Personalities    ◾    219

This is implemented as a grammar. The response to a request consists of

two tokens: [gesture] [consent]. Each of these tokens is then expanded using
a corpus that takes into account mood and diction features. The consent
token breaks down to [consent phrase] [final punctuation], where the final
punctuation will be . if the robot is bored, but ! if the robot is enthusiastic.
Handling all in-game dialogue this way could become overwhelming
fast in a large project; in The Mary Jane of Tomorrow, it works because
there are few initial utterances, and because this level of variability is the
main point of the game experience.
But even in a situation where we need to produce a much higher volume
of content, we can identify dialogue elements that we can usefully treat as
tokens to be expanded in the moment, such as

• Gestures or tone-of-voice indicators that might accompany a speech act.

• Social moves, such as greetings, farewells, compliments, and insults,
that might be included in a longer piece of speech.
• Exclamations.
• Names of other characters or in-game entities, which might be sub-
stituted differently depending on the speaker’s relationship to that
character (“Eunice,” “Dr. Yeung,” “Mom,” or “that genius”).
• Hesitations and framings—If the character has a piece of informa-
tion to convey, do they say it outright or do they frame it with some
hesitation or uncertainty? “I think…,” “I hate to contradict you,
but…,” and so forth, all convey a more cautious personality.

You can also transform generated text after it has been constructed. In
Versu, we had a “drunk character” filter that would create slurred speech
for a character after they’d drunk too much, by converting s- to sh- and
doing other similar conversions. Simple as it was, this effect was a dra-
matic favorite with players.
Stutters, word repetitions, sentence fragmentation, and self-interrup-
tion can also add to the performance, communicating that a character is
nervous, confused, lying, or in some other state of distress.*

* The SpyFeet project of the University of California, Santa Cruz (https://games.soe.ucsc.edu/proj-

ect/spyfeet) digs deeper into some of these forms of speech filtering, and connects these elements
with studies of personality traits in general.
220   ◾    Procedural Generation in Game Design

In Versu, our drunk filter meant we autogenerated profanity when a

character slurred the word sit. It would have been possible to add some
special case checking to make sure that this didn’t occur, but in practice,
we didn’t consider that problem a game killer in the specific case of Versu.
However, as always, it’s a good idea to be ethical about the construction of
output: if you’re altering words before they’re printed, check to make sure
that the generated text is not more offensive than is otherwise appropriate
for your work.

Character and World Interaction

Versu treated its world model more like a stage than like a fully detailed
simulation. The objects in a room were interesting more for their expres-
sive capacity than for any functional purpose—which meant that we often
defined objects with expressive hooks. The main purpose of a vase was
to be smashed in anger. A fireplace might afford the possibility of staring
thoughtfully into the flames or destroying a dangerous letter or stabbing
the coals viciously with a poker. Food at a dinner party could be eaten,
criticized, or (in extreme cases) thrown at other diners.
Providing expressive hooks in the world model augments the sorts
of dialogue-layering techniques described above. Characters seem con-
nected with the space they’re in—and if they’re interacting with props that
the player is also able to move, alter, and destroy, these interactions drive
home that the player is not interacting with something purely scripted.

CREATION
Realization and creation are not independent of one another. A system
that allows dynamic character creation often needs realization techniques
in order to present that character in an interesting way.
But dynamic character creation introduces further constraints, espe-
cially if we intend the player to interact with a character in different ways
depending on the character’s strengths. In this case, designing our vocab-
ulary of character traits becomes a subset of game design, since we are
now in essence constructing a gameplay system with character features as
elements of that system.
Such a system relies on a robust selection of combinable character traits,
skill sets, or behaviors that are

• Orthogonal to one another, or else tagged with information about

mutually exclusive features
 Characters and Personalities    ◾    221

• Mechanically significant
• Easy to communicate to the player individually
• Likely to combine in interesting and evocative ways

Orthogonal
When we say that traits are orthogonal, we’re saying that they apply to dif-
ferent aspects of the character’s life and therefore can be combined freely.
A person’s height is orthogonal to their strength: a person could be tall
and strong, tall and weak, short and strong, or short and weak, without
any inherent contradiction.
On the other hand, if we chose traits such as flirtatious–standoffish
and outgoing–shy, we might find ourselves generating characters who
are both flirtatious and shy. It might be difficult to realize that character
convincingly, since the qualities describe overlapping aspects of a person’s
behavior.
If we really wanted, we could still use those traits, but include a rule
during generation that any generated character who was both flirtatious
and shy should be discarded and rerolled (or restrictively generated in the
first place). However, it’s often easier and more productive to avoid these
problems initially by rethinking the trait axes until we come up with some
that are genuinely orthogonal. For instance, instead of flirtatiousness, we
might measure a character’s interest in romance, giving us a romantic–
aromantic axis to pair with our outgoing–shy axis.
Then we might generate romantic and outgoing characters who would
behave flirtatiously, romantic and shy characters who were hesitant day-
dreamers, aromantic outgoing characters who wanted to be friends with
everyone but date no one, and aromantic shy characters who standoffishly
kept to themselves.
Each of those combinations is interesting and plausible, so we’ve maxi-
mized the descriptiveness of our procedural system.

Mechanically Significant
If a game is heavy in story or meant to be played only once, the fiction
probably carries quite a bit of weight with the player, coloring her experi-
ence of the entire game.
The more the player encounters different scenarios built with the same
ruleset, the more the fictional layer abstracts away unless it is supported by
mechanical content as well. An NPC with an aggressive facial expression
222   ◾    Procedural Generation in Game Design

becomes no different from an NPC with a passive expression unless the

two characters also respond differently to the player.
This is something we discovered in writing characters for the
Austenesque stories in Versu. One of the characters, Mr. Brown, was a
bad-tempered poet loosely based on Lord Byron. We wanted him to be
arrogant, romantic, flirtatious, swift to anger, and a constant source of
irritation.
Initially, we gave him a number of special lines of dialogue to use when
speaking to other characters, allowing him to give flowery speeches that
none of the other characters could use. With repetition, though, the flow-
ery speeches felt like what they were a fictional decoration of the character,
rather than a deep mechanical feature. Brown emerged more powerfully
as an individual character when we translated his dramatic qualities into
a mechanical function instead, and made Brown prefer to escalate situa-
tions. If you insulted him, he’d take maximum offense; if you flirted with
him, he’d come back twice as warmly; if you rejected him, he’d gripe and
sulk.

Easy to Communicate
The more complicated the trait, the less likely it’s going to come across to
the player. Some traits also don’t read well in particular stories because
there are few opportunities to exercise them.
In our Versu stories, a character who was habitually unpleasant to high-
ranking men didn’t read clearly to players: he appeared in a story where
most of the other male characters were rich, and therefore he just seemed
generally surly. There weren’t enough cases where the player could see him
interact with lower-ranking men and perceive a pattern.

Meaningful in Combination
If you present a player with a character description made up of three emo-
tional styles (say, aggressive, friendly, and passive) and three hair colors
(blond, brunette, and black-haired, perhaps), you’ve created a system in
which there are 3 * 3, or 9, possible total combinations.
However, the player will experience this system more as if there were
only 3 + 3, or 6, possible combinations. Because hair color and emotional
style are largely irrelevant to one another, you don’t get a dramatically new
play experience by making a blond aggressive character as opposed to a
black-haired aggressive character. Once players have seen each emotional
 Characters and Personalities    ◾    223

style and each hair color, they’ll feel, correctly, that they’ve exhausted the
system.
Making your elements mechanically significant will help with this, of
course, because then a black-haired aggressive character will (somehow)
play differently than a blond aggressive character. But it’s also productive
to think of elements that resonate with one another as symbols, so that a
pair together takes on a meaning that wasn’t inherently there in the indi-
vidual elements.

RECURRING STRATEGIES

Combining Output from Several Layers of Simulation or Gameplay

In Blood & Laurels, the player could acquire poison from one of the char-
acters and then subsequently poison any NPC in any banquet scene. The
procedural effects of poisoning took hold slowly: at first, the victim would
seem not to react at all, and would go on with what they were already talk-
ing about. But soon afterward, they’d start to choke and sputter, and then
their conversation would be cut off in a dramatic death scene and flurry of
servant reactions.
That delay was part of what made the poison so satisfying to play.
Because you could poison someone at any time in these scenes, and
because the poisoning was interlaced with the rest of the scene’s drama,
it could turn out variously funny, dark, or sad, just as a result of whatever
else happened to be layered with the poisoning death. Cutting instantly to
the poisoning death would have felt much less rich.

Bringing Character into Every Interaction

We want to juxtapose features that come from gameplay context (what
are we currently talking about, and how does this character currently feel
about the protagonist) with long-term character traits (what kind of per-
son is this, and how do they tend to behave in general?)
Writers of screenplays try to write dialogue rich in subtext. Each line
exists to convey information not only about what is happening, but also
about the characters’ relationships and inner life. Dialogue without this
characteristic is called “on the nose”: it’s boring, predictable, and imper-
sonal. In a procedural context, we can layer different aspects of the story
in order to get that feeling of subtext.
224   ◾    Procedural Generation in Game Design

For instance, we might imagine a system where

• The content of the next beat of dialogue is chosen based on the needs
of a prewritten narrative; it’s specific to this one moment and will
never be repeated.
• The emotional content of that beat is chosen based on character rela-
tionships built up via gameplay.
• The text, voice, or graphical rendering of this dialogue and emo-
tional content is determined by the character’s traits and behavior; it
reflects some ongoing realities.

Juxtaposing Events and Interpretation

Players fill in gaps in the narrative with their own explanations of char-
acter motive. This is a tremendously useful opportunity for a designer. In
the simplest case, this means that the player will interpret text output like

The robot sighs. “I’d love to,” she says.

as sarcastic or exasperated, but read

The robot grins. “I’d love to,” she says.

as sincere. If we imagine the tone of voice used in those two performances,

they come out quite different, even though nothing in the text specifi-
cally says that the first tone of voice is sarcastic and the second tone of
voice is genuine. Text is particularly versatile in this respect, but even if
we’re using graphics, we can do some of the same things—for instance, by
combining different gestures with lines of dialogue. The gesture provides
a context for interpreting the dialogue.
The player can build larger bridges as well. Here’s one of the poems the
robot character will compose in The Mary Jane of Tomorrow:

Peasants at the joust

every jaw agape:
defeat is ready to spring.

This robot is trained in medieval diction, so it draws a medieval image

from its corpus of possible images. The final line, meanwhile, is pulled
at random from a separate corpus of proverbs and general situations. By
 Characters and Personalities    ◾    225

putting together a specific, detailed circumstance (peasants gaping) with

a broad interpretive remark (defeat is ready to spring), we invite the player
to come to her own conclusion about what is going on here. Perhaps the
joust is about to go horribly awry; maybe one of the contestants is fighting
badly and in peril of death. But, as the author of these passages, I never
imagined or intended that specific juxtaposition.

Callbacks to Earlier Events

In Versu, if the player did something that caused a character to change
their opinion of the protagonist, the system would not only register a
numerical change in the character’s reaction, but also record a string of
text describing why the NPC had changed their opinion. We could then
reuse the string when the character wanted to report on how they felt
about someone.
For instance, if Fred has decided to insult Miss Bates by saying he
doesn’t have a good opinion of her intelligence, his dialogue might be ren-
dered like this:

Fred: Miss Bates is a fool. She would not stop talking about turtles
at supper.

Callbacks create perceivable consequence: the player can see how ear-
lier events have affected the current state of the system.

PITFALLS

Overgeneralization
A piece of character generation or character behavior code is rarely as
completely universal as one might hope, especially across different genres
or lengths of story.
Versu was designed as a platform for many different stories, and indeed
for many different types of story. The initial suite of stories released were
a murder mystery and a ghost story set in Regency England and reflected
a Jane Austen–derived environment, with characters in some cases lifted
from her books directly and in some cases adapted from similar behavior.
The second set of stories were shorter office comedies set in the modern
day.
We anticipated (correctly) that we would need to create different behav-
ior sets to account for these different genres of story, since a character
226   ◾    Procedural Generation in Game Design

flirting in a modern office setting would likely say things that would be
considered wildly scandalous in the Regency setting.
The less obvious challenge was that we also needed our social actions
to have different impact depending on context. The office comedies were
meant to be farcical, with social situations that rapidly escalated into
more and more ridiculous problems, and gameplay took place over per-
haps half an hour. The Regency stories were meant to be more serious and
dramatic, with character relationships evolving over closer to an hour of
playtime and through many different scenes. The models we developed
for how a character might react to an insult in one were wrongly tuned
for the other.
Multiplying NPC reaction numbers by a scaling factor did not always
resolve this problem, because the Regency social practices were designed
with the idea that characters would spend some time sending social sig-
nals and then confirming that those signals were actually intended. In
the office comedy, by contrast, we needed characters to over- rather than
underinterpret what had been said to them—to start throwing pies right
away, rather than politely trying to clear up any misunderstanding.
In practice, any system of procedural character realization designed for
narrative is likely to be tightly bound to the underlying story genre and
gameplay system. A system created for one context may not easily transfer
to another.

Overrealism
For Versu, we developed a model of the social practice of conversation. It
procedurally expressed ideas about conversation flow taken from stud-
ies of real-life conversational pragmatics: rules such as “it is a violation
of social norms to fail to answer a question you’ve been asked” or “it is
socially normal to keep talking about the same topic that is already under
discussion” or “if there are several conversational partners available, peo-
ple are more likely to respond to those with whom they have the strongest
relationship.”
The latter rule turned out to be realistic but ill-suited to a game. In early
testing of the system, we found that players complained because when
they entered a room, the NPCs in that room might respond to the player
briefly but then go back to talking to one another. When we investigated,
we found that the system was behaving as specified: the player just hap-
pened to be playing a relative stranger walking in on two close friends. The
scenario of the friends acknowledging the newcomer just long enough to
 Characters and Personalities    ◾    227

be polite is true to life, but it violated the players’ expectations of agency

and centrality in the game universe.
When we added a weighting factor to the simulation so that NPCs
would always prefer to interact with a player character, testers said the
results were much more realistic.

Untamed Simulation
In one of the Versu office comedy scenarios we worked on (but never
published), the story included a meeting where all the characters came
together to work on an advertising strategy. In testing, I was taken aback
when two of the characters, who had been flirting all game long, had sex
in the middle of the meeting, in front of their coworkers.
To some extent, this outcome was the fault of not having enough rules
in the simulation: we hadn’t added strong enough controls to indicate that
it was a severe social violation to participate in such intimate behavior in
front of other people, or in a work setting. But it was also a manifestation
of a problem we frequently faced in general, in which the simulation gen-
erated outcomes that damaged the intended behavior of the story.
For Blood & Laurels, we solved this problem by adding scene-by-scene
constraints. In each scene, we would specify which characters were allowed
to die, leave the room, or have sex—those being the extreme points that we
felt could derail the narrative.
In some games, this doesn’t matter. For sandbox games and roguelikes,
there is no pregenerated narrative to destroy; it’s to be expected if all the
dwarfs die in Dwarf Fortress, or if the Sims in Sims 3 are trapped in a
doorless bathroom and starve to death. But in a more narrative context,
we often need to make sure our simulation does not stretch our narrative
framework beyond what we’re prepared to handle.
 IV
The Procedural Future

G ames and software have been procedurally generating interactive

experience for decades, yet at the time of writing, most can agree that
we remain in the infancy of our field’ s potential. While procedurally gen-
erated levels and chatbots are becoming commonplace, the future remains
unknowable. Designers debate whether satisfying stories can be truly gen-
erated, or whether a player’ s perception of content as machine authored
changes its value. Perhaps these questions will be answered definitively,
and perhaps not.
As we’ ve seen throughout this book, procedural generation is a diverse
field even within the confines of game design, as if it resists generalities
by nature. In this final section, we see the nigh-infinite directions we can
continue to explore, whether we create as scientists or as artists. How we
talk about generators, test them, categorize and quantify their properties,
select specific algorithms, or even generate games and generators them-
selves is all touched on here.
Ultimately, the steps forward will be as surprising as they are excit-
ing, and it is difficult to speculate about the next disruptor to the main-
stream conversation about procedural generation. However, the tools used
to achieve that paradigm shift will undoubtedly include those provided
within this section.

229
 Chapter 22

Understanding
the Generated
Gillian Smith
Independent

P ressing the “generate” button and seeing the surprising results

that come from a system you made is one of the joys (and frustrations)
of creating procedural generators. It’s easy to determine the success or fail-
ure of a generator based only on what we witness come from the system
as developers. But unless the generative system is intended to create only
one artifact, at some point we need to understand the generator as a whole:
what types of things it tends to make, whether it makes things that have
sufficiently high quality, or if it has any biases in it that aren’t obvious from
just eyeballing the output manually. As makers of generative systems, we
are making entire spaces of potential content that a player or viewer or
other designer may need to interact with. Instead of needing to do qual-
ity assurance on handcrafted content, we need to do it for every potential
piece of content that could come out of the software we create.
This chapter offers a method for automatically critiquing what is being
generated and, in doing so, better understanding the generator’s capabili-
ties. I’ll be talking about the notion of understanding how expressive a
generator is and some strategies for visually understanding that expres-
siveness. And finally, I’ll close with some general thoughts about how and
why and when it makes sense to grapple with the overall generative space
of a system, rather than individual pieces of content.

231
232   ◾    Procedural Generation in Game Design

EXPRESSIVE RANGE AND GENERATIVE SPACES

Generators are like little minidesigners who operate in a very specific
domain. They have their own style and quirks and strengths and weak-
nesses. A generator’s style is highly sensitive to the data and algorithms
we employ in their creation. Being able to understand this style and how
it changes based on player and/or designer input is a crucial aspect of gen-
erator design.
I like to think about the expressive range of a generative system as the
potential range of content the system might be able to create. A single
instantiation of the system—with a fixed set of parameters and fixed
codebase—has what I’ll be calling a generative space. A system’s expressive
range is made up of many potential generative spaces. When you change
an input parameter or a line of code, you are creating a new generative
space filled with a lot of different pieces of content.
For any given generative space, we might be able to say that a system
can create thousands, even millions, of unique pieces of content that
fill that space. But, how different are those pieces of content from each
other? Is the system biased toward creating certain kinds of content? Are
the few examples you see when you hit the “generate” button represen-
tative of the overall generative space, or accidentally only in one small
corner of it?
It’s impossible to answer these questions by examining small samples
of content that come from a generative system, and looking at every piece
of content is prohibitively expensive. This chapter presents a method for
understanding expressive range automatically, using the Launchpad level
generator as a running example. Launchpad is a parameterized generator
for Mario-like two-dimensional (2D) platforming levels. It supports static
platforms, moving platforms, gaps, springs, enemies that pace on a plat-
form, and stompers.1

QUALITIES OF THE GENERATED

Types of Qualities
If a generative space is a collection of generated content, then let’s begin
by understanding individual pieces of that content. When we look at
generated content ourselves, we are usually judging its value along sev-
eral measures that are custom to the domain we’re working with. The
first step in understanding a particular generative space is to capture
the qualities we care about in pieces of generated content, so that we
 Understanding the Generated    ◾    233

can see how well those qualities are distributed across the space. Maybe
you care about seeing a particular number of challenges in a level, or a
distribution of resources across a map, or a pleasing color scheme. There
isn’t one universal set of qualities to look for, although there are some
types of qualities you can think about for your own system. These types
of qualities are adapted from previous research.2

• Topological: Topological properties are based on the underlying

structure of the content created, independent from how it is skinned.
For a level generator, this might be properties such as the presence of
loops and cycles, or how much the player’s path branches. For a story
generator, it might be the story’s length or the shape of the narrative
arc.
• Experiential: Experiential properties describe how the player will
interact with and experience the generated content. Is there a par-
ticular rhythmic quality to player movement through the space that
you are hoping for, or a method for estimating level difficulty? In a
puzzle game, is there a certainly desirable number of solutions, or
set of characteristics for what a good solution should look like? Are
there particular strategies you want a player to use, and what might
be indicators of those strategies being adopted?
• Aesthetic: Aesthetic properties describe the visual and auditory qual-
ities of the generated content. Qualities like the kind of color palette
used, the extent to which it uses warm or cool colors, and the use of
contrast and salience to highlight particular features are all exam-
ples of artistic properties.
• Semantic: It is important to consider not just the gameplay proper-
ties of content made by generators, but also the real-world, semantic
meaning behind what is created. For example, a character creator
may generate characters from a library of clothing types—how many
of those clothing types read as masculine or feminine?

Example
For some concrete examples, Figure 22.1 shows three example levels from
Launchpad, zoomed all the way out so we can see the entire map at once.
Two qualities that we are interested in based on seeing these levels are
234   ◾    Procedural Generation in Game Design

FIGURE 22.1  Three levels created by Launchpad.

• Linearity: This is a topological property, describing the vertical

change in heights of platforms across an entire level. A level that
is highly linear is one that follows a straight path (be it flat, sloped
upward, or sloped down). A level that is nonlinear is one that follows
a curve, with many changes in vertical orientation. The first two lev-
els in Figure 22.1 have high linearity, as they follow a straight line;
the second has medium linearity.
• Leniency: This is an experiential property, describing how lenient the
level is to players when they fail. A highly lenient level is one with
challenges that do not tend to kill the player if they fail, such as jumps
to new platforms with no gaps underneath them. A nonlenient level
is one that has mostly challenges where the player is likely to come
to significant harm if they fail. The first level in Figure 22.1 has high
leniency, as there is only one challenge that could bring the player
to harm. The next two levels have low leniency, with many gaps that
will kill the player if they fail to jump over them successfully.

Linearity and leniency are the two qualities that will be examples for
the remainder of the chapter, but there are some other example qualities
we could imagine caring about as well:

• Density: An aesthetic quality describing how much of the screen

space is taken up with level elements versus left as open space
• Risk–reward ratio: An experiential quality that describes the average
reward given for each risky maneuver in the level
 Understanding the Generated    ◾    235

Formalizing Qualities into Metrics

In order to automatically assess these qualities and compare many
pieces of generated content using them, we need to develop metrics
that correspond to them. These metrics should be determined using
some algorithm that can give us a score that can be compared with
other pieces of content from the same generator. Often, metrics will be
defined on (or normalized to) a [0, 1] scale, to make it easier to perform
comparisons between individual pieces. Ideally, these metrics will be
efficient to compute for each piece of content, because—as we’ll dis-
cuss in the next section—they will need to be run for many pieces of
content each time you want to assess the current state of the generative
space.
In any formalization of some aesthetic quality into a numerical score,
there will be some aspect of that quality that is not fully captured—and
that’s okay! Our goal here isn’t to come up with the perfect mathematical
definition of fun or beauty or difficulty. Instead, we want to come up with
some formal expression that gives us insight into a very specific quality we
care about. Keeping in mind how well your mathematical definition maps
to the fuzzier concept you are trying to capture will be useful when inter-
preting the generator’s expressive range, as described later in the chapter.

Example
Let’s look at the qualities we pulled out for Launchpad—linearity and leni-
ency—and see how they can be formalized into a metric.

• Linearity: With this metric, we want to capture how much change

in vertical movement there is across a level—whether the player will
feel like they are moving along a series of hills and valleys, or mov-
ing in a straight line. To describe this as a single number, we treat the
midpoint of each platform as a data point, and perform linear regres-
sion to find a line of best fit through the platform midpoints. We can
then calculate how far off from that line each individual platform
is by taking the absolute value of the distance between the actual
height of the platform (y) and the expected height of the platform
if it were to fit on the line (yexp). Summing up all those values and
dividing by the number of platforms (n) gives us a score that tells us,
on average, how far away from the line the platforms are. To boil it
down into an equation, a linearity score for any given level is
236   ◾    Procedural Generation in Game Design

∑
n
y − yexp
p =1
linearity =
n

Because we divide by the number of platforms in the level, this

score can be used to make comparisons between pairs of generated
levels. A larger number means there’s more distance overall from the
line of best fit to the platforms. But to get a number that tells us how
linear a level is in comparison with other potential levels produced
by the same generator, we will need to normalize the score to a [0,
1] range. More on that when we talk about analyzing a generative
space, rather than an individual level.
Note that this number can tell us how linear a level is, but doesn’t
tell us anything about the actual line it follows. A level that is com-
pletely flat will have the same linearity score as a level that follows
a continuous upward or downward slope. For our design purposes,
that was okay. But if we also wanted to know some information about
the kind of line that we’re following, rather than just how well we’re
following it, we could define another metric that uses the slope of the
line of best fit.
• Leniency: Here, we want to create a metric that gives us a sense for
how likely the player is to come to harm in any given level. There
are a lot of factors that could be considered for this, but in the spirit
of keeping with our design goal of creating lightweight metrics that
can be calculated quickly, let’s just consider the hazards that players
encounter and how much harm can come to the player by interacting
with them. Launchpad supports seven major kinds of obstacle that
players interact with: gaps, enemies, long falls, springs, stompers,
moving platforms, and jumps to new platforms with no gap under-
neath. To calculate the leniency of a level, we assign a leniency score
to each of these elements:

Leniency Element
0.0 Gaps, enemies, long falls
0.25 Springs, stompers
0.75 Moving platforms
1.0 Jumps without a gap
 Understanding the Generated    ◾    237

These points are chosen based on design intuition for how harm-
ful the element can be to a player. Gaps, enemies, and long falls have
low leniency because the failure state for interacting with these ele-
ments is death, and we’ve observed them to be the hardest elements
for players to interact with. Jumps with no gap underneath them,
even if technically difficult to overcome, will never be responsible for
player death—thus they are given a high leniency score.
Summing the score for each level component and dividing by the
total number of components gets us a leniency score between 0 and
1 that can be used to compare levels:

∑
n
leniency ( e )
leniency = e =1

n
Again, some information relevant to level leniency is lost in creat-
ing this metric. It does not consider the challenges that come from
ordering or frequency of level elements; for example, a player may be
more likely to come to harm if they face three difficult jumps over
gaps in quick succession, rather than three jumps over gaps that are
spread far apart from each other. The goal with this metric is to pro-
vide a rough estimate based in our design intuition and that will
later be interpreted with this caveat in mind.

Metrics versus Requirements

If you’ve used some generation method that already aims to create desirable
content (perhaps a generate-and-test or optimization-based approach), then
you have already gone through the step of creating some formal, mathemati-
cal definition of desirable features in the content. These may be interesting
metrics to look at on their own: How often does your generator hit the require-
ments you specify for it? But bear in mind that understanding the emergent
qualities of what is created involves examining features that aren’t necessarily
already encoded in the generator. Perhaps some of these metrics will move to
the generator’s fitness function at a later point, if it’s decided that you actually
want all of your generated content to have a particular value for a given metric.
For example, Launchpad uses a generate-and-test approach to creating
levels where it aims to meet specific criteria set by a human designer in
terms of the underlying rhythmic structure. A designer can specify that
they want levels to be made up of fast-paced segments where the player
238   ◾    Procedural Generation in Game Design

takes regular actions, or slow-paced segments where the player has more
time to rest between each challenge. While the way these parameters are
set has an indirect impact on the expressive range of the overall system
and provides different generative spaces for each parameter combination,
the parameters themselves are not directly tied to the metrics of linearity
and leniency.

QUALITIES OF THE GENERATOR

Now that we’ve got a way to automatically assess the quality of individual
pieces of content, we can talk about what qualities we might care about for
the generator itself with respect to the content that it creates.
We can use these metrics to get some insight into the generative space
for an instance of the generator, by considering a generative space to be
the n-dimensional space defined by those metrics. Every piece of content
created by the system can be measured by each metric and plotted some-
where in that generative space. For Launchpad, with its two metrics of lin-
earity and leniency, this means we have a 2D view on the generative space.
With this way of thinking about a generative space, we can now ask
ourselves what kind of qualities we might want to understand with respect
to a particular generative space, as well as the overall expressive range of
the system.

• Variety: We often talk about creating content generators that can

produce a wide variety of content. Now we know what it would mean
to produce high variety within a given generative space: there should
be large coverage across the entire space, with at least one piece of
content at each point in the space. But more importantly, we can
characterize variety. It’s rare for a generator to uniformly produce a
wide variety of content, but we can look at a generative space and see
what kind of variety is produced—perhaps a level generator produces
a high variety of levels with respect to their use of color palettes, but
low variety with respect to a risk–reward ratio.
• Unintentional bias: Related to the sense of uniformity in variety of
content is the notion of consistency. Examining a generative space
lets us see areas where the generator is biased toward creating more
or less content. This may or may not be desirable, but is a helpful
insight to gain before deploying a content generator in a game or
tool. We can gain a sense for how likely a player is to see a generated
 Understanding the Generated    ◾    239

character wearing feminine clothing by looking at the generative

space.
• Responsiveness: Generators can have unpredictable emergent behav-
ior, where a single parameter change propagates and gives unex-
pected results. If our generator permits changing its inputs in some
way (e.g., by altering the data it builds from or twiddling with input
parameters), it is useful to see how the generative space changes as
well, and to build up an understanding of the expressive range of the
generator.
• Content families: Any generator will have a particular style of content
that it typically creates, whether that style is deliberately and care-
fully crafted, or merely an artifact of the generative system. We can
use analysis of a generator’s generative space and expressive range to
understand that style and see if there are particular common “fami-
lies” of content being created: clusters of content that are similar to
each other but far away from other kinds of content.

VISUALIZING EXPRESSIVE RANGE

So we now have a way to think about the generative spaces for our system
and its overall expressive range. What are some useful ways to visualize
that expressive range so we can understand what is being generated? In
this section, I talk about two ways to visualize a generator’s expressive
range: comparing histograms that are cross sections of a particular gen-
erative space, and seeing clusters of similar content.

Histograms
A generative space has dense regions, inhabited by pieces of content that
have similar metrics scores, and sparse regions, where there are few or
even zero pieces of content that will ever be produced with those scores.
A useful way of viewing these regions of differing density is to use a heat-
map-based histogram, where the brightness of a region corresponds to its
density. Visualizing a generative space in this way lets us see at a glance
how varied content is, as well as if there are particular regions that a gener-
ator is biased toward or against. We can easily look at different histograms
side by side to visually scan for changes in the generative space based on
different input data or parameters.
However, visualizing a generative space with a 2D histogram comes
with the trade-off of only letting us view a 2D “slice” of the generative
240   ◾    Procedural Generation in Game Design

space at a time. For visualizing many metrics simultaneously, it may be

worth using a plot, such as a box-and-whiskers diagram, which gives a
high-level overview of the mean, median, and standard deviation of each
metric score at the expense of getting a sense for the overall shape of the
generative space. This kind of visualization can also be useful for rapidly
comparing different versions of the generator side by side.3

Example
Figure 22.2 shows a generative space heatmap for a particular parameter-
ization of Launchpad, produced by generating 10,000 levels and scoring
them with the leniency and linearity metrics described earlier. The linear-
ity metric has been normalized across all generated levels to be on a range
[0, 1], by first calculating all the linearity scores for all the levels, and res-
caling so that the maximum score is 1 and the minimum score is 0.
This particular parameterization generates levels that are chains of
5-second-long, medium-density level segments separated by a resting
platform. The player’s actions follow a regular, rhythmic beat. The his-
togram shows that for this configuration, the generator is biased toward
creating two types of levels, marked by bright green and red areas of the
heatmap. The two types of levels it most commonly creates are levels that
are highly nonlinear with medium leniency and levels that are highly lin-
ear with medium leniency. This heatmap also lets us see that with this

5 regular medium

0.75
Leniency

0.5

0.25

0.25 0.5 0.75

Linearity

FIGURE 22.2  Example generative space histogram for a particular configura-

tion of parameters to the Launchpad level generator.
 Understanding the Generated    ◾    241

configuration of parameters, the generator never makes levels that have a

very high leniency score, and is most likely to create levels with a leniency
score between 0.2 and 0.6.
It’s useful being able to quickly see at a glance what kinds of levels the
system can create. However, the real power of the heatmap visualization
comes where we can compare different visualizations for different con-
figurations of parameters, thus uncovering hidden biases. Figure 22.3
shows a set of generative space histograms for all possible combinations of
the rhythm input parameters for Launchpad—the length of the rhythm,
its type, and its density. These parameters control the “rhythm” that an
ideal player will feel with their hands during gameplay as they hit differ-
ent movement buttons on a controller: from short, regular, fast-paced seg-
ments to long, irregular, slow-paced segments.
Seeing these graphs side by side lets us see some unexpected biasing in
the generator for particular rhythm parameters. In particular, setting the
generator to produce regular rhythms biases the system toward creating
highly linear levels, while setting it to create short swing-beat rhythms
allows it to suddenly create segments with greater variety in leniency. This

FIGURE 22.3  Set of expressive range histograms for Launchpad, each represent-
ing a different configuration of input parameters.
242   ◾    Procedural Generation in Game Design

kind of biasing wouldn’t be possible to see just from a quick visual inspec-
tion of single pieces of content produced by the system, because roughly
the same areas of the expressive range are still covered—just with a very
different frequency in certain regions than before. So hitting the “gener-
ate” button and playing a few levels would show that the same kinds of lev-
els are being created with these parameters, and we would actually need
to play and take note of thousands of levels before we could start to get the
sense that the system is biased.

Distance-Based Clustering
An alternative for visualizing high-dimensional spaces is to determine clus-
ters, or families, of content. While this method takes more implementation
effort up front, it can be a helpful way to visualize over time how different
clusters are changing in size and character. Clusters can show emergent
patterns or families of generated content. For example, Figure 22.4 shows
a cluster map for the Launchpad level generator. Each colored rectangle

FIGURE 22.4  Screenshot of an analysis tool showing clusters of similar levels.

Each square in the top left area corresponds to a cluster of levels, with color cod-
ing for levels that are similar to each other. The bottom panel lets the user scroll
through to see all the levels in that cluster.
 Understanding the Generated    ◾    243

corresponds to a cluster (the size of the rectangle is proportional to how

many levels are in that cluster), and selecting a rectangle shows the levels
that are part of that cluster—the selected cluster in Figure 22.4 shows a
family of upward-sloping levels that have many stompers in them.
In order to perform this kind of cluster analysis, it is necessary to be
able to define the “distance” between two individual levels. In the case
of our example, we defined an “edit distance” between two segments—
the distance between two levels, A and B, is equal to the number of edit
operations that would need to be made to convert level A into level B
(see Ref. 1 for more details on how this was calculated). However, we
could also use the metrics that we defined earlier and define the distance
between levels as the Euclidean distance between them in the generative
space. Any function that can compare two pieces of generated content
and return a number that is small if they are similar and large if they are
dissimilar will do!

CONCLUSION
In this chapter, we’ve seen a systematic method for understanding a gen-
erator’s capabilities in terms of its generative space. Instead of looking at
individual pieces of content by hand, we now have a method for looking
holistically at all pieces of content generated by the system to understand
the full range of what can be created, as well as for diagnosing any unin-
tentional biases in a generative space. While implementing and maintain-
ing metrics and expressive range visualization is additional development
overhead, this method for “understanding the generated” can help with
both the design and development of generative systems. For designers,
having a sense of the range of content a user might see—or not see—can
guide the creation of both the generative system and the larger game con-
text that uses it. For developers, visualizing a generative space and how it
changes during the course of development can help them track down bugs
(e.g., if a small code change suddenly radically alters the generative space)
and understand the emergent ramifications of seemingly small develop-
ment decisions.

REFERENCES
1. Smith, Gillian, Jim Whitehead, Michael Mateas, Mike Treanor, Jameka
March, and Mee Cha. 2011. Launchpad: A rhythm-based level generator for
2D platformers. IEEE Transactions on Computational Intelligence in AI and
Games 3 (1).
244   ◾    Procedural Generation in Game Design

2. Canossa, Alessandro, and Gillian Smith. 2015. Towards a procedural evalu-

ation technique: Metrics for level design. In Proceedings of the Foundations
of Digital Games Conference.
3. Horn, Britton, Steve Dahlskog, Noor Shaker, Gillian Smith, and Julian
Togelius. 2014. A comparative evaluation of procedural level generators
in the Mario AI framework. In Proceedings of the Foundations of Digital
Games. Fort Lauderdale, FL.
 Chapter 23

Content Tools Case Study

Kepa Auwae
Rocketcat Games 

A s discussed in many chapters of this book, handcrafted elements

can create memorable moments in procedural levels and break
player-perceived patterns. But how do you support the creation of those
moments, so that they are influenced nicely by the generator, and prevent
them from becoming repetitive themselves?
In building the level editing tools for Wayward Souls , we were inspired
by Dungeon Crawl: Stone Soup  and Spelunky . The example is for a two-
dimensional (2D) game, but many of these concepts could be applied to a
three-dimensional (3D) game. Through trial and error, we created a highly
functional handcrafted room editor. Our goals for these tools were

• Easy visualization : This is great for coordination, as this way the

editor could be used by anyone on the team. Someone handling the
design could build a basic room based on what would encourage
interesting combat or exploration, and then an artist could load it
up in the editor and do a visual pass. It’ s also great just to make the
development process easier, as it is less cumbersome than typing out
each tile by hand in ASCII into a text file.
• Conditional elements : Some modifications depended on other modifi-
cations to happen or not happen first. For example, we wanted to give
the algorithm a chance to place a pillar in the middle of a room, with a
further chance to place decorations around the pillar with a torch on the
front of it. Nested random chances were ideal to make a modular system
where there were many possibilities that could combine with each other.

245
246   ◾    Procedural Generation in Game Design

• Allow for shaping the room geometry by both addition and subtrac-
tion, based on chance or a conditional being fulfilled. Allow for add-
ing doors or passageways, taking out chunks in the wall. Based on
the previous nested chances, the same room could look very differ-
ent if you saw it multiple times in a single play.
• Allow for easy further scripting, for things like dialogue triggers,
doors being closed or opened by crossing an area, ambush encoun-
ters, cut scenes, and more.
• Be easy to learn for new people. Using layered image files and a
parser for them as a level editing tool means that it’ s easier to visu-
alize and make the modular content. This is again mostly for team
coordination. If we wanted to get a new designer to help make mod-
ular rooms, it would be preferable if they didn’ t have to spend large
chunks of time deciphering complex text files.

We ended up using the usual rooms in Wayward Souls , with the pro-
cedural bits mostly being short hallways. For the current game we’ re
working on, these modular rooms are being attached to and placed in
more purely code-generated areas. This approach is the best of both
worlds. You get freeform generated spaces to explore and maneuver in,
punctuated by a large amount of varied handcrafted modular geometry
(Figure  23.1).

FIGURE  23.1   Generated level.

 Content Tools Case Study    ◾    247

In our previous games, we used a web-based tile level editor, but over
time, we found it was too cumbersome for more complex cases. We had
way too many things in Wayward Souls . Tile types, decorations, monsters,
and other entities would be too much to track. The main tool we ended up
using was Photoshop/GIMP.

SYSTEM OVERVIEW
Every pixel in a map image represented a tile. Every color represented a
different kind of tile. A separate palette file served to associate the colors
to specific tiles.
Data from a layered PSD file would be run through a web-based parser
(Figure  23.2). This processed all the graphic layers and converted them to
tile data, which was stored in an SQLite database on our server. This would
then be assessable in the test build of the game, instantly, for all developers
and testers.
The level image files were layered so that different layers could be flipped
on and off at will. The layer names served as scripting to determine the
conditionals as to when and how often that layer’ s pixels could be drawn
onto the map by the game. The layer name scripting had many different
uses, many of which were compatible to be applied to the same layer:

• Set pixels in that layer to be an entity rather than a tile. Set the pixel
as a specific monster, a breakable piece of furniture, a torch, a door,
or a large art asset.
• Set what percentage chance the layer has to spawn.

FIGURE  23.2   PSD file, ready for processing.

248   ◾    Procedural Generation in Game Design

• Set a name for a group of layers from which only one layer is chosen.
Give the same name to two other layers; the game picks one out of
the three. Have different group names so one layer from each set gets
chosen.
• Set layers as associated with another layer, so it only has a chance to
spawn if the previous layer does, for example, 50% chance to spawn
a torch on a wall that only has a 50% chance to spawn.
• Set pixels drawn on the layer as invisible triggers for the player to
step on, calling a function also set by the layer scripting. Go past this
line, and it will close all doors in the room, and then spawn in waves
of enemies that need to be defeated to open the doors.

EXAMPLE ROOM
Geometry : This is the base room. Every other layer would be drawn
1.
over this layer, should it appear. The geometry is very simple: it’ s
a big square (Figure  23.3). The two light center pixels are possible
spawns of items the player can use. In the corners are spawn points
for possible larger, breakable decorative objects. The darker pixels are
possible spawn points for things that the player can walk over, and
can also be broken in the heat of combat.
Entry : This layer was designed to only appear if a hallway connected
2.
to the room from the east. Not only did it draw a door on the base of
the room, but it also drew spawn points for decorations (Figure  23.4).

FIGURE 23.3   Simple room, in game and in editor.

 Content Tools Case Study    ◾    249

FIGURE  23.4   A new door added to the room, which only appears if the room is
connected from the east.

Part of the wall was indented further out, and another part of the
wall came in to create a shallow alcove in the top right of the room.
3a. Random elements : Now the game chose from a pick-from-one group.
There were three groups to pick from, and the behavior of this one
was to potentially add pillars of various sizes to the room. It picked
that group, and then picked three possible large pillars to spawn
(Figure  23.5).
3b. Random elements : Here’ s the same room, but with a different pick-
from-one group selected (Figure  23.6). Also, new decorations were
randomly picked to be spawned in the room. Also, a new possible
door spawn point was added to the west, above the previous one.

FIGURE  23.5   Large pillars added to the room.

250   ◾    Procedural Generation in Game Design

FIGURE  23.6   Different variant of the same room.

FIGURE  23.7   The room with monsters added, in two different conditions.

Artificial intelligence : Finally, monsters get added to the room by

4.
the game, set by parameters based on the dungeon area and floor
(Figure  23.7). On the right is the same room, but with another set
of different conditionals being chosen at random. With this system,
you could use the same inserted room multiple times on the level,
and it would likely still be different. If you really wanted to avoid
duplicates, you could also track which by chance conditional layers
get used, and put them on a stack that isn’ t used twice.

This could also be applied to other ways to add details and points of inter-
est to code-generated areas in a game. Anything your game supports could
be put into this system. It also makes collaboration easier, improving your
game’ s quality overall. With designers being able to easily add things, you
could have a very large variety of interesting set pieces in an otherwise
code-generated level.
 Chapter 24

Automated Game Tuning

Aaron Isaksen
NYU Game Innovation Lab

As game designers, we use adjustable game parameters to tune a game

to achieve a great player experience. For example, for a two-dimensional
(2D) platformer, we have to tune the size of the character and enemies, how
strong gravity pulls down, how high the player can jump, and so forth. For
a turn-based puzzle game, you might need to tune the probabilities of dif-
ferent pieces arriving on each turn.
Each unique combination of parameter settings creates what we call
a new game variant. These settings directly affect both the designer and
the player, so getting them just right has a major effect on what is possible
in the game, how the game feels, and how enjoyable and difficult a player
might find the game. After all, getting Mario’s jump to feel right is more
about adjusting parameters than coding accurate physics.
Exploring all these possible game variants to find specific settings for
the best player experience was a really time-consuming part of my job as
a game designer. So when I decided to go back to grad school and work on
automated game design techniques and computer-aided game design, my
first project was to come up with a system that could use artificial intel-
ligence (AI) to explore many game variants and report back the difficulty
of each one in a way that matched how human players would feel about the
game. This technique could also find some new interesting game variants
we hadn’t thought of. One of the goals of the project was to make sure the
technique could be used by game programmers that were good coders, but
that they didn’t need to have an advanced degree in computer science or
experience with a lot of mathematics.

251
252   ◾    Procedural Generation in Game Design

Part of what makes this problem so challenging is that there are so

many parameters that can be changed at once, so it’s often impossible to
search exhaustively—imagine adjusting hundreds of independent control
knobs to search for the perfect game. While tuning games, designers rely
on intuition, experience, and user feedback to iteratively search for better
parameters. The designer must estimate player skill; set game parameters;
playtest; evaluate the player experiences using interviews, observations, or
analytics; revise design parameters; and then repeat the process until the
game reaches an appropriate level of difficulty.
Additionally, when a designer and playtesters become experts at the
game, they can lose perspective on how it is experienced by new players.
As the designer repeatedly plays the game, they get better and better, and
the adjustments they make to the game can target those expert players.
This can make the game too difficult or obtuse for new players. So one of
the constraints for our system was to make sure that there was an adjust-
able skill level to control for both low-skilled and high-skilled players.
By changing parameters alone, we can also find surprising and interest-
ing game variants that can inspire new ideas. When working on a game
for a long time, it can be difficult for designers to break out of the current
design and try some creative new ideas. This is especially a problem if all
changes break the game; the designer might need some large changes to
multiple parameters to find something else that works. Algorithms, on the
other hand, don’t get frustrated with temporarily bad results and are will-
ing to keep searching until they find something else worth investigating.
Automated playtesting and visualization help with this process, guid-
ing designers in their exploration to create games best suited to individual
skill levels and play styles.
In this chapter, I describe the method we used to tune game parameters
that can be used by everyday game designers without requiring human
playtesters. Our general approach is shown in Figure 24.1. To explore a
new game variant, we select parameters from a valid range of reasonable
values. Using these parameters, we generate a level and simulate playing
it using an AI.
The trick here is to use an AI that doesn’t try to play the game perfectly,
but that makes the kinds of errors that humans make. Because the AI
doesn’t need to be a champion, you can write a less complex AI than you
might need if competing against human players. I suggest starting with a
simple AI and then add complexity only if needed. We repeat the process
of running simulated players through the game, collecting lots of data.
 Automated Game Tuning    ◾    253

{ pipeGap=7.2,
Select game pipeWidth=3.3,
parameters . . . }

Explore
new game Generate level
parameters

Repeat N times
Optimize for
specific difficulty Simulate level
or
Sample for Record score
visualization in histogram
or After N times

Search for Analyze scores

unique versions

Output d=.372
Repeat as needed difficulty “medium”

FIGURE 24.1  Diagram of how to explore game variants. We explore game space by
simulating game variants and estimating their difficulty using survival analysis.

Then, we look at the data we collected and decide if the game is too easy or
too hard. If it’s too easy, we can make some adjustments to make it harder
and try again. If we feel it’s too hard, we can make it easier. We don’t have
to guess which knobs we should adjust to make it easier or harder, as we
can just try different settings and let the computer quickly figure out the
effect of the changes.
If this overview leaves you with some questions, there are also
deeper details in our publicly available research papers hosted at game.
engineering.nyu.edu/exploring-game-space.

STEP 1: SET PARAMETERS

To begin, we need to decide which game parameters we can adjust, and set
them to some values. For our example in this chapter and for our research,
we decided to work with a game that is popular and has relatively few
adjustable game parameters. We chose Flappy Bird because it is a com-
mercially successful phenomenon, spawning hundreds of similar games,
and the rules are simple to implement and understand.
In Flappy Bird, a player must fly a constantly moving bird without
crashing into a series of pipes placed along the top and bottom of the
screen (Figure 24.2). Each time the player taps the screen, the bird flaps its
254   ◾    Procedural Generation in Game Design

wings, moving upward in an arc while gravity constantly pulls downward.

Each time the bird passes through a pipe gap without crashing, the player
scores a point. Part of the appeal for Flappy Bird is the comically high dif-
ficulty level, especially when typical casual games are easy and forgiving.
Flappy Bird could have been much easier with a few small adjustments,
such as increasing the gap between pipes or decreasing the width of the
pipes, but these would have led to different, potentially less rewarding play
experiences.
Figure 24.2 shows all the parameters that we can adjust to create differ-
ent Flappy Bird variants. The original Flappy Bird, and all our variants, has
a constant value for each parameter during a play session since the game
does not change as the player progresses. However, we’ve also applied our
approach to games that speed up as the player advances, for example, in
games like Canabalt.
The values are self-explanatory from the figure except for “pipe gap
location range,” which is just a number that determines how much the
pipe heights vary from pipe to pipe. Larger numbers mean more changes
from pipe to pipe and are harder because there is more distance to travel
between a high gap and a low gap. By varying the parameter in the figure
within sensible ranges, we can generate all variants of Flappy Bird that use
the same set of rules.

ps: pipe separation

lr: pipe gap location range

pw: pipe width

H: world height

j: jump velocity
bw bh: bird size
pg: pipe gap

v: bird velocity

g: gravity

FIGURE 24.2  Gameplay of Flappy Bird. In Flappy Bird, the player must navi-
gate the bird through a series of pipes without crashing. We modify the labeled
parameters to generate unique game variants.
 Automated Game Tuning    ◾    255

FIGURE 24.3  Two different generated levels given different parameter sets. More
simulations complete the second version, so we can see that it’s easier for the AI
to play. The horizontal lines indicate the target location for the AI to flap.

STEP 2: GENERATE THE GAME LEVEL

Once we’ve set some game parameters, we can generate our new game
variant. This involves placing the bird and pipes in their starting positions
and randomly distributing the pipe gaps. In Figure 24.3, we show two dif-
ferent generated game variants.
Because the levels are generated using a random process, it is important
to generate a new level each time the AI runs, even though the parameters
do not change. Otherwise, if the same random layout of pipe gaps is used
repeatedly, artifacts can arise in our data caused by a particular place-
ment of the gaps. For example, the gap locations can randomly come out
approximately equal for a section of the level, making that section easier.
These problems are averaged out by generating a new level each time.

STEP 3: SIMULATE THE GAME

We now have an AI play the game for us. Accurately simulating human
game play with an AI requires an understanding of how players react to
game events: this is the process of player modeling. The idea is not to make
your AI so good that it can always beat the game, as happens a lot when
people think about AI for games. Instead, we want to model how humans
actually perform—we ultimately care about how humans perform, not the
AIs. As long as the model properly predicts human perception of difficulty
for novice, intermediate, and advanced players, it fits our purposes.
To make the human-like AI, we first begin by creating a simple AI with
perfect motor skills—a perfect player with instantaneous reaction who
would never lose at the original Flappy Bird. Our AI is very simple and
is very fast to run. The AI looks forward to the next pipe and tries to stay
256   ◾    Procedural Generation in Game Design

a bit above the top of the lower pipe, and each time the bird drops below
that target height (the horizontal lines in Figure 24.3) the AI immediately
executes a flap. Whatever AI is used, it should play with very good perfor-
mance on solvable levels, and should mainly only fail on impossible levels,
such as a level with a tiny pipe gap where the bird cannot fit through. You
might want to use an A* or Monte Carlo tree search–based AI for your
game, or write your own simple one, like we did.
We then modify the perfect AI to perform less well by modeling the
main components of human motor skill that impact difficulty in these
types of action games: precision and actions per second.* Adjusting these
values lets us model different player types, since novices react slower and
are less precise than experts. Our model relies on a predetermined value,
which we determined using a user study to figure out how precisely play-
ers can tap an onscreen button, and we can reuse those numbers for other
games too.
When a player plans to press a button at an exact time, they execute this
action with some imprecision. We model this error as a normal distribu-
tion (aka a bell curve), as shown in Figure 24.4. The width of the normal
distribution is proportional to a player’s imprecision. Imprecision is a trait
that all humans have, but it is also related to the time a subject has to react
to an event, called the speed–accuracy trade-off: the less time they have to
react, the less accurately they will respond.
For simplification, our player model assumes precision is an indepen-
dent variable and not dependent on bird speed. We measured precision
as an error, with standard deviation ranging between σp = 35.9 ms and
σp = 61.1 ms, and use this range for game space exploration. Unity, C#,
and JavaScript don’t have built-in functions for generating normally dis-
tributed random numbers (also called Gaussian distributed random num-
bers), but you can search the web for some libraries or functions that will
do this.
Humans can only perform a limited number of accurate button presses
per second. In our user study, we measured an average 7.7 actions per
second, or about once every 130 ms. We also limit our AI to this same
number of actions per second so that it can’t play unrealistically well. We
simplify our model by keeping this constant, although a more complex

* We also modeled reaction time to an unexpected event, but that didn’t matter as much, so I left it
out of this chapter. You can refer to our research papers if you want to implement it for your game.
 Automated Game Tuning    ◾    257

Expert

Novice

Early Perfect Late

timing

FIGURE 24.4  Timing curve graph. Expert players are more precise than novice
players.

model would account for fatigue, since players can’t keep actions per sec-
ond constant for long periods.
This is how we actually coded this approach for use in our system. At
each frame of the simulation, we predict the next time t in the future
when the bird will drop below this ideal flapping location—the ideal
player would flap at exactly this time t. We predict this by just running
a copy of the game as fast as possible without drawing to the screen, and
figuring out how many game loop ticks until the bird will drop below
that line; this takes less than a millisecond to calculate. We then make
the player less precise by adjusting this time t by adding a random nor-
mal distribution with about a 30 ms standard deviation, as shown in
Figure 24.5. By reducing or increasing the width of the bell curve for our
precision model, the AI plays more or less well. To keep the AI from flap-
ping faster than a human could tap, t is limited by the number of actions
per second. If t happens too soon after making another flap, we force t to
be at least 130 ms.
Keep in mind that it’s important that the AIs we are using here don’t
learn or improve. If you make a change and the scores increase, you just
made the game easier for the AIs. You can’t necessarily do that with
human testers, because they are also learning and improving each time
they play the game. So if a human plays two variants and performs better
on the second, it might be that they are just getting more skillful at playing
the game. But AIs that don’t learn are reliable in the sense that they can be
258   ◾    Procedural Generation in Game Design

Early Dead
On time Alive

Late Dead

t-ϵ t t+ϵ

FIGURE 24.5  Effects of variable t on AI performance. Modeling precision by

randomly adjusting the time the simulated player jumps. Moving the jump ear-
lier or later can cause the bird to crash into the lower pipe or upper pipe.

reused repeatedly with the same skill settings to determine how difficulty
changes as you modify game parameters.
For each simulation, we get a score s, equal to the number of pipes that
the AI passes before crashing, and we record each score s for analyzing in
the next step. If the AI reaches a goal score of 20, we store a score of 20
and terminate the run so we do not get stuck simulating easy games where
the AI will never hit a pipe. Although Flappy Bird can theoretically go on
forever, human players tire or get distracted and will eventually make a
terminal mistake, but the AI can truly play forever unless we enforce a
maximum score.
The generate and simulate steps are run repeatedly until we have
enough data samples to adequately analyze. The more times you run the
simulation, the more accurate an answer you’ll get, but it will take longer.
We found doing about 10,000 simulations was plenty, and this only took
about a second for the computer because Flappy Bird is a simple game and
we had our own version that had no graphics to draw. If your game is more
complicated, it will take longer, but you can speed things up by turning
off any graphics rendering and just run the game simulation without any-
thing being drawn. You can also adjust the goal score when you will stop
the simulation—lower goal scores will run faster. The important thing is
to try some different values and see what works best for your game.

STEP 4: ANALYZE THE DATA

Given our game parameters and player model, we can now estimate the
difficulty of the game variant. One way to analyze your game’s difficulty is
to keep track of all the final scores that players get, and then figure out how
likely each score is. We do this using a histogram. A histogram is a type
 Automated Game Tuning    ◾    259

of data structure that keeps track of how many times a value (or a range of
values) has appeared.
To make a histogram, first you keep track of all the final scores in an
array. When you want to figure out the probability of getting each score,
you just go through the array and count how many times each score
appears, and then divide by the total number of scores in the array. If
you’ve got many scores, then you can use bins to keep track of scores in a
range. For example, you might count scores of 49, 10, and 22 as being in
the same bin for scores between 0 and 50.
Figure 24.6 shows some examples of how you can use histograms to
compare different game variants. In Figure 24.6a, we see a pretty hard
game, such as a difficult variant of Flappy Bird. In this game, the most
likely score is 0, the second most likely is 1, and so forth. This means low
scores are very likely and high scores are very unlikely. In Figure 24.6b, we
see the same histogram as in Figure 24.6a, but now compared with a sec-
ond histogram in gray. The gray histogram represents an easier game. We
know this because the probability of getting a low score has gone down,
and the probability of getting a higher score has gone up.
In Figure 24.6c, we see another shape for the histogram, which is rep-
resentative of games that start out easy and then get harder as you score

Hard from the start Gray variant is easier

0.5 0.5
0.4 0.4
Probability

Probability

0.3 0.3
0.2 0.2
0.1 0.1
0.0 0.0
0 2 4 6 8 10 0 2 4 6 8 10
(a) Score (b) Score

Easy at start, gets harder Gray variant is harder

0.18 0.30
0.16 0.25
0.14
Probability

Probability

0.12 0.20
0.10 0.15
0.08
0.06 0.10
0.04 0.05
0.02
0.00 0.00
0 2 4 6 8 10 0 2 4 6 8 10
(c) Score (d) Score

FIGURE 24.6  Examples of final score histograms.

260   ◾    Procedural Generation in Game Design

higher. That gives a low probability of achieving a low score, since the
game starts out easy; a high probability of achieving a medium-valued
score; and a low probability of achieving a high-valued score. In Figure
24.6d, we can see that the game data shown in gray is a harder game than
the game data shown in black. We know this because it is more likely to
get a low score for the gray data than the black data.
Using this histogram, we can use some relatively simple statistics to
determine how difficult the game is. In our research, we used a technique
called survival analysis, which figures out how the score frequencies
change as a player continues in the game. In practice, a simpler method
of just taking the median or mean of the histogram would also be reason-
able and is easier to code. If the median score decreases, then the game is
harder; if the median score increases, then the game is easier.
A couple of situations might come up:

• Impossible games are those games that can’t be played successfully

by any player, no matter how much skill. These games will have a
very low mean score with no cases of high scores in the histogram.
This can happen, for example, if the jump velocity is much stronger
than gravity, so the player crashes into the ceiling, or if the pipe gap
is so small that the bird can’t fit through the gap. We eliminate these
games by first verifying that a perfect player with no time adjustment
can reach a goal score with high frequency.
• Playable games are those games that can be played successfully by
some players but the player won’t get a perfect score. The higher the
mean score achieved by the player, the easier the game. The data in
the histogram tells you how often different scores will be achieved.
• Trivial games exist where all players will be able to achieve the goal
score, assuming a reasonable amount of effort. For example, this can
arise if the pipe gap is very large so it’s unlikely the player will have
trouble fitting through the gap. The player still needs to pay attention
and actively tap the screen, but the game is trivial as long as they are
actively participating. You will find these variants because the histo-
gram will have few low scores, a high mean score, and lots of high
scores. If you have a lot of games that ended at the goal score, you’ll
see that the mean score is quite high, and the histograms will have a
spike at that goal score. In this case, you might find that many play-
ers are able to get very far into your game, and it might be too easy.
 Automated Game Tuning    ◾    261

STEP 5: VISUALIZE THE DATA AND MAKE ADJUSTMENTS

There are a couple ways we can visualize the data we’ve collected, and see
how the game changes depending on the skill of the player (remember that
we can adjust the AI’s skill by making it have more or less error in its time
estimate of when to flap).
Beginning with the original Flappy Bird, we can keep each parameter
fixed and vary just one at a time, exploring a bunch of different values of
that parameter, and estimate the game difficulty of each variant. Figure
24.7 shows a plot of pipe gap versus difficulty for several AIs of different
skill. Each line uses a different value for player precision. Lighter lines in
the figure have a higher standard deviation, and so the AI makes more
errors, and the game is more difficult for the player model. As one expects,
the model predicts that players with less precision will find the same game
more difficult to play, and narrower gaps are harder for everyone.
We can also examine how two parameters are dependent on each other,
which can help designers find interesting relationships. We visualize these
results using dot plots, displaying varying difficulty by the radius and
color saturation of each point.
For example, we see in Figure 24.8 that jump velocity and gravity are
dependent. When gravity is too high or low relative to jump velocity, the
bird crashes into the floor or ceiling. In the middle, gravity and jump
velocity are balanced, and we see that as they increase together, the game
1.00

0.75
Difficulty

0.50

0.25

0.00
20 25 30
Pipe gap

FIGURE 24.7  Line graph of difficulty. Increasing pipe gap decreases difficulty.
Each line shows a different setting for player precision. Lighter lines plot lower
precision, modeling poorer performance, and so are more difficult for the player
model.
262   ◾    Procedural Generation in Game Design

Impossible games
g : gravitational constant 1000
(crash into floor)

es
gam
750 Harder
Medium
ya ble
Easier
Pla
500
Impossible games
Original
(crash into ceiling)
250 Flappy Bird

100 150 200

j : jump velocity

FIGURE 24.8  Graph of parameters that make playable games. Jump velocity and
gravity are related and should be adjusted together.

gets more difficult—faster reaction times are required as the bird is mov-
ing rapidly up and down. Lower values of gravity and jump velocity give
the player more time to react and are easier to play.
By exploring different parameters, the system can find different versions
of the game for you. If you want an easier or harder version, the system can
try a bunch of different parameters to find one easier or harder. Using an
optimizer, you can even have the computer find an exact target difficulty.
We’ve also used this system for exploratory computational creativity,
to find game variants that are as different as possible from existing ver-
sions that have already been explored (Figure 24.9). Unplayable variants
are invalidated, and the system only returns the games that can actually
be played by humans. The designer then can examine the unique games to
find inspiration and new ideas.
Using our system, we have generated interesting and surprising vari-
ants, such as “Frisbee Bird,” which has a very different game feel. This
variant, shown in Figure 24.10, was created by allowing the optimizer to
vary every design parameter, including speed, player width, player height,
jump velocity, gravity, pipe distance, pipe width, and pipe randomness.
The optimizer returned a game with a wide flat bird that moves horizon-
tally very fast but slow vertically and requires short bursts of rapid presses,
followed by a long pause while the bird floats in the air. This unexpected
variant, discovered by our system while using the optimization algorithm,
still relies on the mechanics and rules of the original, but is a significantly
different play experience.
 Automated Game Tuning    ◾    263

3000
Harder

Target difficulty

2000
Gravity

1000

Easier

0
0 100 200 300 400
Jump velocity

FIGURE 24.9  Scatterplot graph of game variants. We can search different game
variants for a target difficulty. Each point indicates a variant tested to find the
target; ´ indicates impossible games; dot size indicates closeness to the target.

FIGURE 24.10  New Flappy Bird variant. “Frisbee Bird” was unexpectedly cre-
ated by our system, an example of computational creativity. We show six simu-
lated paths with jump points indicated by triangles. The player is very wide and
thin, with weak jump velocity and gravity that gives the feeling of floating.

CONCLUSION
I hope this chapter has convinced you that using some simple player mod-
eling and human-like AI players can help you understand the difficulty of
your own game. These techniques can be implemented without any com-
plex libraries or significant experience with statistics. You can analyze the
data using a separate program, perhaps Excel, or doing some simple data
processing in Python (which is how I do it).
The ultimate goal here is not to replace human testers, but to enhance
your game testing with AIs that can reliably play at different skill levels.
264   ◾    Procedural Generation in Game Design

They won’t find the same kinds of issues that human players can find, and
there is no guarantee that AIs will agree with humans about how difficult
the game is. The important part is that you can get a sense of whether your
changes are making the game easier or harder, and that can really speed
up development time when you are tweaking your game parameters.

ACKNOWLEDGMENTS
Special thanks to Andy Nealen, Tanya X. Short, Tarn Adams, Julian
Togelius, Frank Lantz, Dan Gopstein, Steve Swink, Ron Carmel, Mark
Nelson, the NYU Game Center, NYU Game Innovation Lab, and various
fellow researchers, game designers, and reviewers that have influenced this
work and for their helpful suggestions and comments. Finally, I would like
to especially thank Dong Nguyen for making Flappy Bird and for inspir-
ing this research. Images appear here with his permission.
 Chapter 25

Generating Rules
Dr. Michael Cook
Independent 

A s procedural generation becomes more and more important to

people, we look for new ways to make it a part of the things we cre-
ate. Throughout this book, you’ ve already seen a huge variety of ways to
generate content for your games— the artwork, the audio, the worlds your
players walk through, the people they meet, and the problems they are
tasked with solving. Some kinds of content we see generated all the time
in games, like dungeons in a roguelike. Then there are the rarer cases, like
storylines. Then there are kinds of content that are so rare we hardly ever
see them in games at all— like the actual rules your players are following
and playing by.
Depending on the kind of game you’ re making, the rules and mechan-
ics of a game can either define everything about your game or just form a
sort of suggestive space in which the player can freely express themselves.
Think about the block-pushing puzzle game Sokoban , for example— the
player can push blocks by moving into them, but can’ t pull them. They
win by getting all the blocks onto goal tiles. The game is entirely defined
by this set of rules, and by changing the rules, we change the format of
the game. Compare this to a real-world children’ s game like hide-and-go-
seek. The rules here are a setup for creative play, and players might add
their own adjustments to the game according to the situation, who is play-
ing, or what they want to do.
We don’ t see rules or mechanics generated by a game very often, and
it’ s partly because of how fundamental they are. Often, we have a very
strong vision about what we want a game to be like, and altering the rules

265
266   ◾    Procedural Generation in Game Design

too much might completely change that. It’ s also very hard to know how
to even represent things like a game’ s rules— rules don’ t fit into neat tem-
plates like dungeons or quests. They often are unique bits of executable
code that we wrote once to express a certain idea. They’ re also very hard
to test. If a dungeon has a dead end or is a bit boring, it’ s not the end of the
world, and we can try and detect these issues and work around them. If
a ruleset is broken, though, it might make your game unplayable. As any
game designer knows, it can be really hard to see that a game is broken,
even if you’ re a human being.
In this section, we walk through a few basic things you can try out if
you want to play with rule generation in your games. The exciting thing
about this area is that because it’ s relatively unexplored, it’ s a wide-open
playground for new ideas and discoveries— so don’ t be afraid to go off on
your own and try out your own designs, too.

MIX AND MATCH

The process we’ re going to look at breaks rules down into little com-
ponents that we can snap together and rearrange, in the same way that
simple generators snap together parts of mazes, sentences, or songs. The
smallest thing you can start changing are parameters within your game
system. A lot of game designs are only a few changes away from entirely
different game variants— think about all the different match 3 games there
are out there, each with a little tweak. You may have already added lots of
parameters to your game in order to help you balance its design. If you
group all these parameters together into one list, that list defines a variant
of the game. Generating values for each parameter will result in a slightly
different game. Those parameters might include how the player moves,
how much they have to do to finish a level, or how hard the game is to play.
There will be more subtle tweaks than full rule changes, but it’ s a good
place to start.
To really get into rule generation, a nice simple place to start is to build
a simple language you can express rules in without writing code. The sim-
pler the rule is, the easier it is to build a language for. One of the best
things to start with is something like collision— when two things touch
each other, what should the game do? We can do this by having a list of
game objects and a list of simple effects. The list of game objects won’ t have
things like “ enemy”  or “ health”  in it, though. Instead, it should be more
general descriptions like “ turtle”  or “ mushroom,”  because we won’ t know
what’ s an enemy or what heals the player until after the rules have been
 Generating Rules    ◾    267

fixed! For the effects of colliding, we can think about the general events
we want to happen in our game. Can the player gain score? Can the player
die? Can the player get power-ups? All these bits of code become separate
effects.
All our generator has to do then is pick two object types and one effect,
and it’ s got a little collision rule generated. It’ s our job to convert it into
real game code, though, but usually it’ s just a case of having lots of code to
check each building block and point it toward some code you’ ve already
written.
This format of rule generation is really easy to expand by adding in
more options for the generator. We could add extra elements to our rules,
for example. Maybe instead of collisions only having one effect, they have
two effects, one for each of the objects involved in the collision. Now you
can do things like have one object kill another object and survive (like
PAC-MAN  ghosts), or have effects only target one object. We can also add
more types of rules, like rules for when things leave the screen or rules for
when a button is pressed on the keyboard. To get more variety out of this
technique, it’ s important to add things that you might think are useless or
not interesting. Don’ t just add obvious game objects to the collision object
list. Add in the floor. Add in background objects. Add the user interface
in— anything you can think of. If you only put predictable things into a
system like this, you’ re likely to only get predictable things out the other
end.

GAME DESIGN 101

This system of rules and building blocks we’ ve described is sort of a min-
iature slice of game designing, which your generator is exploring by gen-
erating rules. The most obvious way to write your generator is to make it
select rule components entirely at random, maybe generating three or four
of these random rules before adding them to the game. As you can prob-
ably imagine, a lot of these rules combinations will be bad or boring or,
in extreme cases, completely unplayable. A totally random rule generator
isn’ t as good a game designer as you, which won’ t come as a big surprise.
You can help it get better by encoding some of your knowledge about
game design into the system. You’ ve already done a bit of it simply by
writing the rules language— every object you selected and every effect you
wrote down, you decided to add them in because you suspect they will be
relevant to game design. You didn’ t let it adjust the volume of the music, or
the resolution of the screen. You didn’ t let it edit memory values directly
268   ◾    Procedural Generation in Game Design

or delete files from the player’ s computer (at least, you probably didn’ t).
You decided that some things would be better for games than others, and
that’ s an example of you helping the system design games.
But we can do better— we can write heuristics , little rules of thumb or
guidelines for the system to follow that will help it create better rulesets.
These are things that we believe to be true about our rules regardless of
anything else going on in the game. Some of these guidelines we know to
be universally true, no matter what rules come up. For example, if we’ re
making a Candy Crush – style game, there has to be at least one rule that
destroys blocks. Our game is about destroying blocks, so if there aren’ t
any rules that destroy blocks, we can safely say that the ruleset isn’ t very
good, and our generator can scrap it and go back to the drawing board.
These heuristics might seem very simple, but they can be very power-
ful. Even a basic one like “ Can the player do anything?”  can cut out a huge
percentage of bad ruleset designs. Sometimes we can’ t know enough up
front, though. Much like designing a game for yourself, you can’ t tell what
a game will be like just by looking at all the rules written down— that’ s
why people playtest their games or give them to others to play. If we could
get our generator to playtest the rulesets it makes, we could maybe take the
data from those playtests and write even better heuristics.
Right now, there’ s a lot of research funding being put into a field called
general game playing  (GGP). The aim is to develop artificial intelligence
(AI) capable of playing any game, even one it’ s never seen before, by rap-
idly learning rules and strategies. One of the uses for AI like this is to
automatically evaluate games and investigate any problems the designs
might have. It’ s a long road toward that goal, though, and you might not
feel like developing cutting-edge AI systems just to play with rule genera-
tion. Instead of using the cleverest AI possible to play our games and test
them, what we can do instead is use the stupidest  approach possible, and
get our games played entirely randomly by a system that has no idea what
it’ s doing.
This might sound like a waste of time, but even a random agent tells
you something about how the game works. Suppose we run a random AI
player on our generated ruleset 10 times. Our agent can record things as it
plays— we can get it to record if it completes the game, perhaps, or when-
ever it gains score or touches an object. At the end, we have 10 samples of
how the game looks. Suppose we’ re making a game like PAC-MAN . If all
10 of our random players never died, or if they all completed the level in
under 3 seconds, there might be something wrong with our ruleset. We
 Generating Rules    ◾    269

can write more heuristics, more guidelines, that use this new data to think
about what it says about our game. Our random player isn’ t clever, but it
represents a kind of worst-case scenario for gameplay, which is a step up
from just looking at the rules without any kind of gameplay.

THROWING OUT THE RULEBOOK

Lots of people have experimented with rule generation in the past, from
inventing new chess games to making completely automated game
designers, but it’ s still a very new field without many accepted ideas. The
little system we designed, with a series of building blocks for rules, guide-
lines to filter out bad ideas, and a random player to test the game for us, is
a really basic recipe for experimenting with rules. But the exciting thing
is that no one really knows what the best ideas are for creating rules,
mechanics, and systems using procedural generation. There are amazing
new games, entirely new genres, out there waiting to be found. So what-
ever ideas you have about this topic, there’ s a good chance no one’ s ever
tried it before. There’ s only one way to find out how good they are— go
and give them a go!
 Chapter 26

Algorithms and
Approaches
Brian Bucklew
Freehold Games

T he intent of this chapter is to serve as your first-time hiker’s map

of the most well-worn and pedestrian trails in the vast (and mostly
unmapped) wilds of procedural generation. If it does its job, you should
come away with a couple basic tools. First, an idea of where you should
be able to get started without too elaborate a preparation. Second, a set of
landmarks that almost anyone you meet in the field will know.
However, like any public park pamphlet, it won’t attempt to fill in the
vast tracts between marked trails. Eventually, you’ll find that once you
gain confidence and skill, the most rewarding adventure is found among
the brambles.

RANDOM NUMBERS

Pseudorandom Number Generators

Since the goal of most procedural systems is to present a varied experience
from player to player, and from session to session, random numbers play
an important role. Although it’s possible to produce varied choices in a
completely deterministic manner, the most common approach is to use a
series of random numbers that are different each time the game is played
in order to produce the desired variety.
Almost all development environments provide one or more random
number generators to use. These are almost always pseudorandom number

271
272   ◾    Procedural Generation in Game Design

generators. The “pseudo” in this case refers to the fact that these typically
do not return a truly random set of numbers, but only a series of numbers
that appear to be random.
Pseudorandom number generators start from an initial seed value and
then perform a mathematical operation on that seed value that produces
the next value in the random series and increments the seed in some way
so the following call will produce a different (but statistically random)
value. Whenever you give a pseudorandom number generator a particular
starting seed, it will always produce the same sequence of random values.
For example, a particular pseudorandom number generator that is
given the initial seed 5 and gives you 3, 23, 195, and 1 when asked for the
first four random numbers in its series will always return 3, 23, 195, and 1
as the first four numbers when its initial seed is 5. The same is true of all
remaining numbers in the series it generates for the seed value 5.
If that particular pseudorandom number generator is given a seed value
different than 5, the series of numbers it would return when asked would
likewise be different. For example, if you gave that exact same pseudoran-
dom number generator an initial seed of 6 (instead of 5), it might give you
55,129, 4, 51, and 0 (instead of 3, 23, 195, and 1) when asked for the first
four numbers of its sequence.

Making Use of Repeatable Series

The fact that the series of numbers a pseudorandom number generator
generates is completely predictable when it is starting with a particular
initial seed has a number of powerful applications for procedural systems.

• Two players on completely independent computers can share proce-

durally generated worlds by sharing a single seed value. The same “ran-
dom” choices can be made on two completely different computers, no
matter how many choices are needed, as long as those choices are made
in the same order. Many roguelikes, including Caves of Qud, use this
property for daily or weekly challenges to be shared across otherwise
disconnected players. They need only share the same seed value. We’ll
discuss how to generate those shared seed values in the next section.
Many multiplayer games with procedural aspects use pseudoran-
dom series to reduce the amount of information that needs to be
shared over limited network bandwidth. Small seeds can be shared
instead of large swaths of generated data.
 Algorithms and Approaches    ◾    273

Large worlds with information that can be regenerated from seeds

do not need to save a potentially large (or even infinite) amount of
data to disk but can instead regenerate complex systems from those
small initial seeds. The original Elite used pseudorandom series to
generate (and regenerate when revisited) an enormous, apparently
permanent, explorable universe on the very memory-constrained
computer systems of the time (http://www.filfre.net/2013/12/elite/).

Seeds and Hashing

The seed for a typical random number generator is a 32-bit integer. This
value, by default, is most often a simple current timestamp for a random
number generator whose output you aren’t interested in repeating or shar-
ing. This will cause each execution of the program to produce different
output each time it is run. For random number generators whose output
you would like to repeat or those seeds which you’d like to share, a differ-
ent, more structured, strategy is required.
A raw 32-bit integer like 1235100523 is not a particularly memorable or
shareable value. A much more useful method that many games take is to
allow the player to enter a more memorable phrase like “a magic monkey.”
Since the pseudorandom number generators require a 32-bit integer, and
not an arbitrary string, you need some method to convert each different
phrase into a different 32-bit integer.
If you’re technically minded, you might already be imagining a method
to squish a string into a 32-bit integer. Since it’s applicable in very many
areas of computing outside of this discussion, this is actually a well-stud-
ied problem. This process, of taking an input (which may be much longer
than the desired seed value) and creating a value based on it, which is dif-
ferent for each input value, is called a hash function.
A hash function’s job is to take some input data and transform it into
an output value (typically, but not always, shorter in length than the input
data). This output value needs to be the same for any given input. That is,
whenever you give a hash function the value “watermelon,” if the particu-
lar hash function you’re using returns 81, then it should always return 81
every time you give it “watermelon.” Giving the same hash function the
value “lemons” or “watermelon2” should return a value that is statistically
unrelated to the value that you get for “watermelon.”
That doesn’t necessarily mean it won’t return 81 for “lemons”—it may—
but for a good hash, the chances of it being 81 for both “watermelon” and
“lemons” is vanishingly small (e.g., 1 in 2^32 for a good 32-bit hash).
274   ◾    Procedural Generation in Game Design

Another helpful use of the fact that the hash values are statistically
random is that you can use as much, or as little, of the hash function’s
return value as you want since a smaller chunk of bits pulled from a per-
fectly random number are themselves perfectly random. This means if a
hash function generates 128 bits of data, and you want a 32-bit integer as
a pseudorandom number generator seed, you can simply grab the first 32
bits of the value the hash function returns to use as your random number
generator seed.
There are many common hash functions, including MD5, the SHA
family, xxHash, and Murmer. Some, like SHA, produce more statistically
random output, which is useful for cryptography, and some, like xxHash,
produce less statistically random output but with much higher perfor-
mance. Games typically do not need the statistical strength that the slower
hash families provide, so a typical choice for game projects is a fast hash
function like MD5 or xxHash.
Examples of hash functions in use are

• A game allows a player to enter an easily memorable string like

“watermelons,” which is then hashed via MD5 into a 32-bit integer
for the world generation’s pseudorandom number generator.
• A game generates a string describing the current date, like “June 7,
2016,” which is hashed by MD5 into a 32-bit integer to generate a
daily challenge that is the same across all players for any particular
day.
• A game generates a string combining the current world’s seed and
a specific solar system’s identifying information: “Worldseed51234
Galaxy 5 Solar System 118,” which is hashed via xxHash into a 32-bit
integer. This provides a unique seed for that solar system, which
would allow the solar system to be regenerated again in exactly the
same way if the player later visits it again without storing all the pro-
cedurally generated information about the solar system.

Rolling Dice
It is common to talk about random numbers in terms of rolling a number
of dice. These dice are not restricted to the familiar six-sided cube dice.
“1d6” means the result of one six-sided die, which you might find in any
common board game, resulting in a value between 1 and 6. “1d8” means
 Algorithms and Approaches    ◾    275

the result of one eight-sided die. “3d6” means the result of three six-sided
dice added together, resulting in a value from 3 to 18.
Some pseudorandom number generators will only generate an integer
between 0 and the largest representable integer. Since you often want to
generate smaller ranges of values to simulate things like die rolls, it’s very
often useful to generate a range of integers between two given numbers
(say, 1–6 for a d6).
Although many random number generator libraries give you the abil-
ity to specify a minimum and a maximum value for the next number you
generate, if they don’t, you can write your own:

RangeInclusive(max,min)
{
return ((rng.next() % (max-min+1)) + min);
}

It can also be useful to generate a floating-point value between two

given numbers, instead of an integer:

FloatingPointRange(max,min)
{
Return ((rng.next() / MAX_INTEGER) * (max-min)) +

min;
}

Examples of dice in use are

• d100s are commonly used as a percentile success roll, checking if

that roll is less than or equal to the percentage chance of something
happening.
• Small dice pools can be used for any numbers that need to have
highly tunable ranges, for example, weapon damage. One die pro-
duces an equal chance of any number in the range, whereas multiple
dice create a bell curve, making numbers in the middle of the pos-
sible range more likely.

Normal Distributions
It’s often useful to generate a normal (“bell curve” or Gaussian) distribution
of numbers instead of a linear one. Adding several dice of the same type
276   ◾    Procedural Generation in Game Design

together will generate a curve that approximates a normal distribution.

The more dice you add, the closer you get to a truly normal distribution.
A bell curve can be useful for generating many types of content. A bell
curve in one dimension can approximate the natural variety of properties
that many types of things can have, with the lowest and highest values for
those properties being much rarer than the middle values (Figures 26.1
and 26.2).

FIGURE 26.1  A normal distribution in two dimensions can approximate a clus-

ter of plants, or a marksman shooting at a target.

FIGURE 26.2  A normal distribution in three dimensions can approximate a

whole globular cluster.
 Algorithms and Approaches    ◾    277

A normal distribution is described with two terms, the mean and the
standard deviation. The mean is the value at the “peak” of the bell curve.
The standard deviation describes how “wide” the curve is. In a truly nor-
mal distribution, about 68.27% of the values fall within one multiple of
the standard deviation and about 99.73% of the values fall within three
multiples of the standard deviation.
Here’s a basic implementation that will generate a normal distribution
of random numbers with a given mean and standard deviation (std).

GaussianRandomNumber(mean,std)
{
A = rng.nextfloat(); // from 0 to 1
B = rng.nextfloat(); // from 0 to 1
C = SquareRoot(-2 * Log(A)) * Sin(2 * PI * B);
return mean + (std * C);
}

Other, faster methods are known, like the Ziggurat algorithm, if higher
performance is required.
Examples of normal distribution use are

• Generating a random temperature for a particular day, with the

seasonal average at the median, uncommon possible hot and cold
days
• Deciding what time a nonplayer character (NPC) prefers to go to
bed, with 10:00 p.m. at the median but a wide standard deviation,
creating smaller populations of morning larks and night owls
• Deciding the size of a planet in a solar system, with a typical Earth-
size planet at the median but rarer very large and very small planets
possible

Weighted Distributions
Another very useful tool is the weighted distribution. For example, you
might want a common result 9 out of 10 times, and a rare result on the
10th. You can think of this weighted distribution like a bag with nine balls
marked “common” and one ball marked “rare” in a bag. The weight for
each item in a weighted distribution describes the number of balls in the
bag with that item’s result.
278   ◾    Procedural Generation in Game Design

Result Weight
Common 9
Rare 1

WeightedResult(mean,std)
{
Roll = rng.RangeInclusive(1,sumOfAllWeights);
WeightAccumulator = 0;
foreach( currentItem in Items )
{
WeightAccumulator →
WeightAccumulator+currentItem.Weight;
if( WeightAccumulator >= Roll ) return
currentItem.Result;
}
}

A useful variant of this algorithm simulates not returning the ball back
to bag after drawing it, preventing it from being drawn again in future
picks. In order to accomplish this, you simply reduce the weight of the
selected item by one after it is selected.

WeightedResultRemovePickedItem(mean,std)
{
Roll = rng.RangeInclusive(1,sumOfAllWeights);
WeightAccumulator = 0;
foreach( currentItem in Items )
{
WeightAccumulator →
WeightAccumulator+currentItem.Weight;
if( WeightAccumulator >= Roll )
{
result = currentItem;
currentItem.Weight--;
return result;
}
}
}
 Algorithms and Approaches    ◾    279

Examples of weighted distributions in use are

• Picking the template for a random encounter from a list of possible

hand-created templates, with high weight for common encounters
containing a few grunts and skirmishers and low weight for a rare
special encounter with a powerful leader and his retinue.
• Selecting the material for a random weapon, with common ore like
iron having a very high weight and rare diamond swords having a
very low weight.
• Selecting the items for a shopkeeper, with common items having
a high weight and rare items having a low weight. A large table
of items might be possible, and the shopkeeper would pull balls
from that bag without replacing them until his shop inventory is
full.

HEIGHTMAPS
The shape of a lake, the layout of continents, clumps of weeds, and a topo-
graphical map are just a few of the many kinds of map features that are
common in procedurally generated environments. All of them require the
generation of a two-dimensional (2D) grid of values that vary with some
degree of smoothness from cell to cell. These types of 2D grids of smoothly
varying values are called heightmaps (Figure 26.3).

Box Linear Filters

A box linear filter (or box blur) provides a means to create a smooth
heightmap with a controllable method of placing the peaks. A box blur
starts with a grid with cells initialized to a set of values. Typically, a cluster
of high initial values is placed around areas that should contain peaks in
the final result. These initial values are then smoothed across the grid by
setting the value of each cell to the average value of its neighbors. This
smoothing is applied a number of times, with each pass “spreading” the
initial peak values out across the grid.
A matrix of weights is often used to weight the contributions of each
cell’s neighbors to the average in order to control the shape of the resulting
280   ◾    Procedural Generation in Game Design

FIGURE 26.3  Example heightmap created with Perlin noise layering. Whiter
pixels have higher values.

smoothed area. This is an example set of weights that creates a “hill-like”

shape when applied to each cell of a grid:

1 3 1
3 5 3
1 3 1

Midpoint Displacement
Square-diamond midpoint displacement generates a heightmap by start-
ing with random values for the corners of a grid. The center grid square
of that square (M in the grid below) is then given a value of the aver-
age of those corner values plus or minus a small random variance. Then
the middle grid of each edge (E in the grid below) is given a value of the
average of the two corners and midpoint plus or minus a small random
variance. This creates four subsquares, and the process continues on these
subsquares iteratively until the whole grid is populated. Typically, during
 Algorithms and Approaches    ◾    281

each iterative step the amount of variance applied to each new point is
reduced exponentially.
The corners are given a random value:

2 E 8

E M E

10 E 5

The center cell is given a value equal to the average of the corners plus
or minus a small random amount:

2 8

10 5

The midpoint of each edge is given a value equal to the average of the
corners and midpoint plus or minus a small random amount:

2 8 8

5 9 6

10 4 5

This continues recursively for each of the four new subsquares.

Perlin and Simplex Noise

Perlin and simplex noise are algorithms that produce gradients that have
rolling waves that appear at similar rates in every direction you look, with
valleys of 0 and peaks of 1. You can visualize this looking something like
an egg crate foam pad, with the peaks randomly placed but statistically
evenly distributed across the whole surface. These noise functions are not
282   ◾    Procedural Generation in Game Design

limited to 2D heightmaps, but can also produce one-dimensional (1D),

three-dimensional (3D), and higher dimensionalities of noise (Figure 26.4).
Because most surfaces don’t have an egg crate–like topography, it’s
typical to combine several layers of noise, each having a different fre-
quency and amplitude. The first layer is usually a fairly high-amplitude,
low-frequency layer, with big peaks and valleys that are very broad. You
can imagine this as laying out the overall mountain ranges and valleys.
Additional layers that are each higher frequency but lower amplitude
are added. Since they are lower amplitude, they don’t affect the overall
structural height as much, but the higher frequency allows them to add
interesting detail to the existing peaks and valleys. You can imagine the
highest-frequency, lowest-amplitude noise layers creating little potholes in
the road and anthills on the ground of the larger formations that were laid
down by the earlier layers.
Simplex noise has many advantages over Perlin noise. It’s faster to gen-
erate. It is visually isotropic in all directions, whereas Perlin noise has
some directional artifacts. However, its popularity is curtailed because it
is patented under U.S. Patent 6,867,776. A similar open-source algorithm

FIGURE 26.4  Example of a single 2D layer of Perlin noise; whiter pixels have
higher values.
 Algorithms and Approaches    ◾    283

called OpenSimplex was developed in 2014 that shares many of the posi-
tive qualities of simplex noise without patent encumbrance.

SEQUENCE GENERATION

Lindenmayer Systems (L-Systems)

L-systems produce a fractal-like sequence by taking a seed string and
repeatedly applying a set of transformation rules to that string (Figure 26.5).

Input string: A
Rule 1. (A→AB)
Rule 2. (B→A)
Initial value: A
(Iteration 1: A becomes AB via rule 1.)
AB
(Iteration 2: A becomes AB via rule 1; B becomes A via rule 2.)
ABA
(Iteration 3: A becomes AB via rule 1; B becomes A via rule 2; A becomes
AB via rule 1.)
ABAAB

FIGURE 26.5  Weeds generated by drawing instructions generated by L-systems.

284   ◾    Procedural Generation in Game Design

(Etc.)
ABAABABA
ABAABABAABAAB

The individual elements can be interpreted in a variety of ways. One

common approach is to transform each element into a different type of
input to a drawing program. For example, A = move forward 1 step, B =
turn right 45°.
Examples of L-systems in use are

• Generating road divisions and building placement sequence for a

city layout
• Generating the branching headwaters of a river system
• Generating the branching road systems and satellite villages around
a major trade hub

Markov Chains
Markov chains produce a sequence by utilizing a directed graph of nodes.
Each edge has a percentage weight assigned to it. The result of a Markov
chain is generated by starting a particular node. That node’s value becomes
the first value of the sequence.
The next node is chosen randomly from among all the edges directed
away from the current node. That next node’s value is used as the second
value of the sequence. Then the third node is chosen from the edges lead-
ing away from that node, and the chain continues in that way until the
sequence reaches some termination condition. That condition might be,
for example, a particular length of sequence or a node that’s flagged as an
end node.
In order to generate a sequence of weather for a week, we can use the
simple Markov chain shown in Figure 26.6.
Let’s say we randomly choose to begin on the rainy node. So the first
day in our week is rainy.
The two edges leading away from rainy are a 25% chance of the rainy
node and a 75% chance of the sunny node. Let’s say we choose randomly
between those two options and choose to move to the sunny node, so the
second day of our week is sunny.
 Algorithms and Approaches    ◾    285

25%
Rainy

75%
25%

Sunny

75%

FIGURE 26.6  Simple Markov chain sequence generating a sequence of rainy and
sunny days.

The two edges leading away from sunny are rainy with a 25% chance
and sunny with a 75% chance. We choose one of those two edges at random
to generate the weather for our third day. We continue moving between
these two nodes until we’ve generated seven days of weather, and then our
sequence is complete!
If you have an existing sequence of data, you can construct a Markov chain
from it in order to create more data that shares similar stochastic properties.
Let’s say we have a sequence of the letters ABCD. Each of these letters
becomes one of the nodes in our chain.

ABCAABACBACDB

We then look at each letter and determine the number of times each
node is followed by each of the others.

A is followed by A, B × 2, and C × 2 (A 20%, B 40%, C 40%)

B is followed by C and A × 3 (C 25%, A 75%)
C is followed by A, B, and D (A, B, and C 33.3%)
D is followed by B (B 100%)

With that information, we can construct our graph.

286   ◾    Procedural Generation in Game Design

20% 40%

A 75% B

25%
40% 33.3% 100%

33.3%

C D

33.3%

FIGURE 26.7  Chain constructed from the sequence ABCAABACBACDB.

The example chains in Figure 26.7 are “order 1,” which means each sin-
gle element is used as a node in the chain. The order number is the number
of elements per node. An order 2 Markov chain would operate on pairs
of elements instead of single elements. So in Figure 26.7, our nodes in a
second-order chain would be AB, BC, CA, and so forth.
Examples of Markov chains in use are

• Generating music from note patterns

• Simulating the volatile mood swings of dwarves
• Generating random recipes

FILLING SPACE
Random Walks
Random walks are produced by starting at a given point and then tak-
ing steps in random directions. Many natural processes, like molecular
motion, can be described by random walks, and they can be useful for
generating all sorts of naturalistic paths and features.

One-Dimensional Random Walks

The simple 1D random walk shown in Figure 26.8 is generated by moving
randomly either up or down a single unit for each step to the right. This
simple technique could be used, for example, to produce a sequence of
hills and valleys for a platform game.
 Algorithms and Approaches    ◾    287

FIGURE 26.8  A 1D random walk.

FIGURE 26.9  Two steps in a 2D random walk, first east then north.

Two-Dimensional Random Walks

Two-dimensional random walks are produced by starting at a specific
location and then stepping a random direction. In Figure 26.9, each step
can be either north, south, east, or west.
It’s very common to make many attempts to generate a path via a
random walk and simply discard those paths that don’t meet your nec-
essary criteria (Figure 26.10). For example, when using a random walk
in order to generate a river system, you might discard any path that
moves into a cell that is higher in altitude than it started, or that crosses
a desert tile.
A powerful method of maze generation is a series of loop-erased ran-
dom walks (Wilson’s algorithm). During a loop-erased random walk,
any time a walk crosses itself during generation, that walk is thrown out
entirely and a new walk is attempted.
To generate a maze via loop-erased random walk, first a start and an end
point are selected. A loop-erased random walk is then made between the
start and end points. Then a random point on that random walk is selected
and a new loop-erased random walk is made until that walk reaches any
point that has already been walked by a previous walk. Again, if that walk
crosses itself while it is being generated, it is thrown out and a new one is
started. This process continues until all the grid spaces that belong to the
maze are filled by random walks.
288   ◾    Procedural Generation in Game Design

FIGURE 26.10  A 2D random walk after hundreds of steps.

Examples of random walks in use are

• Creating river systems between mountains and bodies of water

• Creating road systems between points of interest
• Creating systems of caverns
• Connecting the doors of rooms generated by a separate method
• Creating sewer layouts

Cellular Automata
Cellular automata are a broad category of systems that operate on a graph
of discrete cells. They define a set of states for each cell and a set of rules
that describe how the state of each cell changes based on the state of adja-
cent cells. These rules are typically (but not always) applied to every cell
in a stepwise fashion, with each cell computing the next state based on its
 Algorithms and Approaches    ◾    289

current neighborhood, and then all the cells changing in one step to the
next state defined by the rules.
A common and familiar cellular automaton is Conway’s Game of Life,
which is played out on a regular 2D grid with two states for each cell, alive
and dead. The rules for each time step are described in terms of the cur-
rent state of the cell and how many of the four cardinal direction neigh-
bors are in the alive state.
Rules for Conway’s Game of Life

Current State + Alive Neighbors New State

Live + 0 or 1 alive neighbors Dead
Live + 2 or 3 alive neighbors Alive
Live + 3 or 4 alive neighbors Dead
Dead + 3 alive neighbors Alive

An example of a simple cellular automaton that is commonly used to

generate procedural areas is a simple system that is able to generate cave-
like maps from an initial random grid of walls. This cellular automaton
operates on a 2D grid like Conway’s Game of Life. The cells can be either
walls or open space, and the grid is typically initialized with some per-
centage of random wall cells.
Example Rules for a Simple 2D Cave Map Generator

Current State + Wall Neighbors New State

Open + 6–8 wall neighbors Wall
Wall + 0–4 wall neighbors Open

These rules, when applied repeatedly, tend to smooth out the stand-
alone walls and concentrate the denser clusters, resulting in a cave-like
appearance.
Figure 26.11 shows a cave map being iteratively generated by cellular
automata in Caves of Qud. The field is generated with a two-wide border,
and then each interior cell has a 55% chance to contain a wall. The cave
rules described above are then executed twice on that random field, result-
ing in a cave-like set of chambers.
Examples of cellular automata in use are

• Simulating growth patterns for spreading plants and forests

• Simulating spreading fire
290   ◾    Procedural Generation in Game Design

FIGURE 26.11  Stage of cellular automata level generation.

 Algorithms and Approaches    ◾    291

• Simulating migration and proliferation patterns for populations of

animals

Settling
A settling algorithm takes a set of assorted shapes that are generated
with some overlap. Those shapes are given simple physics rigid body
representations, and a physics simulation that causes the overlapping
shapes to “push” away from each other is run until they no longer over-
lap. This provides a straightforward means to arrange a large set of pieces
with widely varying sizes and shapes into a set that is connected but not
overlapping.
Examples of settling in use are

• Generating a pile of randomly sized dungeon rooms with larger

physics colliders than the interior rooms themselves and allowing
them to settle out, and then connecting them with hallways
• Generating a pile of very randomly shaped areas and allowing them
to settle out, creating a cave system
• Generating a pile of straight and elbow pieces and allowing them to
settle out into a sewer system

Wang Tiles
Wang tiles are a mechanism for defining a set of tiles and the way they
each connect to each other side by side as they are laid down across a
plane. They can be imagined as squares with a color for each side. The tiles
are placed so that the color of each side of each newly placed tile matches
the color of the sides of each other tile that it touches.
In games, these colors often signify a particular kind of connection
between tiles. For example, “blue” tile sides might be those with a door
along that side, and “green” tile sides those with a hallway, so that when
tiles are aligned based on the color, they produce a properly connected set
of areas.
One particularly interesting subset of Wang tiles for procedural games
is those carefully select sets of tiles that can tile a plane without creating
repeating patterns (aperiodic tilings) (Figure 26.12). Not all possible com-
binations of tile number and side color will produce aperiodic tiling, but
many such sets have been discovered. These aperiodic sets can be as small
as 11 tiles (Figure 26.13) (http://arxiv.org/abs/1506.06492).
292   ◾    Procedural Generation in Game Design

FIGURE 26.12  A 13-tile Wang tile set that will not repeat its pattern across a
plane.

FIGURE 26.13  Example of nine tiles from the set tiled properly in a 3 × 3 block.

Examples of Wang tiles in use are

• The placement of hand-built tiles with predefined edge connections

to form large 2D platformer worlds
• Placing hand-built detail tiles inside a room to add detail to the room
• Using Wang tiles to create nonrepeating variation in plant type and
tree density in a larger forested area, with each Wang tile having a
hand-designed population and density of plants and trees
 Algorithms and Approaches    ◾    293

PARTITIONING SPACE
It’s often the case in generating procedural content that you have an open,
undifferentiated space that you need to partition into a set of regions.
This might be, for example, dividing a landmass or star map into political
areas, dividing that country into states, or taking a space and generating a
set of units for pathfinding or decision making.

Binary Space Partition

A binary space partition takes a given space and splits it in half, and then
takes the two areas that were created and splits those in half, and repeats
until some threshold is reached. Some commonly used thresholds are a
particular number of splits performed and the average size of the resulting
areas (Figures 26.14 through 26.17).
Caves of Qud builds the internal structure of many types of ruins by
using a simple binary space partition of that space (Figure 26.18). A simple
volume outline for the building is created, and then a random column
or row in the building is selected and a wall is created, splitting the area
in two. Each time the area is split, a door is placed that connects the two
spaces along the splitting line.
Since we know cutting an area in half and placing a door that connects
both creates two areas that are definitely connected, and since we know
each additional split of those resulting subareas ensures the resulting two

FIGURE 26.14  Step 1: A region is split by a wall along a random column.

294   ◾    Procedural Generation in Game Design

1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2

FIGURE 26.15  Step 2: A portal is placed at random somewhere along the wall.
The newly created subregions are marked 1 and 2.
3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4
4 4 4 4 4 4
1 1 1 1 1 1 1 1 1 1 1 4 4 4 4 4 4
1 1 1 1 1 1 1 1 1 1 1 4 4 4 4 4 4
1 1 1 1 1 1 1 1 1 1 1 4 4 4 4 4 4
1 1 1 1 1 1 1 1 1 1 1 4 4 4 4 4 4
1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2

FIGURE 26.16  Step 3: Each subregion is now split at random and a portal
placed along the wall, typically restricted to a place where it won’t create a T or
X junction.

7 7 7 7 7 3 3 3 3 3 4 4 4 8 8
7 7 7 7 7 3 3 3 3 3 4 4 4 8 8
7 7 7 7 7 3 3 3 3 3 4 4 4 8 8
7 7 7 7 7 3 3 3 3 3 4 4 4 8 8
4 4 4 8 8
5 5 5 1 1 1 1 1 1 1 4 4 4 8 8
5 5 5 1 1 1 1 1 1 1 4 4 4 8 8
5 5 5 1 1 1 1 1 1 1 4 4 4 8 8
5 5 5 1 1 1 1 1 1 1 4 4 4 8 8
5 5 5 1 1 1 1 1 1 1
5 5 5 1 1 1 1 1 1 1 2 2 6 6 6
5 5 5 1 1 1 1 1 1 1 2 2 6 6 6
5 5 5 1 1 1 1 1 1 1 2 2 6 6 6
5 5 5 1 1 1 1 1 1 1 2 2 6 6 6
5 5 5 1 1 1 1 1 1 1 2 2 6 6 6

FIGURE 26.17  Step 4: After a third round of partition, the area starts to resemble
a set of connected rooms.
 Algorithms and Approaches    ◾    295

FIGURE 26.18  Set of ruins in Caves of Qud, generated primarily via binary space
partition.

areas are themselves connected, we can be sure the process guarantees

that all the rooms created during the space partition will be reachable
from one another.
Examples of binary space partitions in use are

• To create a traditional Rogue-style room layout, you can place a

smaller rectangular room inside the borders of each generated
region and connecting to the adjacent regions with hallways based
on portals.
• Filling an area with walls and doors to form connected rooms.

Voronoi Diagrams
A Voronoi diagram is constructed by taking a set of seed points on a
plane and partitioning that plane so that the region containing each seed
contains only the points on the plane closest to the seed (Figure 26.19).
A variety of distance metrics can be used.
Since Voronoi diagrams work to partition a continuous plane of space,
they are most useful for partitioning continuous, uninterrupted spaces,
like continents; however, there are variations of Voronoi diagrams that
take obstacles into account.
Examples of Voronoi diagrams in use are

• Generating the areas of farmed fields from the placement of farm-

houses in a countryside
296   ◾    Procedural Generation in Game Design

FIGURE 26.19  Simple Voronoi diagram.

• Generating territorial control of dragons from the placement of their

lairs
• Generating areas of varying encounter difficulty from the placement
of major dungeons

Dijkstra Maps
Dijkstra maps partition space in a manner similar to that of Voronoi dia-
grams; however, they work on a discrete graph instead of a continuous
plane. In our examples, we use a square grid, which is a graph where each
node is connected to its four or eight neighbors, but the algorithm sup-
ports any graph.
A Dijkstra map is generated by taking a set of seed points and calculat-
ing the distance from each node in the graph to the closest seed point. This
is typically done via a breadth-first search from each seed point. During
the search, each node is typically marked with the seed point it is closest
to and the distance to that seed, forming a region surrounding each seed.
Dijkstra maps have wide applicability to terrain analysis and pathfind-
ing. Caves of Qud utilizes Dijkstra maps with randomly placed seeds in
order to generate a uniform population of uneven spaces (Figure 26.20).
Random seeds are placed on the map until the largest region reaches a
target size. The resulting Dijkstra map regions are each populated with a
similar number of objects, providing an even distribution of map contents
irrespective of the map’s layout or amount of open space.
 Algorithms and Approaches    ◾    297

FIGURE 26.20  A 20-seed Dijkstra map regionalization of a cave system in Caves

of Qud used for terrain analysis and to create an evenly distributed population.

Examples of Dijkstra maps in use are

• Finding the furthest point from the entrance of a cave system to

place treasure or a boss encounter
• Evenly distributing encounters throughout a nonuniform space
• Analyzing areas to place internal features, for example, a set of square
buildings or tents in a nonuniform space
• Filling national territories around a set of capital cities on a map

Tree Mapping
Tree mapping is a technique used to take a list of values and then parti-
tion a rectangular 2D field into a set of rectangular division (Figure 26.21).
Each of the rectangular divisions has an area proportional to one of the
target values in the list. There are several algorithms for producing tree
maps that trade off a consistent visual aspect ratio of the divisions with
the stability of the map layout when changes are introduced to the list of
values.
Examples of tree maps in use are

• Creating a basic city street grid for a set of buildings and city blocks
of known size
• Creating the rooms of a side-view building given a list of room sizes
298   ◾    Procedural Generation in Game Design

FIGURE 26.21  Typical tree map.

PUTTING IT ALL TOGETHER

Let’s use some of the techniques described above to generate an adorable
little galaxy!
First, we need all the stars in a galaxy. We can use a normal distribution
in 3D to generate the positions of every star in a little globular cluster of a
few thousand stars—or a few million stars if we have a little more time to
work! While we work, we’ll assign each star a random seed. We might just
give each one a totally random value, or we might base it on the physical
location in the globular cluster, hashing the position to produce a seed.
We don’t yet have to generate most of the information about the star,
since we can now generate it based on each star’s seed whenever we need it.
However, we probably want to generate just enough information about the
stars to display them. Probably the size and color of the star are enough.
We can use a little weighted distribution for each star in our galaxy to
pick the type, temperature, and size.

Weight Star Type Temperature (K) Size

1 Super giant 1a 3,000–30,000 Mega huge!
1 Super giant 1b 3,000–30,000 Huge!
2 Bright giant 3,000–30,000 Big
5 Giant 3,000–30,000 Big
20 Main sequence 1,000–30,000 Medium
2 White dwarf 4,000–30,000 Small
 Algorithms and Approaches    ◾    299

So now we know the size of each star and we can pick the display color
based on the temperature.
When the player clicks on a star, we can seed the random number
generator with a value like (starseed+hash(“starname”)) and then use a
Markov chain of syllables we’ve created from a catalog of star names to
generate a random name for the star. This is fast enough that we can just
do it in real time whenever we mouse over a star.
When the player zooms into the system, we can continue to add detail.
A weighted table to tell us if it has one or more companion stars, perhaps a
weighted table for the number of inner rocky planets, and then one for the
outer gas planets. Then we can select a seed for each planet, perhaps a hash
of the system name and orbital (“Algotauri 4”) and generate the overall
visual aspects of the planet for display.
When the player zooms in to a planet, we can generate more detail, for
example, using Perlin noise layers to generate the topography, L-systems
to generate the layouts of the cities of the creatures that occupy it, binary
partition to create the rooms in an individual building, a weighted table
to tell you what the contents of a particular room are, and a Markov chain
to tell you what is on the notepad that is in a desk in that particular room.
Procedural generation opens endless doors in endless hallways to end-
less untracked worlds.
Happy hiking!
 Chapter 27

Meaning
Dr. Mark R. Johnson
Independent

H ow might randomness be given meaning? Can the placement of

elements of a procedurally generated game ever have the kind of
meaning present in other games, especially those notorious for deep nar-
rative significance in the nature and location of every component, such
as the Souls series? Can meaningful thematic elements be offered up to a
player through procedural generation? How might deep and implicit con-
nections between procedurally placed game elements be forged? Will the
worlds and spaces and gameplay challenges created by procedural genera-
tion ever be able to transmit the same kind of meaning as a handmade
game?
An increasing number of games are moving toward offering complex
stories and depicting detailed worlds, fundamentally dependent on the
creation of meaning. Although roguelikes and other procedural games are
known primarily for their mechanics and gameplay rather than these less
formal elements, there is no reason that designers of procedural games
need to fall behind in this aspect of game design. Procedural generation is
absolutely capable of producing meaning at a number of scales and with
varying levels of designer effort and involvement, depending on the scale
and scope of the desired result.
In this chapter, we look at how meaning is created, the concept of
meaninglessness, and two methods of how meaning can be produced
during play. Players can assign meaning to elements lacking any designer-
intended meaning, just as players can readily fail to note the deeper mean-
ing of game content specifically placed by designers with the purpose of

301
302   ◾    Procedural Generation in Game Design

creating meaning. Lastly, we’ll explore the procedural creation of mean-

ing in Ultima Ratio Regum (URR) as a case study¸ the most fundamental
underlying component of the game’s design. In the process, we’ll exam-
ine the design goal of emphasizing meaning over more traditional aspects
of game design, the use of meaning as a game mechanic and its intrigu-
ing relationship to the practice of seeking out game “spoilers,” and what
I believe the procedural generation of meaning can bring to game design
in the future.

MEANING IN GAMES
What is meaning? On the most highly abstracted level, meaning in games
(and any work of art, fiction, or media) is created through forging connec-
tions between elements of that experience. A phrase has meaning when
the reader remembers another character uttering the same phrase before;
a company name has meaning when the viewer recalls seeing their logo
on a van earlier in the film; the visual design of a boss or enemy has mean-
ing when it is similar or comparable to others seen previously. When this
is not the case, in-media information serves only a role as background
or scene setting. Background is an important role, without doubt, but
does not produce elements that are truly meaningful to the experience
of the work, because they could be exchanged for other comparable ele-
ments (switching the logo on one van seen for a moment with a different
logo) and the salient elements of the media artifact would go unchanged,
because that element has nothing to do with the message being conveyed.
Meaning is fundamentally created through the connection of otherwise
disparate elements.
A common source of meaning in games is through the distribution of
information about what now tends to be called “lore.” This is often textual
in-game content that is optional to the player’s success over the mechan-
ical challenges of the game, but serves instead to flesh out the fictional
world in which the action takes place.
In procedural games, this takes many forms. A game like Tales of
Maj’Eyal adheres to a model of scattered lore that adds up to a large body
of data about a story, and a story can appear at first to be one thing and
then morph into something else as more information is acquired. Dungeon
Crawl Stone Soup instead draws on classical mythology for its lore, posi-
tioning its monsters within various mythologies, ranging from ancient
Greek to Egyptian and Jewish to Hindu, in many cases giving hints about
how these monsters behave based on these classical origins. The Binding
 Meaning    ◾    303

of Isaac offers puzzling fragments of a story that appear gradually over

many playthroughs, and slowly lend increasingly substantial meaning to
the different items and monsters found within the game’s dungeons, epis-
temologically repositioning “mere” grotesque monsters and unsettling
body horror world design to become part of a deeper, darker, and more
profound story about (depending on your interpretation) religion, child
abuse, parenting, and imaginary worlds. In each of these cases, meaning
exists outside the formal mechanics of the game, as provided by extra nar-
rative and thematic information, and this information both provides an
understanding of a story or a setting and changes how the player thinks
about those same mechanical elements in their transition from an isolated
gameplay element to a component of a greater whole.
Investigation of such lore is contingent upon a kind of “deal” made
with the player—players must be persuaded that there is sufficient mean-
ing there to actually merit their time. One could readily imagine a game
that procedurally generates cryptic descriptions and utterances of the sort
found in Souls games, without having the underlying system or hidden
narratives required to actually tie these together into a greater thematic
whole. Players would discover that these clues actually meant nothing
and were just selections of words and phrases selected by an algorithm
for their potential for obfuscation, instead of hinting the player toward
deeper truths about the game world, and would lose interest in attempting
to uncover the game’s meaning—because the signifiers of supposed mean-
ing are actually meaningless. There is therefore no meaning to be found
by examining them.
However, such forms of meaning do not always only add to the game,
as in some cases they can almost serve the opposite role and highlight the
constrained nature of a game’s world, rather than (or as well as) bringing a
player’s attention to the wider thematic or historical elements of a fictional
space.
As Erik Champion has previously noted in his work Critical Gaming:
Interactive History and Virtual Heritage, an amusing example can be
found in the Elder Scrolls games. Although the game’s books and the
ambient comments of nonplayer characters do much to flesh out the game
world and its history, factions and cultures, there is a rather odd observa-
tion to be made: the world described in the game’s books is almost infi-
nitely richer than the game actually experienced by the player. There are
vast numbers of events and occurrences in the game’s lore books that can-
not be experienced by the player, and reading all the books can come to
304   ◾    Procedural Generation in Game Design

make the quite repetitive core gameplay of the games themselves feel sur-
prisingly paltry in contrast to the rich world hinted at, but never directly
experienced. Lore in this case does bring meaning and background to the
game world, but it also undermines that same meaning by highlighting
the artificiality of the game world and the limited numbers of interac-
tions that can take place in the game’s possibility space, and reminding the
player that what they read is pure fiction, one step disconnected from the
possible interactions in the world they’re currently exploring.
Meaning is therefore about striking a balance between bringing together
all the elements within a particular fictional world to come to resemble
more than the sum of their parts and avoiding bringing to the player’s
attention what isn’t there—a meaningful game world must be somehow
hermetically sealed, giving out just the right amount of information. This
is a difficult balance that few games get exactly right, showing that mean-
ing is a complex question that extends into all elements of game design,
from the mechanical to the thematic and the narratological to the experi-
ential, as embodied by each individual player.

MEANINGLESSNESS IN GAMES
Meaningless occurrences in games take many forms, but here I’d like
to focus on particular kinds in specifically procedural generation—the
unnecessary item and the mostly uninteresting map. In games with very
explicit and linear differences in the quality of items, where there are rarely
interesting decisions to be made about using Weapon A over Weapon B
in Context C, for example, then acquiring Weapon A after Weapon B,
if Weapon B is undeniably stronger and there are no contexts in which
Weapon A will actually be better, makes the acquisition of Weapon A an
effectively meaningless moment, and turns Weapon A into a meaningless
item (from a gameplay mechanics perspective).
This meaninglessness is not inherent to the item, but rather an emer-
gent property from the elements of the context in which it was acquired—
Weapon A might have been deeply meaningful if found at the start of the
game and in the process giving the player’s character a significant boost,
and it was an item the player had become used to using for hours and hours
of subsequent gameplay, but at the end of the game, it becomes merely an
unnecessary addition. This is perhaps most marked in the Angband family
of roguelikes—in many of these games, late-game item decisions become
almost impossible without actively ignoring or removing from consider-
ation a large volume of the acquired loot.
 Meaning    ◾    305

This problem of the meaningless item arises often in roguelikes and

other games that utilize procedural generation, in which items have their
statistics (semi-)randomized, and therefore only a small portion will be
relevant to the player. Some games try to “tilt” the generation of items
toward the player character’s strengths and weaknesses in an attempt to
raise the volume of found items that will actually have some meaning to
the player, but there will always be procedurally generated elements that
lack any meaning to the player in the current gameplay context.
Similarly, on a game with a large expanse of procedural terrain, much of
the terrain is only present to split up points of interest, such as settlements,
dungeons, or events. Several games, including URR (as noted later), have
developed various ambient elements of the game world (weather, flora,
fauna, random events, etc.) designed to alleviate the disinterest stemming
from these large tracts of uninteresting land. It is nevertheless the case
that large volumes of these game spaces are meaningless or, at best, derive
their only meaning from other elements surrounding them, rather than
any gameplay experience they themselves offer.
In my upcoming theoretical monograph on unpredictability in games,
I call this the “granularity of interest” possessed by a game, which is to say
the volume of experienced or present content that is actually intriguing,
relevant, and meaningful to a player. If almost all elements are in some
way interesting and meaningful, then the game has an extremely high
granularity of interest, since the player doesn’t have to look far from one
point of interest to find another, for points of interest are closely packed.
By contrast, if large volumes of the game space serve little direct purpose
other than spacing out the interesting or meaningful aspects, then the
game has a low granularity of interest.
All games that integrate procedural content generation can be usefully
understood on this scale, as granularity of interest affects the ebb and flow
of player action and engagement in the game, ranging from a consistent
high level of engagement and interest (high granularity) to a sinusoidal
wave of gameplay between moments of rest and moments of interest, more
intense as a result of the previous rest (low granularity).
A low granularity of interest therefore does not necessarily mean that
the entire game, viewed as a whole, is automatically less interesting than
one with a high granularity of interest; indeed, this is a possibility explic-
itly taken by the developers of the procedural galaxy game No Man’s Sky.
Rather than attempting to make every single generated planet interesting
and every single region of every single planet interesting, they have instead
306   ◾    Procedural Generation in Game Design

accepted that many regions of many planets will be relatively uninterest-

ing to the player, and that in many cases entire planets will lack any real
interest for the player. However, they present this as a positive—the more
worlds are bland and ordinarily, the more that the interesting and unique
worlds will actually come to stand out. As exploration is one of the game’s
central draws, it appears that they instead rely on the hopeful appearance
of a kind of emergent player behavior, whereby players find interesting
worlds and share them with others, and the presence of uninteresting
worlds gives compelling meaning to this exploratory pursuit of interest-
ing worlds, and their sharing with other players.
It is often difficult to create handmade worlds where every instant is
interesting, and such a problem certainly remains present when dealing
with procedural content generation. Meaninglessness is therefore hard to
avoid in procedural generation, but can be adapted to one’s advantage if
presented as a contrast that highlights the meaningful elements, rather
than irrelevant aspects of a game experience that, in certain contexts, will
simply never be of note.

DESIGNER AND PLAYER MEANING

We can perceive two methods by which meaning can be experienced by a
game’s player—either it can be specifically placed by a game’s designer and
then recognized by the player, or elements not intended to have any mean-
ing can be seen as possessing meaning by a player, whether “correctly” or
spuriously.
Many games utilizing procedural generation have a large database of
fixed content that is placed first (or placed with the highest priority), and
then procedural content is set up around it. Excellent examples of this
approach include fixed characters and bosses in narrative-driven rogue-
likes like Ancient Domains of Mystery and Tales of May’Ejal, and the
long chain of secret actions in Spelunky—reaching the final boss involves
combining specific items, locating a particular area, dying in a particular
place, and so forth. This sequence remains the same regardless of what the
game’s algorithmic generation systems produce, much like the bosses and
important characters in other roguelikes, and systems of this sort enable
designers to author points of meaning and importance into a game while
still making the overwhelming majority of the game procedurally gener-
ated. This system requires that the rest of the content works “around” it,
as the secret actions must always work and elements must spawn every
time—they are not randomly available or unavailable.
 Meaning    ◾    307

A related but distinct system to the one outlined above is the creation
of handmade optional content segments worked into otherwise proce-
dural areas. In addition to Chapter 7, you may also take, for example, the
“vaults” in Dungeon Crawl Stone Soup parlance. In Dungeon Crawl Stone
Soup, the vaults range from unique and distinctive gameplay challenges
to areas designed to give the player a little bit of information about the
game’s fictional universe. One might encounter a shrine to a particularly
bloodthirsty god with the remains of sacrifices strewn around it, or a vault
designed to tell a particular story, such as the tragic and rather amusing
tale of the deceased bakers, which I encourage all readers to go and seek
out. Handmade segments of gameplay such as vaults are typically nones-
sential, will appear on some playthroughs of a procedural content gen-
eration game but not others, are generally used to develop thematic or
“background” elements rather than the immediate narrative of a game,
and are placed on top of the game’s procedural generation systems as an
optional and unpredictable extra.
There are also situations in which players will assign their own meaning
to particular events. In Chapter 13, we explored briefly how players could
come to regard actions of particular artificial intelligence (AI) actors in
some roguelikes as being far more “intelligent” than they actually were.
Similarly, in a playthrough of a roguelike where the player character has
always been weak and been unlucky with procedurally generated items,
for example, the sudden discovery of a tremendously powerful weapon
or highly effective piece of armor will immediately transform the player’s
experience of that playthrough and come to be seen as a moment of great
meaning in that playthrough.
This is very visible when we look at the player-created practice of writ-
ing “Yet Another Stupid Death” or “Yet Another Victory Post” forum
posts that offer handwritten narratives of a character’s eventual success
or death within a permadeath roguelike. These player stories assign their
own meaning to particular stages of the game that had no designer inten-
tion and were entirely or predominantly procedural, but which came to
strongly affect the outcome of the game, such as gaining a powerful item,
defeating a particular enemy, or making a critical error.
Meaning is also often assigned in procedural games through encour-
aging players to take greater risks to achieve greater rewards, and these
playthroughs—when successful—are moments for great celebration and
pride. In Dungeon Crawl Stone Soup, for example, the player must acquire
three “runes” to win, but there are 15 available runes in total, and a victory
308   ◾    Procedural Generation in Game Design

with more runes is seen as more prestigious than one with fewer. NetHack,
similarly, has a range of “conducts”—such as the vegetarian conduct that
forbids the player character from consuming meat. Practices of this sort
are another way to give even greater weight and meaning to a player’s deci-
sions and to the eventual outcome of victory, if achieved, and to mark out
one playthrough of a roguelike from all the other playthroughs that sur-
round it.
Such practices are rare in player cultures surrounding non procedurally
generated games (such as speed running and high score tables), whereas it
is hard to find a roguelike without a wealth of player-created (or designer-
created) extra challenges that give players the feeling that a particular
playthrough in a particular style was especially meaningful, and represen-
tative of their skill, knowledge, and abilities in the game in question. These
many forms of meaning combine to offer players a range of “anchors” for
their experiences in these randomized worlds, providing ample sources of
meaning and consequence for their activities, and undermining the popu-
lar idea that randomness is necessarily devoid of any deeper importance.

MEANING IN QUALITATIVE PROCEDURAL GENERATION

The creation of meaning is a central part of my own game design work,
and I’d like to close this chapter with a critical reflection on my own suc-
cesses (and failures) in this area thus far, and how I’ve sought to build on
previous concepts of both designer-assigned and player-assigned meaning
to develop a system for procedurally generating worlds rich with meaning,
and making player understanding of that same meaning a central game-
play mechanic. URR absolutely hinges on the creation of meaning, but in
a way distinct from the techniques outlined in the earlier sections.
Fundamental to URR’s world-building processes, and the creation of
meaning within them, is the idea of “chains of meaning.” Nothing (or at
least, very little) is intended to appear in the world that is only related to
itself, or only related to one other element.
For example, flora and fauna are not fixed, but rather procedurally gen-
erated in such a way as to be meaningful to other elements. There are huge
databases of options, in which there are various options that enable and
disable other options, for the generation of wildlife and plant life in each
region of the planet, and modeling their spread. Thus, each region then
ends up with its own distinctive selection of animals and plants, each of
which comes with a procedurally generated image based on its “abstract”
traits, and a name. These serve to orient the players as they move around
 Meaning    ◾    309

the world, and also to produce extra variety as the players move around
the planet. Games that generate huge procedural worlds often struggle
with making different areas of the world feel different, and although ter-
rain (temperate, desert, tundra, etc.) goes a long way, URR deploys this
procedural flora and fauna as an additional detail that varies widely as the
player moves around the world.
The crucial point, however, is that animals and plants (just one example
of a vast number of elements I could have selected for this discussion) do
not only serve this purpose—they also create and influence meaning in a
range of other in-game elements. The covers of books might depict local
animals. Healers in the area will speak of local plants and their regen-
erative properties. Those encountered in distant lands might speak of the
national animals or plants of their homeland.
Perhaps one of the most striking and effective examples of this system
thus far implemented is the generation and reproduction of shared aesthet-
ics distinct to each (secular) culture and each religion (with some bleeding
of visual styles and practices between the two). Everything within a par-
ticular culture is designed to share commonalities of shape and color—
everything from chairs to shop signs and city layouts to ornate floor tilings
in one civilization will share a particular geometric form (squares, octa-
gons, circles, etc.), while everything in a religion, from religious vestments
to altars and prayer mats to holy books, shares particular sets of colors
reproduced in different but clearly connected ways across each item.
The point of these systems is to show the player the meaning in the
generated world, instead of telling them about it. When viewing a sigil
embossed on a shield, for example, the game never says, “This is the coat
of arms of House B.” Instead, the world is simply presented to the player,
along with the understanding that nothing in the game world is discon-
nected, and that meaning is there to be found if the player is able to iden-
tify it, and the player is left to do the rest.
If these elements were generated merely as background informa-
tion, however, their meaning would be limited, no matter how detailed.
However, as well as the technical and artistic interest and challenge of
developing a system of this sort, it is also arguably the most fundamental
gameplay system in the game. The player’s central objective is the uncov-
ering of a global conspiracy, clues to which are distributed through the
world’s cultures: a hint might be found in the corner of a painting, or a
particular line in a particular novel, or hidden in a chest in the quarters of
310   ◾    Procedural Generation in Game Design

a mercenary, in the stained glass window of an ancient religious building,

or in a piece of poetry.
All the interwoven meaning of URR is a method for the player to track
down these clues. A novel containing a clue doesn’t exist in isolation, but
might be mentioned by philosophers, influential in public opinion about a
recent war, sold in translation in foreign lands, or named after a particular
local animal. All these connected elements enable the player to zero in
on tracing down a copy of that work, and thereby acquiring the clue in
question.
Discovering meaning is therefore the game’s core gameplay system,
and one that hinges on both designer-led meaning (creating procedural
systems to create systems with shared meanings) and player-led meaning
(the ability to identify points of meaning, connect them, and extrapolate
from them).
When we think about “discovering meaning” as a game mechanic,
there is actually an illustrative comparison here to playing roguelikes
with and without spoilers. Many classic roguelikes are so complex, and
contain so many obscure and sometimes unclear mechanics, that many
players elect to play roguelikes “with spoilers”—this means that they have
a “wiki” or guide or equivalent source opened up in their Internet browser
while they play, and upon discovering a new monster, a challenging boss,
or some item they haven’t found before, or a decision they haven’t pre-
viously encountered, they will look up all the relevant information and
make their decisions based on that information.
Crucially, therefore, this is information the player is not intended to
have—if the game doesn’t give out that information upon a first encounter,
and this is indeed the player’s first encounter, then the game’s designers
must have intended for this decision to be made without complete knowl-
edge about its potential outcome(s). Some classic roguelikes, probably
most noticeably NetHack, are popularly thought to be almost impossible
to complete without the use of spoilers, and only the tiniest number of
people have ever (claimed to have) achieved such a feat.
Playing how the designers intended, therefore, in this case means
discovering the meaning behind the game’s systems. This is a form of
meaning contingent upon understanding more traditional statistical and
mechanical game systems—what does this stat mean, what does this item
do, and how does this monster attack?—rather than the kinds of mean-
ing I’m describing here in my own work, but it is still asking the player to
uncover meaning as part of playing the game.
 Meaning    ◾    311

All games of course have some kind of learning process, but the per-
madeath of roguelike games is what encourages many players away from
learning “the hard way” that Medusa will immediately petrify you if she is
given line of sight, or that certain enemies in Dungeon Crawl Stone Soup
can banish you to the abyss far before you are ready, and thereby away
from the intended gradual discovery metagame and toward immediate
knowledge.
There are two different kinds of knowledge at play here—NetHack
knowledge applies to every playthrough, since Medusa will always be a
threat, and URR knowledge, which only applies to a single playthrough,
aside from the meta-knowledge of how knowledge in URR is acquired. In
URR, players can only take to the next playthrough the understanding
that they are meant to acquire knowledge, not what that knowledge is,
and thus a wiki or guide becomes almost irrelevant and the centrality of
“discovering meaning as a game mechanic” is maintained even with the
presence of spoilers.

CONCLUSION
An opposition between meaning and randomness need not always be pres-
ent. Unchanging meaning can be easily added into procedural games
through the use of static markers around which the rest of the world is
generated, while “changing” meaning that is unique to every playthrough
can be implemented through utilizing deeply interconnected procedural
generation systems that do nothing in isolation, and always draw upon
each other when producing their outcomes.
In turn, this final point allows for the development of intriguing and
highly unusual game mechanics, focused on discovery, understanding,
and problem solving. They are similar to those used in many puzzle and
adventure games, but replayable across however many playthroughs are
desired. Even the creation of riddles or puzzles whose widely distributed
yet closely related elements, ideally, should be indistinguishable from
something that was handmade. Meaning can therefore be a central part
of procedural generation, but it requires designers to move beyond the
mental model of the “random” procedural generation that distributes
affordances with little thought to their connections, and instead embrace
the ability to distribute moments of meaning within the randomness, and
from these build a story, a background, or entire gameplay mechanics.
Index

A ambient behavior, 137

conversations, 138–141
Additive vs. parametric rendering, 170
dialects, 138–140
Advanced, and indirect rendering, 171
system, 140–141
Aesthetic properties, 233
emergent phenomena, 137–138
AI, see Artificial intelligence (AI)
movement and combat, 135–137
Akalabeth, 99
unpredictability and, 134–135
Algorithms, and approaches, 271–299
Art toys, 161–173
heightmaps, 279–283
aesthetics for, 173
box linear filters, 279–280
building, for experts and novices,
midpoint displacement, 280–281
161–163
Perlin and simplex noise, 281–283
design and construction of, 166–172
partitioning space, 293–297
connectivity and meshes, 168
binary space, 293–295
data and transformations, 168
Dijkstra maps, 296–297
forces and acceleration, 169
tree mapping, 297–298
gestural curves, 168–169
Voronoi diagrams, 295–296
inputs, 166–167
random numbers, 271–279
points and rotations, 168
making use of repeatable series,
rendering, 169–172
272–273
experience of, 163–164
normal distributions, 275–277
outside generator, 172–173
pseudo, generators, 271–272
trading control for power, 164–165
rolling dice, 274–275
Austen, Jane, 225
seeds and hashing, 273–274
Automated game tuning, 251–264
weighted distributions, 277–279
analyzing data, 258–260
sequence generation, 283–286
generating game level, 255
Lindenmayer systems (L-systems),
setting parameters, 253–255
283–284
simulating game, 255–258
Markov chains, 284–286
visualizing data and making
space filling, 286–292
adjustments, 261–263
cellular automata, 288–291
random walks, 286–288
settling, 291
B
Wang tiles, 291–292
Argument Champion, 46–50 Ban lists, 47–48
Artificial intelligence (AI), 133–141, 210, Bar, 181
250–252, 255–258, 261, 268 “The Battle for Mount Hyjal”, 147

313
314   ◾    Index

Beat, 178–179 combining output from

Beats per minute (BPM), 178 several layers of simulation/
Become a Great Artist in Just 10 Seconds, gameplay, 223
163, 172 juxtaposing events and
Binary space partition, 293–295 interpretation, 224–225
The Binding of Isaac, 302–303 source material, 216–217
Blood & Laurels, 217, 223, 227 Chords, 181
Bogost, Ian, 201 Civilization, 99, 100, 103, 105
Borderlands, 53 Code, 154
Bounding box, 75 Coen, Ethan, 204
Box linear filters, 279–280 Coen, Joel, 204
BPM, see Beats per minute (BPM) Cogmind, 137
Brogue, 83 Combination, 31
Brough, Michael, 163 Complicated Wires puzzle, 123
Bully Pulpit Games, 203 Compound queries, 127
Computer-generated images (CGIs), 210
ConceptNet, 46–50
C
Conditional lock, 91
Canvas, 170 Consumable keys, 93
Catlateral Damage, 57–58 Content tools case study, 245–250
Caves of Qud, 38–39, 199, 200, 201, 202, example room, 248–250
206, 289, 293, 296 system overview, 247–248
Cellular automata, 288–291 Continuously Escalating Total, 148–150
Centaur chess, 165 commentary, 149–150
CGIs, see Computer-generated key characteristics, 150
images (CGIs) pseudocode, 148
Champion, Erik, 303 Craft modules, 30
Characters, and personalities, 215–227 Critical Gaming: Interactive History and
creation, 220 Virtual Heritage, 303
easy to communicate, 222 Crusader Kings 2, 187, 188
meaningful in combination, Crypt generation, 66–68
222–223 Dungeonmans, pseudocode, 70–71
mechanically significant, goal, 66
221–222 room data, 66–68
orthogonal, 221 Cursed Chests, 143, 149
pitfalls, 225–227
overgeneralization, 225–226
D
overrealism, 226–227
untamed simulation, 227 Dangerous lock, 91
realization, 217–220 Darkest Dungeon, 202
layering dialogue features, Deep Blue, 165
218–220 Defuser, 119–120, 122
selecting dialogue, 217–218 Delaunay triangulation, 172
world interaction, 220 Density, 234
recurring strategies, 223–225 DeSantis, Dennis, 183
bringing character into every Desktop Dungeons, 107, 112–113, 114,
interaction, 223–224 115, 116
callbacks to earlier events, 225 core element, 112
 Index    ◾    315

design goal and resource model, G

112–113
Gabriel Knight, 109
Desolus, 158–159
Game design 101, 267–269
Deterministic generator, 109
Game of Life, 289
Diablo, 6, 52
Gamma World, 206
Diablo III, 143, 149
General game playing (GGP), 268
Digital art toys, 163
Generative artwork, 153–159
Dijkstra maps, 296–297
perception of intent, 156–159
Directed graph model, 191
techniques, 154–156
Distance-based clustering, 242–243
Generative spaces, 232
Doubledoom, 116
Generative tools, 165
Dungeon Crawl Stone Soup, 135, 245, 302,
Gestalts, 29–30
307, 311
assembly mechanisms and, spaces,
Dungeon generation pseudocode, 69–70
30–32
Dungeonmans, 63, 66, 68, 69
plotting desirable, 38–39
crypt generation pseudocode, 70–71
Gestalt spaces, 30–32
dungeon generation pseudocode,
GGP, see General game playing (GGP)
69–70
Google, 49, 168, 171
Dwarf Fortress, 9–10, 99, 103, 187, 188,
Google milking, 48
201, 213, 227
GPU, see Graphics processing
unit (GPU)
E Grammars, 188
Graph grammar, 192
Elder Scrolls, 303
Graphics processing unit (GPU), 158
Elite, 48, 99, 273
Graph transformation, 89
E major scale, 180
Grid-and-tile-based approach, 25–26
Emergent narratives, and story volumes,
Grid patterns, 19
199–207
Guinness Book of World Records, 53
Fiasco, 203–205
grist for narrative mill, 207
motivation, 199–201
H
unpacking themes, 205–207
voice in machine, 201–203 Hash function, 273
Emergent phenomena, 137–138 Heightmaps, 279–283
Evans, Richard, 216 box linear filters, 279–280
Excellence in Design Award, 107 midpoint displacement, 280–281
Experiential properties, 233 Perlin and simplex noise, 281–283
Experts, 119, 122 Henry, Desmond Paul, 153
Expressive range, 232 Herringbone Wang tiles, 20
Histograms, 239–242
Hitpoint Progression, 150–151
F
commentary, 151
Façade, 187 key characteristics, 151
Fiasco, 203–204 pseudocode, 150–151
Fixed-in-place keys, 93 “Holdout” missions, 143
Flappy Bird, 253–254, 255, 258–259, 261 A House of Many Doors, 210, 211,
Freehold Games, 33, 206 212–213
Frontier: Elite II, 10 Hunting and horse, 212
316   ◾    Index

I handcrafted integration, 63–71

crypt generation, 66–68
IBM, 165
Dungeonmans crypt generation
IGF, see Independent Games
pseudocode, 70–71
Festival (IGF)
Dungeonmans dungeon generation
Impossible games, 260
pseudocode, 69–70
Impressionism, 165
practices, 68–69
Improvise for Real, 183
standard dungeons, 63–66
Independent Games Festival (IGF), 107
Level generation, 83–95
Infinity mode, 5
cycles, 84
Initial rewrite rules (IRRs), 192, 193
implementation, 89–91
Input–model–rendering model, 171
lock-and-key attributes, 91–93
IRRs, see Initial rewrite rules (IRRs)
as binary barriers, 91–92
consumable/persistent keys, 93
J fixed-in-place keys, 93
particular/nonparticular keys, 93
Joubert, Rodain, 112 permanent locks, 92
Joy Exhibition, 166 safe lock, 92
single purpose/multipurpose keys,
K 92–93
valves and asymmetrical locks, 92
Karth, Isaac, 165 patterns, 87–89
Kasparov, Garry, 165 tilemaps, 93–95
Kazemi, Darius, 47 using graphs to express cycles, 84–87
Keep Talking and Nobody Explodes, Lindenmayer systems (L-systems), 162,
119, 121 283–284
Kunzelman, Cameron, 199 Linearity, 234, 235
Little galaxy, 298–299
Lock-and-key attributes, 89, 91–93
L
as binary barriers, 91–92
Launchpad, 232, 233, 235, 236–238, consumable/persistent keys, 93
240–242 fixed-in-place keys, 93
Leap Motion, 168 particular/nonparticular keys, 93
Left 4 Dead, 134, 187 permanent locks, 92
Leniency, 234, 236–237 safe lock, 92
Level design single purpose/multipurpose keys,
architecture generation, 73–82 92–93
assign region types, 80 valves and asymmetrical locks, 92
calculating bounding box, 75 Loop-erased random walk, 287
generating regions, 81–82 Lore, 302
make adjustments, 81 L-systems, see Lindenmayer systems
place connections, 78–80 (L-systems)
skim perimeter regions, 76–78
splitting box into regions, 75–76
M
case study, 57–62
overview, 58 McClure, Andi, 163
rules, 59–60 Making Music, 183
working, 60–61 Mark, 170
 Index    ◾    317

Markov chains, 284–286 O

The Mary Jane of Tomorrow, 217,
Oculus Medium, 171
218–219, 224
One-dimensional random walks, 286–287
Maya, 164
“On the nose”, 223
MD5, 274
OpenSimplex, 283
Meaning, 301–311
Orthogonality, 35–36
designer and player, 306–308
Overshadowing module, 37
in games, 302–304
meaninglessness in games, 304–306
in qualitative procedural generation,
P
308–311
Microsoft, 50 Panoramical, 172
Microsoft 3D Pipes, 170 Parker, Felan, 203
Midpoint displacement, 280–281 Particular keys, 93
Milk, Chris, 167 PCG, see Procedural generation (PCG)
Minecraft, 99, 100, 102 Pentatonic major scale, 181
Minesweeper, 110, 111 Perlin, and simplex noise, 281–283
Model-driven engineering, 85, 89 Permanent locks, 92
Modular design, 29–41 Permutation, 31, 111
enabling play, 33–38 Persistent keys, 93
equivalence of impact, 37–38 Persuasive Games: The Expressive Power of
mechanics as shared substrates, Videogames, 201
33–35 Photoshop, 164
orthogonality, 35–36 Picbreeder, 173
gestalts, 29–30 Pitch, 176–177
assembly mechanisms and, spaces, Pitfall, 14
30–32 Playable games, 260
plotting desirable, 38–39 Player vs. environment (PvE) games
inserting memorable asymmetry, 40–41 Poetry generation, 209–214
modules, 29–30 Postprocessing, 156
Modules, 29–30 Procedural audio, in Skipping Stones,
Morse code puzzle, 122–123 175–177
Motif, and repetition, 181–183 pitch, 176–177
Multipurpose keys, 92 implementation, 177
Murmer, 274 sampling, 175
Mushroom-11, 11 Procedural composition, in Skipping
Stones, 177–183
beat, 178–179
N
chords, 181
NetHack, 137–138, 308, 310, 311 motif and repetition, 181–183
No Man’s Sky, 8, 102, 305 scale, 179–181
Nonconditional lock, 91 steps, 179–181
Nonparticular keys, 93 Procedural enemy waves, 143–152
Nonplayer character (NPC), 139, 140, Continuously Escalating Total,
162, 191, 215–216, 217, 221–222, 148–150
226–227, 277 commentary, 149–150
Normal distributions, 275–277 key characteristics, 150
NPC, see Nonplayer character (NPC) pseudocode, 148
318   ◾    Index

Hitpoint Progression, 150–151 metrics vs. requirements, 237–238

commentary, 151 types of, 232–234
key characteristics, 151 visualizing expressive range, 239–243
pseudocode, 150–151 distance-based clustering, 242–243
Spawn by Timer, 144–146 histograms, 239–242
commentary, 144–146 Procedural logic, 119–132
key characteristics, 146 application, 121–122
pseudocode, 144 approach to, 120–121
Spawn on Completion, 146–148 background, 119–120
commentary, 147 improving process, 127–129
key characteristics, 147–148 queries, 127–128
pseudocode, 146 solutions, 128–129
Procedural generation (PCG), 135, 136, puzzles and their rules, 122–126
137, 138, 140–141 countable problems, 123–124
aesthetics in, 23–27 generating, 125–126
blurring boundaries, 26 trivial case, 122–123
establishing rules, 25–26 Wires puzzle, 124–125
result, 27 random rules, 129–131
Sunless Sea, 24 dealing with degenerates, 129–130
ethical, 43–54 validation, 130–131
Argument Champion, 46–50 rules list, 131
Borderlands, 53 Procedural worlds, 97–106
ConceptNet, 46–50 history, 98–101
Diablo, 52 qualitative generation, 103–106
future, 54 reasons to generate, 101–103
talking in code, 44–46 expansive/complex, 102
Tay, 50–51 exploration, 101–102
managing output, 13–21 gameplay variation, 103
planning, 3–5 ProcessAudioSources, 178
drafting content, 4 Pseudorandom number generators,
integral, 3–4 271–272
modal, 5 Puzzle, 107–117
segmented, 5 Desktop Dungeons, 112–113
risks, 5–8 core element, 112
authored experience, 7 resource model, 112–113
multiplayer, 7 generation approaches, 109–111
overreliance, 8 backward from goal state, 110
quality assurance, 6 heuristics, 110–111
random, 8 permutations, 111
time restrictions, 6–7 random start state, 109–110
use, 8–12 players, 113–117
unique, 10–12 generating hope, 115–117
utilitarian, 8–10 guaranteeing solvability, 114–115
Procedural generators, 231–250 procedurally generating, 107–109
expressive range and generative desired outputs, 108–109
spaces, 232 puzzle-spaces, 108
qualities, 232–239 solutions, 126
formalizing, into metrics, 235–237 Puzzle-spaces, 108
 Index    ◾    319

Q Sequence generation, 283–286

Lindenmayer systems (L-systems),
QA, see Quality assurance (QA)
283–284
Qualitative procedural generation,
Markov chains, 284–286
103–106, 308–311
SHA, 274
Quality assurance (QA), 6
Sims 3, 227
Queries, 125–126, 127–128
Single-purpose keys, 92
compound, 127
Skipping Stones, 175, 177–183
contexts, 127–128
procedural audio in, 175–177
pitch, 176–177
R sampling, 175
procedural composition in, 177–183
Random elements, 249 beat, 178–179
Random numbers, 271–279 chords, 181
making use of repeatable series, motif and repetition, 181–183
272–273 scale, 179–181
normal distributions, 275–277 Sky Rogue, 15–16, 17
pseudo, generators, 271–272 Smith, Alvy Ray, 169
rolling dice, 274–275 Snapjaws, 199
seeds and hashing, 273–274 Snowball effect, 145
weighted distributions, 277–279 Sokoban, 265
Random walks, 286–288 Souls, 303
Red Hook Studios, 202 Space filling, 286–292
Red Rock, 199 cellular automata, 288–291
Reed, David, 183 random walks, 286–288
Repeatable series, 272–273 settling, 291
Reversible locks, 92 Wang tiles, 291–292
Rewrite rules, 188 Space partition, 293–297
Risk–reward ratio, 234 binary, 293–295
River Raid, 99 Dijkstra maps, 296–297
A Rogue Dream, 48, 49 tree mapping, 297–298
Rolling dice, 274–275 Voronoi diagrams, 295–296
Rules generation, 265–269 Spawn by Timer, 144–146
experimenting, 269 commentary, 144–146
game design 101, 267–269 key characteristics, 146
mix and match, 266–267 pseudocode, 144
Rule validation, 130–131 Spawn on Completion, 146–148
commentary, 147
key characteristics, 147–148
S
pseudocode, 146
Safe lock, 92 SpeedTree, 158
Sampling, 175 Spelunky, 10, 18, 46, 53, 245
Scale, 179–181 Spirograph, 165
Scratch, Nick, 199 Spore Creature Creator, 172
Secondary rewrite rules (SRRs), 192, Sproggiwood, 33, 34–35, 36, 38–39, 40
194–195, 196 SRRs, see Secondary rewrite rules (SRRs)
Seeds, and hashing, 273–274 Standard dungeons, 63–66
Semantic properties, 233 area gen code, 64
320   ◾    Index

goal, 63 Uncertain lock, 91

room data, 63–64 Unexplored, 83–84, 86, 87, 89, 91,
starting environment, 64–66 93–94
Starcraft 2: Wings of Liberty, 143, 145 Unsafe lock, 92
Steam Early Access, 38 URR, see Ultima Ratio Regum (URR)
Stellaris, 50
Story, and plot generation, 187–207
game world, 190–191 V
simulation, 195–196 Valves, and asymmetrical locks, 92
grammars and story grammars, 188–190 “Varieties of Digital Painting”, 170
metric-guided generation, 196–1997 Veale, Tony, 48
model, 191–192 Venn diagram, 123
rule design, 192–194 Versu, 216, 217–218, 220, 222, 225,
secondary rewrite rules, 194–195 226, 227
Story volume, 202 Voronoi diagram, 172, 295–296
Strangethink, 166
Stream Friends, 199
Sudoku puzzle, 109, 111 W
Sunless Sea, 24
Walker, Austin, 199
Super Meat Boy, 6
Walker, Brian, 83
Wang tiles, 291–292
T Wayward Souls, 245, 246, 247
Weighted distributions, 277–279
Tales from Suburbia, 205 Weighted random selection, 76
Tales of Maj’Eyal, 302 Weighted selection approach, 129
Tay, 50–51 When in Rome 2, 216, 217
Temporary doors, 92 Windforge, 73, 76, 79–82
Text Rain, 172 Windows Solitaire win animation, 170
Tilemaps, 84, 93–95 Wires puzzle, 124–125
Tilt Brush, 168, 171 World of Warcraft, 143, 147
Topological properties, 233
Transformational graph grammars, 86–87
X
The Treachery of Sanctuary, 167
Tree mapping, 297–298 xxHash, 274
Trivial games, 260
TSR, 206
Y
Two-dimensional (2D) grids, 279
Two-dimensional random walks, 287–288 YouTube, 11

U Z
Ultima Ratio Regum (URR), 103, 138, 140, Zelda, 92, 93
141, 302, 308–311 Zubmariner, 27