A Whirlwind Tour of Computer

Simulation

Dave Goldsman
Georgia Tech
Atlanta, GA, USA
[email protected]
www.isye.gatech.edu/~sman
5/13/19

1
0. Getting to Know You
Goals:
* Contact Info
* Course Info

2
Contact Info
Instructor:
Prof. Dave Goldsman
School of ISyE, Georgia Tech
[email protected]
www.isye.gatech.edu/~sman
(404)822-8949
Office Hours: Anytime you can get me! 
3
Course Objectives
 Intro to simulation models and simulation
studies.
 Organization of simulation languages.
Modeling with Arena – or any comprehensive
simulation package with animation
capabilities.
 Statistical aspects including input analysis,
random variate generation, output analysis,
and variance reduction techniques.
4
Prerequisites
 You ought to know some probability
and statistics.
 You should be familiar with some
programming language and maybe
even a spreadsheet package.
 Good News: Don’t panic! I’ll make the
course as self-contained as possible!

5
Suggested Texts / Resources
 Banks, J., Carson, J.S., Nelson, B.L., and Nicol, D.M.,
Discrete-Event System Simulation, 5th ed., Prentice
Hall, Upper Saddle River, 2010.
 Law, A.M., Simulation Modeling and Analysis, 5th ed.,
McGraw-Hill Education, New York, 2015.
 Kelton, W.D., Sadowski, R.P., and Zupick, N.B.,
Simulation with Arena, 6th edition, McGraw-Hill, New
York, 2015.
 Free Arena software download:
https://www.arenasimulation.com/academic/students

6
Course Notes
 I’ll provide pretty extensive notes on a
course website.
 This doesn't mean that you can simply
print out the notes and skip the fun
class…
 …so we’ll also do some hands-on work.

7
Computer Programming
We’ll do a reasonable amount of
computer programming. You’ll have some
choice, but you can expect to use:
 A spreadsheet package, e.g., Excel.

 Some spreadsheet add-ons.

 A “real” language, e.g., Matlab or

Python.
 A simulation language, e.g., Arena.
8
Syllabus – Let’s Get Going!
 Introduction
 Calculus, Probability, and Statistics Boot
Camp (Law, Chapter 4)
 Hand Simulations; Spreadsheet
Simulations
 General Modeling Concepts (Law, 1&2)
 Verification and Validation (Law, 5)
 Is your simulation doing what you think?
9
Syllabus – Fun with Arena
 Arena Basics (KSZ, Chapter 4)
 A Generic Call Center in Arena (KSZ, 5)
 An Inventory Model (KSZ, 5)
 A Manufacturing Center (KSZ, 6)
 Entity Transfers in Arena (KSZ, 7)
 More-Advanced Arena Stuff (KSZ, 8)

10
Syllabus – Randomness
 Random Number Generation (Law, 7)
 generate “randomness” on a computer
 Random Variate Generation (Law, 8)
 single random variables
 multivariate random variables
 random processes
 financial models

11
Syllabus – Statistical Issues
 Input Analysis (Law, 6)
 What should drive the simulation?
 Output Analysis (Law, 9)
 Analyze what comes out of the simulation.
 Comparing Systems (Law, 10)
 Which system is better / best?
 Variance Reduction + Other Cool Stuff

12
Module 1 – A Whirlwind Tour!
1. Intro to Simulation
2. Some Easy Examples
3. Generating Randomness
4. Analyzing Randomness
5. Some Bigger Examples
6. Selecting a Simulation Language

13
1. Intro to Simulation
 Models are high-level representations of the
operation of a real-world process or system.
 Our concern will be with models that are
 Discrete (vs. continuous)
 Stochastic (vs. deterministic)
 Dynamic (vs. static)
 How can you “solve” a model?
 Analytic methods
 Numerical methods
 Simulation methods

14
Examples of Models
 Toss a stone off of a cliff. You can model its
position via the usual physics equations –
analytical models.
 Model the weather. Too tough for exact
analytical models, so you might use numerical
methods.
 Add a little randomness, and you may have to
resort to a simulation model (plenty of
examples coming up).
15
What is Simulation?
 Simulation is the imitation of a real-
world process or system over time.
 Simulation involves the generation of an
artificial history to draw inferences
concerning the operating characteristics
of the real system that is represented.

16
Simulation is…
 One of the top three industrial
engineering / operations research /
management science technologies
 Used by academics and practitioners on
a wide array of theoretical and applied
problems
 An indispensable problem-solving
methodology
17
 What is It Good for?
 Describe and analyze system behavior
(systems can be real or conceptual)
 Ask “what if” questions about the system
 Aid in system design and optimization
 Pretty much anything can be simulated
 Fixed systems with no “customer” arrivals (e.g.,
estimate stock option prices), or
 Service systems like Queueing Centers, Manufacturing
Processes, Supply Chains, Health Systems, etc.
18
Reasons to Simulate
 Will the system accomplish its goals?
 Current system won’t accomplish its goals.
Now what?
 Need incremental improvement
 Create a specification or plan of action
 Resolve disputes
 Solve a problem, like a bottleneck
 Sell an idea
19
Advantages of Simulation 
 Can study models too complicated for analytical or
numerical treatment
 Study detailed relations that might be lost in the
analytical or numerical treatment
 Use as a basis for experimental studies of systems
 Use to check results and give credibility to
conclusions obtained by other methods
 Reduce design blunders
 Really nice demo method
 (Sometimes) very easy
 …
20
Disadvantages 
 Sometimes not so easy
 Sometimes time consuming / costly
 Simulations give “random” output (and
lots of misinterpretation of results is
possible)
 To do a certain problem, better
methods than simulation may exist.
 …
21
History
 1777 – Buffon’s Needle Problem

 Early 1900’s – Beer and Student’s t

distribution

22
More History
 1946 – Ulam, Metropolis, von
Neumann, and the H-Bomb

 1960’s – Industrial Applications

 Manufacturing
 Queueing Models
23
Recent History
 Development of Simulation Languages
 Easy-to-use modeling tools
 Graphics  Harry Markowitz

 Rigorous Theoretical Work

 Computational algorithms
 Probabilistic and statistical methods

24
Origins: Mfg/Material Handling
 Simulation is the technique of choice
 Calculates movement of parts and
interaction of system components
 Evaluates flow of parts through the system
 Examines conflicting demand for resources
 Examines contemplated changes before
their introduction
 Eliminates major design blunders

25
Typical Questions
 What will be the throughput?
 How can we change it?
 Where are the bottlenecks?
 Which is the best design?
 What is the reliability of the system?
 What is the impact of breakdowns?

26
What I Used to Think Simulation
Was…

27
Actual Applications
 Manufacturing
 Automobile Production Facility
 Carpet Production Facility
 Queueing Problems
 Call Center Analysis
 Fast Food Drive-Thru
 Fast Food Drive-Thru Call Center
 Airport Security Line
28
Ex: Generic Queueing System
 Combination meat processing center /
fast-food restaurant / amusement park

 Denis (8) and Mina (6)

29
 Application: Supply Chains

supplier manufacturer distributor retailer customer

supplier manufacturer distributor retailer customer

• Think of supply chain as a network where any

node can potentially connect to any other node.
• In order to be lazy, let’s have things flow from
left to right. Looks a bit like a flowchart, eh? 30
 Supply Chains (cont’d)
 1985: Professor and CEO of SC company:
“Why would anyone want to simulate a SC?”
 Now: Many folks are developing SC tools
with simulation capability.
 Use simulation to determine how much value-added
a forecasting application provides.
 Use simulation to analyze how SC randomness or
model errors affect a proposed solution. Is it
robust? What’s the best solution?
31
Applications (cont’d)
 Inventory and Supply Chain Analysis
 Financial Analysis
 Portfolio Analysis
 Options Pricing
 Traffic Simulation
 Airspace Simulation
 Service Sector
 Health Systems 32
Health Systems Applications
 Patient Flow in a Hospital
 Hospital Room Allocation
 Optimization of Doctor / Nurse
Scheduling
 Procurement of Supplies
 Disease Surveillance
 Propagation of Disease Spread
 Humanitarian Logistics 33
Surveillance Application
 Use simulation to monitor certain time
series
 Predict issues as or before they happen
 Is a disease in the process of becoming
an outbreak?
 When is something out of the ordinary
occurring?
 Take advantage of HUGE data sets.
34
Who Is This Good-Looking Guy?

35
 Dr. Harold Shipman
 Killed his patients using heroin overdoses
(2nd ranked serial killer in UK)
 Was caught after carelessly revising a
patient’s will, leaving all her assets to himself
(don’t use your own typewriter next time)
 His computer records were doctored to show
his patients had needed the morphine.
However, the software recorded the dates of
these modifications.
 Hung himself in prison, never confessed 36
Dr. Harold Shipman

37
 How Does Simulation Help?
 Surveillance relies on a series of sequential
statistical hypothesis tests, where the null
hypothesis is “no disease” (or “no murder”)
 The statistics involved in the test have very difficult
distributions, even under the null hyp H0
 Simulation can be used to approximate the
probability distributions of the statistics under H0
 If sampling reveals deviations from this simulated
distribution, you reject H0
Thanks to Prof. Bill Woodall for this example! 38
π

2. Some Easy Examples

 Happy Birthday
 Let’s Make Some Pi
 Fun With Calculus
 Random Numbers Gone Bad
 Queues ‘R Us
 Stock Market Follies

39
Happy Birthday
 How many people do you need in a
room in order to have a 50% chance
that at least two will have the same
birthday?
 9
 23
 42
 183
40
41
 Let’s Make Some Pi
 Use Monte Carlo simulation to estimate π.
 Idea:
 Area of a unit square is 1.
 Area of an inscribed circle is π/4.
 Probability that a dart thrown at the square
will land in the circle is π/4.
 Throw lots of darts. Proportion that will land
in circle should approach π/4.
 Multiply proportion by 4 to estimate π. 42
43
Fun With Calculus
 Use simulation to integrate
f(x) = sin(πx) over [0,1].
 Idea:
 Sample n rectangles.
 Each is centered randomly on [0,1] and
has width 1/n and height f(x).
 Add up areas.
 Make n really, really big.
 Sum of areas approaches integral of f(x).
44
45
Random Numbers Gone Bad
 See what happens when you use a bad
random number generator.
 Idea:
 Simulate heights vs weights.
 Should be a 2-D bell curve (normal
distribution) with most observations in the
middle and some on the outside.
 Do observations “look” random?
46
47
48
Queues ‘R Us
 Single-server queue at McWendys.
 Customers show up, wait in line, get
served first-in-first-out.
 What happens as arrival rate
approaches service rate?
 Nothing much?
 Line gets pretty long?
 Hamburgers start to taste better?
49
Queues ‘R Us (cont’d)
 Can analyze queues via simulation.
 Can analyze via numerical or exact
methods.

 Fun fact: Notice anything interesting

about the word “queueing”? How
about “queueoid”?

50
51
52
Stock Market Follies
 Simulate a portfolio of various stocks
 Stock prices change randomly from year
to year, with various volatilities
 Can consider different mixes for
portfolio
 Simple spreadsheet application

53
54
Taking a Random Walk
 Take a normal step up or down every
time unit and plot where you are as
time progresses
 This “random walk” converges to Brownian
motion
 Einstein and Black+Scholes won Nobel
Prizes for this research.

55
 3. Generating Randomness
 Need random variables to run the simulation,
e.g., interarrival times, service times, etc.
 Generate Unif(0,1) pseudo-random numbers
(PRN’s)
 Use a deterministic algorithm
 Not really random, but appear to be
 Generate other random variables
 Start with Unif(0,1)’s
 Apply certain transformations to come up with just
about any other type of random variable
56
Unif(0,1) PRN’s
 Deterministic algorithm
 Example: Linear Congruential Generator
 Choose an integer “seed,” X(0)
 Set X(i) = (a X(i-1) + c) mod(m), where a,
c, and m are carefully chosen constants,
and mod is the modulus function.
 Set the ith PRN as U(i) = X(i)/m

57
Unif(0,1) PRN’s
 Pretend Example:
 Set X(i) = (5 X(i-1) + 3) mod(8), with X(0) = 0

 Then X(1) = 3 mod 8 = 3

 X(2) = 2, X(3) = 5, X(4) = 4, X(5) = 7, etc.

 So U(1) = X(1)/m = 3/8

 U(2) = 2/8, U(3) = 5/8, etc.

 Numbers don’t look particularly random.

58
 (4,7)
(7,6)
(2,5)
(5,4)
(0,3)
(3,2)
(-2,1) (6,1)
(1,0)

The random
numbers fall
mainly on the
planes

Plot of (X(i-1) , X(i))

Unif(0,1) PRN’s
 Real Example
 X(i) = 16807 X(i-1) mod(231 -1)
 U(i) = X(i) / m
 This generator is used in a number of
simulation languages
 Has nice properties, including long “cycle
times”
 Better generators are out there

60
Generating Other Random
Variates
 Start with U(i) ~ Unif(0,1)
 Apply some appropriate transformation
 Example: -(1/λ) ln(U(i)) ~ Exp(λ)
 Inverse transform method – can use this
for various important distributions
 Many other more-sophisticated methods
available, e.g., Box-Muller method for
normals
61
4. Analyzing Randomness
 Simulation output is nasty. Consider
consecutive customer waiting times.
 Not normally distributed – usually skewed
 Not identically distributed – patterns
change throughout the day
 Not independent – usually correlated
 Can’t analyze via “usual” statistics
methods!
62
 Analyzing Randomness
 Two general cases to consider
 Terminating Simulations
 Interested in short-term behavior
 Example: Avg customer waiting time in a bank
over the course of a day
 Example: Avg # of infected victims during a
pandemic
 Steady-State Simulations
 Interested in long-term behavior
 Example: Continuously running assembly line
63
Terminating Simulations
 Usually analyzed via Independent
Replications
 Make independent runs (replications) of
the simulation, each under identical
conditions
 Sample means from each replication are
assumed to be approximately i.i.d. normal
 Use classical statistics techniques on the
i.i.d. sample means (not on the original
observations) 64
Steady-State Simulations
 First deal with initialization (start-up) bias.
 Usually “warm up” simulation before collecting
data
 Failure to do so can ruin statistical analysis
 Many methods for dealing with steady-state
data
 Batch Means
 Overlapping Batch Means / Spectral Analysis
 Standardized Time Series
 Regeneration

65
Steady-State Simulations
 Method of Batch Means
 Make one long run (vs. many shorter rep’s)
 Warm up simulation before collecting data
 Chop remaining observations into contiguous
batches
 Sample means from each batch are assumed to be
approximately i.i.d. normal
 Use classical statistics techniques on the i.i.d.
batch means (not on the original observations)

66
Batch Means Example
 Get a confidence interval (CI) for the
mean of an autoregressive process
 This process is highly correlated
 CI’s via classical statistics (“CLT” method
on next page) result in severe
undercoverage
 Look at batch means and overlapping
batch means
 BM and OBM do better than CLT.
67
68
5. Some Bigger Examples
 Immunization Clinic
 Maternity Clinic
 Guinea Worm Disease
 Airline Routing Structure
 Pandemic Disease Propagation

69
Immunization Clinic
Partnership of Immunization Providers

Use simulation to
 Study operations of an immunization clinic
 Model generic clinic
 Read interarrival and service distributions
from spreadsheet
 Study alternative clinic configurations
70
71
72
6. Selecting a Simulation
Language
 More than 100 languages out there
 Commercial
 Freeware
 Maybe 5 to 10 major players
 Cost considerations
 Ease of learning
 Ease of use
 Classes, conferences
73