Augmented Computational Design

Pirouz Nourian*1 , Shervin Azadi2 , Roy Uijtendaal3 , and Nan Bai4

1
University of Twente, 7522 NH, Hallenweg 8, Enschede, NL
2
Eindhoven University of Technology, 5612 AZ, Het Kranenveld 8, Eindhoven, NL
3
Nieman Raadgevende Ingenieurs, 3542 AB, Atoomweg 400, Utrecht, NL
4
Delft University of Technology, 2628 BL, Julianalaan 134, Delft, NL
*
Corresponding Author: [email protected]
arXiv:2310.09243v1 [cs.AI] 13 Oct 2023

October 16, 2023

Abstract
This chapter presents methodological reflections on the necessity and
utility of artificial intelligence in generative design. Specifically, the chap-
ter discusses how generative design processes can be augmented by AI
to deliver in terms of a few outcomes of interest or performance indica-
tors while dealing with hundreds or thousands of small decisions. The
core of the performance-based generative design paradigm is about mak-
ing statistical or simulation-driven associations between these choices and
consequences for mapping and navigating such a complex decision space.
This chapter will discuss promising directions in Artificial Intelligence for
augmenting decision-making processes in architectural design for mapping
and navigating complex design spaces.
Keywords: Artificial Intelligence, Design Space Exploration, Generative
Design, Augmented Intelligence, Probabilistic Design 1

1 Introduction
The core of the performance-driven computational design is to trace the sensitiv-
ity of variations of some performance indicators to the differences between design
alternatives. Therefore any argument about the utility of AI for performance-
based design must necessarily discuss the representation of such differences, as
explicitly as possible. The existing data models and data representations in the
field of Architecture, Engineering, and Construction (AEC), such as CAD and
BIM are heavily focused on geometrically representing building elements and
facilitating the process of construction management. Unfortunately, the field of
AEC does not currently have a structured discourse based on an explicit repre-
sentation of decision variables and outcomes of interest. Specifically, the notion
of design representation and the idea of data modelling for representing “what
needs to be attained from buildings” is rather absent in the literature.

1. This is the author version of the book chapter “Augmented Computational Design: Me-
thodical Application of Artificial Intelligence in Generative Design.” In Artificial Intelligence
in Performance-Driven Design: Theories, Methods, and Tools Towards Sustainability, edited
by Narjes Abbasabadi and Mehdi Ashayeri. Wiley, 2023

1
 This treatise proposes to systematically view the differences between design
alternatives in terms of decision variables, be they spatial or non-spatial. Based
on such an explicit formulation of decision variables, we set forth a framework
for building and utilizing AI in [architectural] generative design processes for
associating decision variables and outcomes of interest as performance indicators
in a reciprocal relationship. This reciprocity is explained in terms of the duality
between two quintessential problems to be addressed in generative design, i.e.
the problem of evaluation of design alternatives (mapping), and the problem of
derivation of design alternatives (navigation).
Starting from an explicit representation of a design space as an ordered
pair of two vectors respectively denoting decision variables and performance in-
dicators, we put forth a mathematical framework for structuring data-driven
approaches to generative design in the field of AEC. This framework highlights
two major types of applications for AI in performance-driven design and their
fusion: those capable of augmenting design evaluation procedures and those
capable of augmenting design derivation procedures. Moreover, we introduce
the reciprocity of “flows” and “manifolds” as an intermediary notion for going
beyond the so-called form-function dichotomy. Discussing these notions neces-
sitates the introduction of a mathematical foundation for the framework rooted
in multi-variate calculus.
The main advantage of this explicit formulation is to enhance the explain-
ability of AI when utilized in generative design by introducing meaningful and
interpretable latent spaces based on the reciprocal relationship between mani-
folds and flows. The balance of predictive/deterministic power and interpretabil-
ity/explainability is discussed in the concrete context of an illustrative example.
A chain of key concepts will be introduced in this chapter, starting from the
notion of decision-making in design, the nature of design variables, the specifics
of spatial decision variables, the notion of design space, and the two dual actions
in the exploration of design spaces: mapping and navigating.
Whilst the introduced framework is quite general, a particular class of Prob-
abilistic Graphical Models (PGM), Bayesian Belief Networks (BBN), is intro-
duced to provide a concrete illustrative example of the utility of AI in AEC. For
a deeper insight into this particular approach to data-driven design, the read-
ers are referred to the two classical books about PGM: Pearl (1988) & Koller
and Friedman (2009). The illustrative example is a BBN trained for making a
data-driven replica of the building energy model used by the Dutch government
in order to obtain a rough meta-model to be used in mass-scale policy analysis,
e.g. for advising the government on the relative utility of energy transition sub-
sidies and planning measures. This example is chosen not because the BBNs are
the most advanced models or the most accurate models for approximating such
large functions. The choice is rather pragmatic in that this model has proven to
be promising from the stance of predictive power while retaining a basic level
of theoretical interpretability and intuitive appeal.
The chapter is structured as follows: we first present a historical context
to establish the necessity of such a data-driven generative design framework;
continue with conceptualizing and mathematically formulating the structure of
the framework, dubbed as Augmented Computational Design (ACD); present
an illustrative example demonstrating the utility of the framework; and con-
clude with a discussion on its outlook, open questions, and avenues for further
research.

2
2 Background
Here we revisit the utility of AI for data-driven generative design by highlighting
some key gaps of knowledge in the field of AEC and briefly mentioning overar-
ching frameworks in computational design and AI that can be used to address
these gaps.

2.1 Relevance of AI in AEC

The earliest attempts for enhancing accountability and predictive power in com-
putational design can be traced back to the notions of Scientific Architecture
(Friedman 1975) and the Sciences of the Artificial (Simon 2019). Both of these
seminal books explicitly discuss the necessity of forming some kind of a specific
spatial and configurative form of design knowledge, the core of which boils down
to being able to explicitly represent the main subject matter of spatial design
as “spatial configurations”. One of the first phenomenological and systemic de-
scriptions of design processes explicitly referring to the notion of performance
is the “Function, Behaviour, Structure” framework of Gero and Kannengiesser
(2004), in which the overused notions of form and function are elaborated in
terms of expected and required behaviour/functionality from a system (dubbed
as the function), its design as a form or configuration (dubbed as structure),
and its performance (dubbed as behaviour). The idea of design as a process
of generating the representation of a spatial structure is explicitly discussed in
this framework and the difference between desired behaviour and the actual be-
haviour of the structure is discussed as the performance drive for the process.
What can be observed in this phenomenological framework, predating most re-
cent advancements in computational design, is the fundamental belief about
the innate necessity of creativity in terms of the cognitive capability of design-
ers for proposing structures capable of working as desired, based on some kind
of tacit knowledge. Congruently, an anthropological description of design pro-
cesses refers to the old duality between form (structure) and function (purpose)
of designed artefacts, and the fact that [in the absence of explicit knowledge and
representation schemes], as Kroes (2010) has put it, designers are traditionally
trained to produce solutions (draw them) through a “logical leap” often without
even understanding or paying any attention to the design requirements or sup-
posed levels of quality attainment. Suppose we wanted to evaluate (compare)
two different alternative designs for a hospital, (Jia et al. 2023), or a home,
how do we want to represent the designs digitally for a computer to evaluate
them? Let us discuss an analogical example: if we wanted to compare two pieces
of music in terms of their beauty, it would be very straightforward to digitize
their notations and feed them to a machine, because the musical notation is
already discretized (digitized), regardless if it is written on paper or etched on
the cylinder of an old-fashioned winding music-box (Zeng et al. 2021). However,
doing the same, that is comparing two buildings, would be a much more difficult
challenge especially because there is currently no [discrete/textual] notation for
spatial design that can capture the features of spatial configurations.
Instead of the extensive emphasis on the product of architecture as the
shapes of buildings, we turn our attention to the processes of design and put
a lens of “design as [discrete] decision-making” on the debate to avoid com-
monplace reduction of design to the production of design drawings. This view

3
forms the basis of the generative design paradigm as extensively articulated by
Nourian, Azadi, and Oval (2023) & Veloso and Krishnamurti (2021). Similarly,
the challenges, opportunities, and promising ways of utilization of AI (particu-
larly deep-learning and generative models) for goal-oriented design explorations
have been discussed extensively in (Regenwetter, Nobari, and Ahmed 2022) &
(Regenwetter and Ahmed 2022).

2.2 Historical Context

In this section, we first give a very brief history of the most important and
relevant developments in AI; then lay the foundation of a formulation of ar-
chitectural design as a matter of decision-making; discuss the mathematical
implications of this paradigmatic frame for generative design; elaborate on the
notion of decision-making and the duality of derivation and evaluation problems;
and discuss two statistical approaches to design, a possibilistic approach utiliz-
ing Fuzzy Logic or Markovian Design Machines and a probabilistic approach
utilizing Bayesian Belief Networks or Diffusion Models.
We are currently witnessing an era of exponential success in the field of arti-
ficial intelligence that has been evolving for more than 50 years (See Figure 1).
Meanwhile, it is common knowledge that progress is slow in terms of innovation
and scientific knowledge development in the field of AEC.
As has been extensively argued before by Simon (1973) and Azadi and
Nourian (2021), once an unambiguous language is adopted for discussing the
classification of problems, we can see that many of the problems of the AEC
can be dealt with very adequately (and possibly painstakingly) through conven-
tional mathematics, physics, and computer science. In other words, the utility
or the necessity of employing AI for dealing with problems that can be dealt
with through conventional mathematical or computational procedures is not
only pointless from a resource-efficiency stance but also questionable from the
point of view of interpretability, transparency, and explainability. To assess the
potential applications of AI in AEC w.r.t. these questions, we highlight the
history of AI (see Figure 1) and refocus on the scope of AI (see Figure 2), at
least as could possibly pertain to AEC.
Once a problem is formulated adequately, two major determinants can be
considered as to whether it would be sensible to apply AI or not: whether
the data schemata of the problem are structured (vectorized) or unstructured
(textual/visual), and whether the underlying associations between the inputs
and outputs can be modelled through first principles (governing laws of physics,
typically stated in differential equations), stochastic processes, or Agent-Based
Models. If the problem data are unstructured or the conventional modelling
approaches do not have the capability of capturing the complex associations
between the inputs and the outputs of interest, then, especially when inter-
pretability can be sacrificed over the necessity of predictability, utilizing AI is
quite sensible. The example that we shall discuss in this chapter may seem
somewhat questionable according to these points but on the other hand, it is
too overwhelmingly large and complex that no conventional approach can deal
with it at the aimed level of abstraction. In this case, the ambition of the project
is on such a high level of abstraction for policy analysis that the inaccuracies
and ambiguities of the purely data-driven approach can be justified because of
the insights that can be gained from the meta-statistical model.

4
 Morphogenesis 1952
Alan Turing, Artificial Intelligence Paul Werbos
Mathematics, Back Propagation
Computer-Science 1974
+ Theoretical Biology Machine Learning
GPU
Any Computing
Technology that
Soft Computing
enables Statistical
mimicing Methods
human Enabling AlphaGo 2017
behaviour Learning Fuzzy, Neural,
Deep Learning
without Explicit Evolutionary, &
Programming Probabilistic Multi-Layered
Computing Artificial Neural
Networks

Jeff Hinton,
50

60

70

80

90

00

10

0
AlexNet, CNN,

2
19

19

19

19

19

20

20

20
Demis Hassabis,
AlphaGo, DRL...
Early AI, focused on
Building Expert
Systems, Symbolic Slow Take Off of Machine
AI, Prolog Learning after the AI Winter,
Aliases: Knowledge-Based
PyTorch
Systems, Computational
Intelligence

Accord.NET

TensorFlow

Scikit-Learn

NumPy

LISP ProLog MATLAB Python

Figure 1: Highlights in the history of Artificial Intelligence

2.3 Design as Decision-Making

The commonly overstated notions of difficulty or ill-defined nature of design
problems, see e.g. Simon (1973), can be attributed to the fact that most design
tasks are expected to produce a very concrete geometric description of an object
to be built (the form), given only a very abstract description of what the object
is supposed to be used for, how it should work, and what would be desirable for
it to achieve, all of which are often described quite vaguely (the function), q.v.
Kroes and Meijers (2006).
Hillier was one of the few shrewd theorists who understood that, at least
after the separation of structural design from architecture in the 19th century,
q.v. Giedion (2009), what distinguishes building buildings from architecture is
the art and science of configuring spaces, as stated in “Space is the Machine”,
(Hillier 2007). Once one realizes that the so-called architectural form is not
only a single shape of an iconic object but also a set that includes the shapes
of spaces and eventually the constituent segments of a building then we can
distinguish the superior importance of spatial configurations. As obvious as it
may sound, it seems to be necessary to emphasize that architectural design is not

5
 High-level Task Description Interpolation & Extrapolation Policy Formulation & Prescription
Low-level Task Description Clustering Classification Regression Dim. Reduction State Space Search

Technical Purpose Function Approximation with Imprecision & Uncertainty Optimization & Control

Knowledge Representation & Artificial Intelligence

Reasoning (KRR) Symbolic AI

Pattern Recognition & Machine Learning Metaheuristics

Knowledge Discovery Artificial Neural Networks & Nature-Inspired
(Data Mining) (Deep Learning) Intelligence
Supervised (Evolutionary & Swarm
Computer Vision Unsupervised Learning Optimization)
Learning
Network/Graph Data Mining
Manifold Learning Fuzzy Logic

Collective Intelligence
&
Natural Language Processing Probabilistic
Dimensionality Reduction
(NLP) Graphical Models

Automated Planning & Simultaneous Reinforcement Learning

Scheduling Localization & Mapping

Optimal Control (Robotics)

Robot Motion Planning Multi-Agent Systems

General Game Playing (GGP) Stochastic Games

Content Generation Generative Models

Statistical AI
we could know but
General Goals Telling us things that we kind of know, but faster and at scale! Telling us things that we do not know already!

Figure 2: An Euler diagram of the scope of Artificial Intelligence

merely about sculpting a shape but configuring spaces to accommodate some

human activities. This involves some puzzling tasks such as packing, zoning,
and routing spaces of various functions, which we hereinafter refer to as the task
of configuring buildings (Azadi and Nourian 2021). For the problems of shape
and configuration to be transformed into decision problems, they need to be
discretized rigorously. In short, we can call a massing problem a shape problem
and a zoning problem a configuration problem (see Figure 5).
The mainstay of the generative design paradigm is a rigorous reformulation
of a design problem as a discrete topological decision problem rather than a
geometrical problem (Nourian, Azadi, and Oval 2023). Therefore discretization
is the process of breaking down the integrated design problem into multiple
smaller yet interdependent decision problems. An example of such discretization
can be a voxel grid that provides a non-biased and homogeneous representation
of spatial units, each of which poses a decision problem of function allocation
(Nourian et al. 2016) &(Soman, Aditya, Azadi, Shervin, and Nourian, Pirouz
2022).
Moreover, to ensure the correspondence of these discrete decisions we need to
include the topological information about their neighbourhood to represent their
spatial inter-dependencies; similar to topology optimization (O’Shaughnessy,
Masoero, and Gosling 2021). At the limit, such discretization can also be used
to model a continuum of solutions and provide a frequency-based or spectral
representation system, similar to (Marin et al. 2021), for spatial design much
like the musical notation that is based on notes.

6
 Decision Space Performance Space Performance Space Decision Space

thousands of choices integral consequences integral consequences thousands of choices

Mapping Navigating
of from
Choices Consequences
to to
Consequences Choices

Mapping Design Spaces Navigating Design Spaces

How do the different consequences emerge from How should we combine the choices to get to the
combinations of choices? desired consequences?
Endocding Decoding
↫ ↬
Primal Evaluation Problem Dual Derivation Problem

Figure 3: The duality of evaluation and derivation problems in generative design

Additionally, it is important to note that design decisions have a strong

spatial dimension, however, they can include the social dimension to represent
the preference of stallholders and enable consensus-building (Bai et al. 2020).
Given a view of design as a matter of decision-making, we can readily see two
important types of practical questions that will shed light on the relevance of
AI for decision-making:
1. how to map/learn the associations of some hundreds or thousands of con-
stituent choices of a compound design decision (function approximation
and dimensionality reduction for ex-ante assessment of the impact of de-
cisions)?
2. how to navigate a gigantic decision space with thousands of choices
and their astronomically large combinations with a few important
consequences in the picture?
The proposed notion of design as decision-making makes a point of depar-
ture for the rest of the chapter in that it highlights two essential problems of
equal importance and significance that can be tackled by AI and their duality:
Firstly, evaluation problems can be portrayed as mapping problems in Machine
Learning and Deep Learning, where the approximation power of Artificial Neu-
ral Networks (ANN) can be exploited in regression and classification settings.
Secondly, derivation problems can be portrayed as navigation problems in gener-
ative models, concerned with navigating from a low-dimensional representation
of performance indicators towards disaggregated design decisions.
Encapsulating the complex and often non-linear associations of many de-
sign decisions with a few outcomes of interest or performance indicators is here
dubbed as a mapping problem. Inverting a map (as an approximated function,
e.g. in the form of an ANN), can thus be viewed as an enhanced or augmented
form of design, where the designer is navigated towards many small decisions
just by pointing towards certain data points within a low-dimensional perfor-
mance space (see Figure 3). It must be apparent that a navigation problem
in this sense is much harder to solve; almost always impossible in the absolute
sense due to an arbitrarily large increase in information content and thus a
combinatorial explosion of possibilities.

7
 A Spectrum of Generative Design Methodologies & Generative Models for Content Generation

Design Rules Knowledge Data

Exploratory Performance-Based (Associative)

Rule-Based Model-Driven Policy-Driven Data-Driven

Multi-Agent Systems Generative Models

Graph/Mesh Rewriting Grammars Topology Optimization
& &
Generative Design Methods & &
Markov Decision Processes Bayesian/Markovian
Symbolic AI/Expert Systems Shape Optimization
Reinforcement Learning Probabilistic Graphical Models

Formulation of Design Rules Mathematical Derivation Devising Design Games Systematic Data Collection
Computational Design
as from & &
Practice
Topology Rewriting Methods Governing Equations Scoring Mechanisms Black-Box Training

Rule-Based Models White Box Models Grey Box Models Black Box Models

Grammatical Analytical/Differential Gamified Statistical-Probabilistic

Graph Theory, Numerical Methods Stochastic Games Diffusion Processes

Mathematics Algebraic Topology, Partial Differential Eq. Markov Processes Markov Chains
Cellular Automata Multivariae Calculus Agent Based Models Manifold Learning

Mostly out of scope of AI Within the scope of AI

Figure 4: The spectrum of generative design methods and their relation to AI

methods

2.4 AI for Generative Design

Given the formulation of main generative design tasks as mapping and navi-
gating, we focus on a particular set of AI methods that are distinguished as to
their relevance for these tasks in high-dimensional design decision spaces. More
specifically, within the spectrum of generative design methods (Nourian, Azadi,
and Oval 2023), we focus on data-driven mapping and navigating strategies. As
shown in figure 4, for brevity, we will only focus on the data-driven approaches
to design on the right-hand side of the spectrum. Notwithstanding the other
possible applications of [different kinds of] AI in this generative design spec-
trum, such as Reinforcement Learning in Policy-Driven design (playing design
games), approximation of evaluation functions in topology or shape optimiza-
tion, and Expert Systems in grammatical design, our framework here is focused
on the statistical AI paradigm and so we only discuss the purely data-driven
approaches to generative design.
Two subtle issues must be noted here: firstly, instead of discussing the utility
of the wondrous application of generative models for the entertainment industry,
we shall reflect on how the generative processes based on diffusion or dimen-
sionality reduction can be controlled for attaining high-performance designs in
an explainable manner. Secondly, the model-driven approaches to performance-
based generative design (topology optimization in particular) based on first
principles, are already utilizing something important from the realm of nature-
inspired computing called Hebbian Learning, which is already in the scope of
[statistical] AI. This point, however important, generally interesting, and rela-
tively unknown, falls out of the scope of this chapter.

8
3 Framework
The emphasis on the decision-making approach to design entails that design
tasks can be formulated as a set of [typically unstructured] questions about the
form and materialization/construction of an object (a building) to be answered.
In this chapter, we focus on the questions that pertain to form.
We propose a mathematical framework for generative design that relates
multiple strands of work together. We use design space for referring to an or-
dered pair of two vector spaces: a decision space containing vectors or data
points representing design configurations in the form of x ∈ (0, 1]n and a perfor-
mance/quality space containing vectors or data-point representing combinations
of outcomes of interest in the form of o ∈ [0, 1]q .
The mathematical lens that we shall put on the issue is to redefine both of
these notions to provide a much more specific and workable idea for discussing
the utility or futility of applying AI to design problem-solving. It is hopefully
easy for the reader to accept that a regular discretization of a so-called design
space (which is an unfortunately common misnomer, but here somewhat prag-
matically useful) provides a straightforward and simple discretization of design
decisions as vectors in the form of x := [xi ]n×1 ∈ (0, 1]n or x ∈ {0, 1}n , where
n is the number of discrete cells in the design space in which virtually any
conceivable shape can be constructed at a certain level of resolution. Without
loss of generality, however, the decision variables are not necessarily spatial and
can be assumed to be relativized float variables in the range of minimum and
maximum admissible parameter values of the functions that together result in
the shape and configuration of a building. Even if a multi-colour (multi-label,
multi-functional) space is the subject of the design problem, then multiple cat-
egories/colours of such vectors can be seen together as a matrix of decision
variables, whose rows have to add up to 1 (see Figure 5).
Once this terminology is established, it is easy to observe that, in Machine
Learning (ML) terms, the problem of performance-based design can be seen
as two problems that are dual to one another, a multi-variate regression prob-
lem for figuring out an approximation function that can map a few outputs
to many inputs (here we call this mapping or the evaluation problem), and a
pseudo-inversion problem for finding the combination of inputs that could re-
sult in desired output data points (here we call this navigation or the derivation
problem), see Figure 3.
When approached as a data-driven problem-solving task, both problems are
somewhat hard and impossible to solve in the absolute sense of the word, unless
we think about them as loss minimization or approximation problems. The nav-
igation/design problem is much harder than the mapping/evaluation problem.
The main idea here is to advocate for training (fitting) meta-models (neural
networks) to sets of sampled pairs of inputs and outputs to first approximate a
complex design space as a map between decision data points and performance
data points, and then find the pseudo-inverse of this map or navigate it in the
reverse direction to be able to find designs (decision data points) that perform
in a desired way. In other words, mathematically, we look at the performance-
based design process as a pairing between a decision space and a performance
space, where a map is conceptualized as a function f : (0, 1]n 7→ [0, 1]q ) such
that o = f (x). The pseudo-inverse map is thus dubbed as f −1 : [0, 1]q 7→ (0, 1]n ,
such that x = f −1 (o).

9
 For the sake of brevity and also generalizability to non-spatial design prob-
lems, we shall focus on massing problems and leave coloured configuration prob-
lems out of the picture momentarily (see Figure 5 for the distinction).
Furthermore, by considering two abstract and high-level descriptions of a
design task in our proposed regular discretization frameworks, we can formulate
two mathematical tasks:
• Mapping Design Spaces: approximating the function that can model the
associations between the many input design variables and a few outcomes
of interest;
• Navigating Design Spaces: approximating the inverse function that can
guide the generation of valid configurations in the decision space given
desired data points in the performance space.

3.1 Design Space Exploration

Here we explain the mathematical meaning of the two dual problems that to-
gether can be called design space exploration tasks: mapping and navigation.

Mapping The problem of mapping associations between a large set of inde-

pendent input decision variables and dependent output performance indicators
is key to performance-driven design. Any explainable and accountable design
methodology should have the capacity to guarantee the attainment of some
quality or performance indicators. From a mathematical and statistical point
of view, we might prefer to have an explainable and interpretable model of such
relations that can be fitted into our data or ideally a simulation model to predict
outputs from input data. However, in some cases, especially where a multitude
of very different quality/performance indicators are involved, and when one does
not have an established basis for simulation modelling, statistical (data-driven)
modelling seems to be the only option. And so, when the complexity of the
model passes a certain threshold of non-linearity and a multitude of input out-
puts, we might prefer to trade interpretability for predictive power. That is
exactly where ANNs as families of adjustable non-linear functions stand out
as viable function approximators. Training a network is practically a matter
of minimizing a loss/error function by adjusting the parameters of a family of
functions (that is set out by the so-called architecture/structure of the ANN).
Even though this approximation is inherently non-linear and global, it is
illuminating to think of an alternative [locally] linear approximation based on
the Jacobian Matrix. Suppose that o = f (x) := [fk (x)]q×1 = [fk ([xi ]n×1 )]q×1
is a vector of multiple scalar functions of vector input variables. Then a basic
idea of approximation is to approximate this function locally around an input
data point by its Jacobian. This matrix operator gives the basis for a hyper-
plane equation that provides the n-dimensional Euclidean tangent space of the
underlying function, very much like a multi-variate regression hyperplane, albeit
the latter would be fitted to the entire dataset.
Note that the Machine Learning task here would be a multi-variate regression
task in this case, i.e. predicting the dependent given the independent variables.
To understand the difficulty of the mapping then consider that the Jacobian
matrix J := [Jk,i ]q×n = [ ∂f T
∂xi ]q×n = [∇ fq ]q×1 would just provide the best local
k

linear approximation of an otherwise globally non-linear map from Rn to Rq ,

i.e. n decision variables to q quality criteria or performance indicators.

10
 The Jacobian approximation is numerically computable provided the under-
lying function is smooth and differentiable. For brevity, as commonly done, we
have omitted the fact that the Jacobian can be evaluated at a certain input
data point and that it is expected to be the best linear approximation of the
function in question in the vicinity of that point. If we abbreviate the notation
for the Jacobian as such a functional, then we can denote the approximate lin-
ear function at any given data point as follows (as a first-order Taylor Series
expansion): o(x)|x∼xo ≃ J(x0 ) (x − x0 ), or simply put, as o ≃ Jx, if we assume
x to represent the vector of difference between the input data point with the
centre of the neighbourhood.
The Jacobian approximation is also illuminating for another important rea-
son: it allows us to approximate the Jacobian in a different sense, i.e. in the
sense of dimensionality/rank reduction using the Singular Value Decomposition
(SVD) to see a clearer picture of the main factors playing the most significant
roles in attaining the outcomes of interest, in other words identifying the inputs
variables to which the outcomes of interest are most sensitive. Even though
we do not explicitly perform this operation in our demonstrative example us-
ing the SVD, it is still illuminating to see what SVD can do for this insightful
approximation and dimensionality reduction for two reasons:
• The SVD approximation of the Jacobian allows us to make a cognitive and
interpretable map of the most important causes of the effects of interest
• The SVD approximation of the Jacobian allows us to conceptualize a
pseudo-inverse function to navigate the design space from the side of per-
formance data points.
The SVD (low-rank) approximation of the Jacobian Matrix can be denoted
as below:
J := UΣVT , (1)
where Uq×q := [uk ]1×q and Vn×n := [vi ]1×n are orthogonal matrices (i.e.
UUT = Iq×q and VVT = In×n ), and Σ is a matrix of size q × n with only
p = min{q, n} non-zero diagonal entries denoted as σc and called singular val-
ues, which are the square roots of the eigenvalues of both JT J and JJT , sorted
in descending order, q.v. (Martin and Porter 2012).
X
J≃ σc uc vcT , (2)
c∈[0,r)

where, r ≤ p. It must be noted that the sum is not meant to be exhaustive in

that the sum of the first significant terms would achieve the purpose of dimen-
sionality reduction of the decision space by showing a low-dimensional picture of
the correlations between decision variables and their performance consequences.
So, instead of decomposing the Jacobian up to p, we can choose to have a lower
dimensional approximation up to some arbitrary smaller number r.

Navigation Navigating a high-dimensional design space from the side of the

performance space towards the decision space for deriving design decisions (see
Figure 3) is clearly a very challenging task, almost always impossible in the
absolute sense of solving the equation Jx = o, if the decision variables x are the
unknowns.
It is easy to see that the Moore-Penrose pseudo-inverse of the approximated
Jacobian matrix can be computed as a matrix of size n × q by easily using the

11
SVD factorized matrix:
J† := VΣ† UT , (3)
where Σ† is simply formed as a diagonal matrix of size n×q with the reciprocals
of the singular values. Similarly, the approximate pseudo-inverse of the Jacobian
can be computed as: X
J† ≃ σc−1 vc uTc . (4)
c∈[0,r)

However, in the same way, a minimal loss approximate solution exists for such
equations when the matrix is rectilinear, J† o is expected to be the least-square
solution to the linearized Jacobian approximate of a navigation problem. Even
though the system might in theory have a solution, the odds of finding a unique
solution are practically very skewed towards having an indeterminate system
with many more inputs than outputs, and thus the system will have many
approximate solutions rather than a unique exact solution.
This is of course in line with the intuition of most human beings about
the inherent difficulty of design problems for which there is no unique solution.
Note that in all these theoretical treatments we implicitly assumed that all data
points within the decision space correspond to valid designs whereas in reality,
it might be more difficult to ensure finding valid solutions (feasible in the sense
of complying with constraints) rather than good solutions. In other words,
constraint solving tends to be more difficult than optimization in a feasible
region of decision space.

3.2 Spatial Design Variables

If the question of the design problem directly pertains to the shape of the con-
figuration of an object, we can still construct decision variables to be handled
within the proposed framework for mapping and navigating design spaces.
The idea of bringing spatial decision variables in a generative design process
is to consider first the nature of the objects being designed as manifolds, i.e.
locally similar spaces (homeomorphic) to Euclidean spaces of low dimensions
(2D planes or 3D hyper-planes) but globally more complex, possibly having
holes, handles, and cavities (shells). Three types of these manifolds are of special
interest for generative design, such as those that conduct walk flows (explicit
or implicit pedestrian corridors in buildings and cities), light flows (rays of
sunlight, sky-view, or other visibility targets), and force flows in structures. Our
conceptual framework proposes that these flows are conducted within spatial
manifolds as below:
1. Walkable Space Manifolds (2D) conduct walks (accessibility questions)
2. Air Space Manifolds (3D) conduct light rays (visibility questions)
3. Material Space Manifolds (3D) conduct forces (stability questions)
This consideration allows us to see the way this object is supposed to func-
tion would largely be determined by how this manifold is configured in that the
way the manifold in question conducts the flows of walks in a walkable floor
space, flows of light rays in a visible air space, or flows of forces in a reliable
material space. Thus we can highlight the specific concept of flow in a network
representation dual to the discrete representation of a manifold as an unambigu-
ous alternative intermediary instead of any vague notion of function to study
and measure.

12
 0 Morton Binary Index
Morton Code= (d)
1 (ZYXZYX) Problem Elements
[decimal]
2
assuming XX=00
3
4 voxels:
1010 1011 1110 1111 10
36 38
11 14
52 15
54

11

11
100100 100110 110100 110110

Binary Coordinates along the

1000 1001 1100 1101 32
8 34
9 12
48 13
50 colours:

10

10
100000 100010 110000 110010

0010 0011 0110 0111 2

4 3
6 20
6 22
7 universe:

01

01
000100 000110 010100 010110
Z-Axis

0000 0001 0100 0101 0 1

2 16
4 18
5 design domain:

00

00
000000 000010 010000 010010
0 1 2 3 4

00 01 10 11 00 01 10 11 design/configuration: voxels labelled with

Binary Coordinates along the colours
Voxel:
Morton Code:

design question: what should be the colour of each voxel in ?

38
11

34
9
Massing Problem (Binary [3D] Image) Zoning Problem (Coloured [3D] Image)
3
6
Y-Axis
1
2
s
xi
-A

size of binary decision-space: size of coloured decision-space:

X

0 1 2 3 4

Figure 5: An illustrative discrete design domain and its associated decision space
distinguished for shaping/massing and zoning/configuring problems

Apart from mathematical elegance, this approach also provides multiple

computational advantages that are very much in line with the recent advance-
ments in the field of generative models in AI. In a nutshell, the discrete represen-
tation of the so-called design space not only provides a workable representation
of all possible forms but also a workable representation of some inherent func-
tional properties of the represented manifold that should logically determine
how it could function as a building (or a structure). The manifold representa-
tion can be mathematically denoted as a polygon mesh of vertices, edges, and
faces M = (V, E, F ) (for a 2-manifold) or a polyhedral mesh of vertices, edges,
faces, and cells M = (V, E, F, C) that can have a dual graph representation in
the form of Γ = (N, Λ).
This description should principally sound natural if we articulate the purpose
of a design task as follows: finding the ideal form (configuration and shape)
of a manifold to conduct some flows in a desirable pattern. In this way, we
are diverting our attention from the containers of space (the building) into
what it contains (i.e. the space and its spatial configuration). This change of
focus allows us to see the direct correspondence between the so-called form and
function of a design or, better put, the form (i.e. configuration & shape) and
expected quality/performance of a spatial configuration.
In what follows, we shall go much beyond the vector data inputs consisting
of only numerical variables, especially in the context of our illustrative example.
In fact, without loss of generality, the ideas of mapping and navigating design
spaces in an approximate sense go beyond decision variables pertaining to con-
tinuous decision variables and those pertaining to the spatial configuration and
geometric shape of spatial manifolds. The same ideas can be applied to design
problems that are about decision-making in a much more general sense as dis-
cussed above. Note that the illustrative example that we have demonstrated
at the end of the chapter has a heterogeneous mix of spatial and mostly non-
spatial decision variables as well as a mix of categorical and numerical decision
variables.

13
3.3 Statistical Approaches to Design
Amongst the statistical approaches to design, we can distinguish the possibilistic
approaches from the probabilistic ones.
• probabilistic approaches: BBN, Variational Auto-Encoders (VAE), and
Diffusion Models
• possibilistic approaches: Markovian Design Machines, Fuzzy Design (see
MAGMA below)

Possibilistic Approach The essence of the possibilistic approach to design is

using a multi-valued or non-binary logic framework for making design decisions,
typically in the sense of making discrete choices about discrete segments of space,
for example, the Markovian Design Machines of (Batty 1974), the Spatial Agents
Academy of (Veloso and Krishnamurti 2020), and MAGMA (Multi-Attribute
Gradient-Driven Mass Aggregation) through Fuzzy Logic as introduced briefly
in Nourian (2016) and Soman, Aditya, Azadi, Shervin, and Nourian, Pirouz
(2022). Both of these methodologies apply non-binary logic from a possibilistic
point of view, in the sense that they take design inputs that are valued in the
range of [0, 1] but treat them as possibility measures rather than probability
measures. The two big ideas behind these two methods are the utilization of
Markov Chains, Markov Decision-Processes, and Fuzzy T-Norms for coping with
uncertainty and human-like reasoning in simulated negotiations between spatial
agents.

Probabilistic Approach The probabilistic models briefly mentioned here

are all related to the concept of conditional probability, the Bayes theorem, and
[generalized] stochastic processes that resemble Markov Chains (Weng 2021;
Nourian 2016). In a nutshell, the core of these models is about updating some
posterior probabilities indicating beliefs about the truth of some statements by
prior probabilities multiplied by the likelihood of compelling evidence, scaled
by the probability of the existence of the evidence. When probabilistic neuron-
like nodes in Probabilistic Graphical Models are combined, these new posterior
probabilities or probability distributions can be fed into other layers of a net-
work to create ANN architectures. A basic idea here is to gradually reduce
the dimensionality of input data into an abstract low-dimensional representa-
tion (encoding, or mapping, albeit into a typically unintuitive and interpretable
latent space) and then gradually use the inverse of the forward diffusion-like
processes to denoise a vector in the low dimensional hidden space. The lat-
ter process is called denoising or decoding and it matches our description of
navigation processes, albeit without the direct control of the meaning of the
latent space vectors. A breakthrough in this domain can come from enhancing
the explainability of the latent space low-dimensional representations. This idea,
however interesting, falls way outside the scope of this short treatise. Therefore,
here we only provide a theoretical minimum for understanding the demonstra-
tive example (i.e. a shallow Bayesian Belief Network).

14
4 Demonstration
In this section, we will present a demonstration of the utility of the proposed
framework to indicate how a discrete decision-making approach can facilitate
generative design processes. As a disclaimer, it must be noted that this example
is not chosen for technical reasons related to AI but rather due to its real and
societal and environmental important purpose for policy analysis concerning
energy transition planning actions at the level of a country, and sustainability
strategies at the building level.

4.1 Case Study

Understanding the energy performance of architectural designs is crucial in
ensuring a sustainable future. Building Energy Modelling (BEM) is a multi-
purpose approach used by designers and policymakers for checking building code
compliance, certifying energy performance, subsidy policy making, and building
management. The Dutch government has recently introduced the NTA 8800 cal-
culation model for quantitatively determining the energy performance and code
compliance of buildings (NTA 8800 2022). The NTA 8800 aims to provide
a transparent, verifiable, and enforceable building energy performance model,
based on the European Energy Performance of Buildings Directive (EPBD),
the European Committee of Standardization (CEN), and the Dutch Normal-
ization Institute (NEN) published standards (Nen 7120+C2:2012/A1:2017 NL
2017). These regulations describe methods to calculate the energy performance
of buildings, set energy requirements for new buildings, and make agreements
about energy label obligations in existing buildings. The NTA 8800 only con-
cerns building-related measures, as expressed in the EPBD, Annex A (Union
2021).
The NTA 8800 document has been implemented as an MS Excel tool by the
Dutch government (commissioned by Nieman B.V. consultants). This calcula-
tion model translates the public European standard document into a calculation
tool. The calculation tool is not publicly available and it is not documented.
Since we were given temporary and bounded access to this model we chose to
approximate it and construct a meta-model out of it. The model consists of 269
unique input parameters about the spatial and technical building design config-
urations, based on which the model returns three scalar response values about
the energy performance of the building design; BENG 1 (maximum permissible
energy demand in kW h/m2 y), BENG 2 (maximum permissible primary en-
ergy consumption in kW h/m2 y), and BENG 3 (minimum permissible share of
renewable energy use as a percentage). The acronym BENG refers to some na-
tional performance indicators for Nearly Zero-Energy Buildings (Bijna Energie
Neutrale Gebouwen in Dutch).
NTA 8800 model has three main limitations: (1) it can only process and
compute information about a single specific scenario at a time; (2) it returns
scalar values about the energy performance that is untraceable to input parame-
ters; and (3) missing input values could result in errors or non-realistic response
values. These three limitations make the model impractical for designers and
policy analysts, particularly in the early stages of design. This impracticality is
because, in conceptual design and policy analysis, designers need to (1) explore
and iterate various options simultaneously; (2) need feedback on the degree of

15
influence of each design decision; and (3) cannot provide detailed information
yet about later stage design choices, such as the technical systems.
The framework of Augmented Computational Design (ACD) is particularly
useful here as it allows us to relate the aggregated performance changes of the
few NTA 8800 outputs of interest to the changes in the many design deci-
sion parameters of its input. In this particular case, we adopt a probabilistic
meta-modelling (function approximation) approach based on the methodology
suggested by (Conti and Kaijima 2021).

4.2 Methodology
Meta-Modelling Meta models are models that describe the structure, be-
haviours, or other characteristics of related models, providing a higher-level
abstraction for constructing and interpreting complex numerical models that
approximate more sophisticated models often based on simulations. A meta-
model is to serve as a simplified, computationally efficient model of the model
(Conti and Kaijima 2021), also referred to as a surrogate model (J. P. C. Kleijnen
1975). The process of creating a meta-model is referred to as meta-modelling
(Gigch 1991). Some alternative meta-modelling techniques include interpola-
tion methods such as spline models (Barton 1998), polynomial regression (J. P.
Kleijnen 2008), or Krigging (Ankenman, Nelson, and Staum 2010).
Within the ACD framework, such meta-models provide structured ways to
perform the two most important tasks of the generative design: mapping and
navigation.
In general, a standard meta-model can be described as: o = f (x) ≃ g(x),
where o is the aggregated simulation response, f denotes a computational
simulation-based model conceptualized as a vector function and g is the
approximated model function (cf. Figure 3.) With this notation, the objective
of meta-modelling is to build the g in such a way that it produces reasonably
close values of o. In the case of augmented computational design, meta-
modelling can be adopted as a methodology of design mapping that provides
a differentiable and ideally reversible g that can be used in the navigating
process. In other words, the meta-modelling should structurally relate the
choices and consequences in such a way that the choices can be derived from
the desired consequences; hence providing a data-driven basis for generative
design. The next part demonstrates a probabilistic meta-modelling approach
to navigation tasks in high-dimensional design decision spaces, based on the
methodology introduced by Conti and Kaijima (2021).

Bayesian Belief Networks A Bayesian Belief Network (BBN) is a kind of

Probabilistic Graphical Model (PGM) that is effectively an ANN in the form
of a Directed Acyclic Graph (DAG) with neuron-like nodes that can compute
Joint Probability Distributions (JPDs) from input probability distributions or
discrete Probability Density Functions (PDF), which is then attributed to an
output probability distribution through a Conditional Probability Distribution
(CPD) computing posterior probabilities/beliefs through the Bayes theorem,
hence the name Bayesian. The set of edges of a BBN forms the model archi-
tecture or structure that represents the particular probabilistic dependencies
between the discrete probability distributions attributed to the starts and ends

16
of the nodes. This structure is typically set by the modeller based on their knowl-
edge of the process; while the conditional probability distributions (transition
probability matrices) are learnt from the experimental data. BBNs can help
us semi-automatically reason about uncertain knowledge or data (Peng, Zhang,
and Pan 2010). This makes it possible to perform probabilistic inference, such
as computing the JPDs of some outputs (effects) given some inputs (causes).
The name of these ANN comes from the idea of updating beliefs or hypotheses
posterior to observing evidence; more precisely, utilizing the Bayes theorem for
updating conditional probabilities in network structures, in a fashion similar to
modelling and evaluating Markov Chains, albeit with the difference that Markov
Chains operate as uni-partite networks but each neuron of a BBN is a bipartite
network coupled with an outer product calculator.
The neurons of a BBN consist of two layers, the first of which can be
dubbed a presynaptic layer that combines input discrete probability distribu-
tions (through an outer product) and forms a JPD and then flattens the JPD
to form a vector-shaped probability distribution. The second, i.e. the synaptic
layer then is a CPD, i.e. practically a rectangular probability transition matrix
that maps this flattened JPD to the output probability distribution. A BBN
then consists of such neurons connected in a DAG. Training a BBN means find-
ing the entries of the CPD in such a way as to minimize the loss of recovery
of the output probability distribution from the input distributions. The appeal
of BBNs is twofold: on the one hand, they allow the inclusion of expert knowl-
edge and intuition in the architecture of the network and on the other hand the
training of the network makes the network adapted to the objective data. In
this case, we limit the architecture of the network to a single layer of neurons
to keep the network invertible.

Network Architecture:
probability of sending emails to customer
A Directed Acyclic Graph (DAG)
probability of having meetings with customer
probability of sales happening
probability of revenue brackets

presynaptic layer:
form a Joint Probability Distribution (JPD)
as
a flattened outer product of inputs
synaptic layer:
form a Conditional Probability Distribution (CPD) Nodes with a column sum equal to 0 are input nodes
as
Nodes with a row sum equal to 0 are output nodes
a rectangular transition probability matrix

Figure 6: An illustrative example of a Bayesian Belief Network, eliciting the nature

of nodes and the network architecture, an example adapted from here

Workflow Research from Conti and Kaijima (2018) illustrates the four pro-
cess steps involved in developing a BBN meta-model. In this use case, we alter
this methodology as creating a BBN with all 269 input parameters is infeasi-
ble. We add an intermediary step of sensitivity analysis to identify the most
influential input parameters before constructing the BBN. Thus we follow these
steps in order (see Figure 7): (1) sample the input parameter space, (2) run
simulations to generate the output values, (3) sensitivity analysis and selection

17
of influential input parameters, (4) train the BBN, and finally (5) evaluate the
model’s robustness. As Conti and Kaijima (2021) highlights, it is important to
model a shallow BBN as a complete bipartite graph connecting all input nodes
to all output nodes, effectively limiting the topology to two layers. This would
allow us to make a reversible approximation that can be used to derive the nec-
essary input configuration for any desired performance output. Additionally,
the fixed values can also include some of the input variables turning them into
design constraints.

Figure 7: Overview of the workflow adopted from the framework introduced by

(Conti and Kaijima 2018)

Step 1: Sampling the Parameter Space We need to set up a Design of Ex-

periment (DoE), to generate simulation data to study the relationships between
various input variables and output variables (Hicks 1964). The experiment in-
volves running several simulations at randomised input configurations (Sacks
et al. 1989). Before running the simulation, it is important to carefully select
a sampling method, to determine these input configurations, since the chosen
strategy influences the quality of the meta-model (Fang, Li, and Sudjianto 2005).
Since it is assumed that the decision space is unknown, the intention is to be
inclusive of all regions of the decision space as possible. The sampling algorithm
should generate a well-varied response data set that captures all the information
about the relationships between the input parameters and the responses. For
this study, 20.000 quasi-random input samples were generated based on Sobol’s
sequences (Sobol’ 1990) to ensure the homogeneity of the samples.

18
Step 2: Run NTA 8800 Simulation Model Vectorization is an important
part of the ACD, we represent the decision variables and outputs of interests
as vectors (See Section 3.) Each sample point can be interpreted as a vector
of scalar input values x. Each batch of such vectors is fed to the NTA 8800
model to generate the vector of corresponding building performance outputs
o. After running the primary simulation model for the sampled input data
points, the response data is collated and linked to the input samples to form an
input-output dataset for regression modelling (as in Machine Learning).

Step 3: Sensitivity Analysis The creation of a meta-model from 269 pa-

rameters, each with scalar input values, requires a simulation of all possible
combinations (the number of options to the power of 269). Even limiting the
number of options for each parameter to two, results in an immense number
of possible combinations, calculated at 5.39 × 1080 . To contextualize the mag-
nitude of this number, it is more than the estimated number of atoms in the
observable universe.
The sheer magnitude of this number makes storage and training of Bayesian
Belief Networks (BBNs) infeasible. Therefore, in this study, we use global sen-
sitivity analysis to apportion the uncertainty in outputs to the uncertainty in
each input factor over their entire range. This allows us to remove the param-
eters with the lowest influence on energy performance. The sensitivity analysis
method is implemented in the workflow based on the SALib library (Herman
and Usher 2017).
This results in a meta-model with fifteen parameters instead of 269, making
it feasible to store and train the BBN, but on the other hand, reducing the
accuracy and scope of the model. However, the most influential fifteen parame-
ters are responsible for 90,45% to 92,30% of the final energy performance score.
Hence, we decided on the inclusion of the fifteen specific parameters to construct
the BBN meta-model.

Step 4: Build a BBN meta-model Building a BBN meta-model is a pro-

cess of associating the probabilistic relationships of inputs and outputs. These
relationships may be characterized by a high degree of non-linearity and possibly
multiple interactions and correlations between model parameters. Consequently,
there are two main steps in this process: (1) learning the network topology as a
DAG structure and (2) estimating the CPD attributed to the neuron-like nodes
of the network.
In this demonstration, we adhere to a particular network topology to ensure
the reversibility of the trained model (Conti and Kaijima 2021). Accordingly,
this BBN has only two layers: one corresponding to the input and one cor-
responding to the output. However, effectively only a single layer of neurons
operates in the middle of the two layers. In this case, the selected parameters
from the sensitivity analysis results are represented by the input nodes, and the
BENG 1, BENG 2, and BENG 3 parameters are the output nodes (see Fig-
ure 8). Therefore, we skip the topology learning step in the conventional BBN
modelling; because the topology of this particular network is assumed to be a
complete bipartite DAG. In particular, we use the pgmpy Python package to
model the network topology (Ankan and Panda 2015).
The next step is to estimate the nodes’ CPDs from the input-output dataset.

19
The CPDs for the nodes can be learnt directly from the input-output simula-
tion data generated in steps 1 and 2, using the Maximum Likelihood algorithm.
Additionally, we discretize each variable range into a fixed number of intervals.
All numerical input distributions generated using a space-filling approach, like
Sobol’s sequence or Latin Hypercube, are sampled based on continuous ranges,
and should therefore be discretized. Discretization is done by dividing the in-
terval of the parameter over a fixed number of ranges between the minimum
and maximum values.

Figure 8: The single layer BBN; Right: BENG 1, BENG 2, and BENG 3
parameters are the output nodes; Left: Most sensitive input parameters as input
nodes.

Step 5: Validating the Meta-Model To assess how our trained BBN ap-
proximates the original NTA 8800 model, we use a cross-validation approach
in combination with Normalized Root Mean Square Error (NRMSE) and Mean
Absolute Percentage Error (MAPE) (James et al. 2013). Cross-validation splits
the generated input-output dataset (step 2) into a training set and testing set,
before building the BBN (step 3). The BBN is trained on the training set and
assessed based on the testing set. However, to obtain a more reliable estimate of
the model’s performance, the dataset is split into several subsets or folds, with
each fold used as both a training set and a testing set. This research adopts a
k-fold cross-validation technique, where k refers to the number of groups that
the data set is split into. We set k = 10 based on experimentation to ensure a
low bias and a modest variance. The model is then trained on k − 1 of the folds,
and the remaining fold is used for testing. This process is repeated k times, with
each fold used for testing once. The performance of the model is then evaluated
by averaging the performance across all k runs.
Following the approach suggested by Conti and Kaijima (2021), we computed
the mean difference of the predicted and actual output values and normalize the
RMSE values through division by standard deviation to achieve NRMSE.
To calculate the accuracy of the meta-model, it is recommended to use mul-
tiple metrics to get a comprehensive evaluation of the model’s accuracy. Hence,
NRMSE is combined with the MAPE metric. MAPE measures the average
absolute percentage difference between the predicted and actual values. It is
a measure of the magnitude of the errors in the model’s predictions. Lower
NRMSE and MAPE values indicate better model performance. The larger the

20
error between the two, the higher the NRMSE and MAPE values will become.
Therefore, the NRMSE and MAPE results will indicate how dispersed the pre-
diction data is compared to the actual model response.

4.3 Results
This section presents the numerical results obtained from the experiment of
NTA 8800 meta-model.

4.3.1 BBN Validation Results

Here we elaborate on the results of the cross-validation technique in combination
with NMSRE and MAPE based on the test data set (s=1100). The interpreta-
tion of what is considered an acceptable NRMSE and MAPE score depends on
the specific problem and the context in which the meta-model is being used. In
general, it is recommended to compare the NRMSE and MAPE scores of the
meta-model with the baseline models and the state-of-the-art models in the field.
This can provide a benchmark for what is considered acceptable performance
in the specific context of the problem.
In our case, the BBN does not compete with other models but rather com-
petes with consulting building energy specialists in estimating the building en-
ergy performance in the early design stages. However, to assess the proficiency of
our model in capturing the underlying relationships using solely the 15 selected
parameters, we employ the following benchmarks: The NRMSE values should
be in the range of (0.20%, 0.60%) as baseline, and in the range of (0.10%, 0.30%)
as state-of-the-art (Bui et al. 2021); The MAPE values should be in the range of
(0.10, 0.30) as baseline, and in the range of (0.05, 0.15) as state-of-the-art (Khan
et al. 2021).
The NRMSE of the trained BBN for BENG 1, BENG 2, and BENG 3
respectively is 0.82%, 1.52%, and 0.47%. This indicates that except for the
BENG 3 indicator, the prediction of the model is not accurate enough. The
MAPE of the trained BBN for BENG 1, BENG 2, and BENG 3 respectively
is 0.35, 0.28, and 0.33. This indicates that the predictions of the model are on
the upper threshold of being acceptable as baseline models. Absolute prediction
difference can be seen in Figure 9.
100y ′
The prediction accuracy is calculated by P A = 100∗y−y ′ Where P A is the

prediction accuracy, y ′ is the prediction of the meta-model and y is the NTA

8800 estimation.

4.3.2 Toy Problem

Here we present a test case that demonstrates the effectiveness of the BBN meta-
model in building design. The study involves two toy problems that showcase
the advantages and utilization of the meta-model. The toy problems regard two
common design challenges that cannot be solved using the currently available
tools, such as the NTA 8800. The first problem involves predicting the BENG
1 energy performance of a typical Dutch dwelling during the early design stage.
The spatial characteristics of the building are fed to the meta-model. As output,
the meta-model returns a range and the confidence level of that range.

21
 Figure 9: Histogram of BBN prediction and NTA 8800 outputs (top) and their
comparison (bottom) for BENG1, BENG2, and BENG3, in order from lft to right.

In this toy problem, the meta-model predicts the BENG 1 value to be within
the range of (0 − 50)kW h/m2 .y, with a confidence level of 100% (see Figure
10). To validate this result, we cross-checked the predicted result with the final
configuration of the dwelling using the original NTA 8800 model. The NTA 8800
model returns a value of 39.8kW h/m2 .y, confirming the prediction capability
of the meta-model.
The second problem reverses the first problem and involves the ex-ante de-
termination of the most probable design configuration that satisfies a specific
energy performance goal. In this example, the BENG 3 value of a typical Dutch
dwelling design (35%) does not satisfy the minimal requirements (50%). Since
this problem arises in the final design stage, some input parameters cannot be
changed anymore. In this case, the architects and engineers are limited to mod-
ifying only the area (AreaPV) and power (PPV) of the PV panels. Since the
minimum required performance goal for BENG 3 is 50%, we set the goal value
to a range of 60 − 80%. Given the binning approach employed, it should be
noted that the AreaPV value of 5 depicted in the figure corresponds to a range
of [40, 50]m2 , while the PPV value of 5 corresponds to a PV Power range of
[200, 250]W/m2 . Accordingly, the meta-model advises increasing the PV area
to [40, 50]m2 , and the PV power to [200, 250]W/m2 (See Figure 10).
This discretisation allows for a clearer representation of the recommended
parameter values within the specified ranges, facilitating the interpretation and
practical implementation of the BBN meta-model outputs. Since these ranges
are the maximum of both scales: the meta-models advise can be interpreted as
maximizing the PV area and PV power to reach the goal BENG 3 value of 60%
to 80%.

4.4 Discussion
In the end, to validate this result, we finish the loop by calculating the final
configuration of the dwelling with the original NTA 8800 model. The NTA
8800 returns a value of 71%, confirming the reverse inference capability of the
meta-model. These results illustrate in a simple and digestible example how
the BBN meta-model is capable of providing valuable insights and assisting
architects and engineers in navigating the multidimensional decision space.
By using the numerical Design of Experiments and Sensitivity Analysis we
are effectively conducting a dimensionality reduction task similar to the low rank
SVD as introduced before. As mentioned earlier, the mention of this particular

22
Figure 10: Output Recommendations of the BBN Meta-model for Achieving BENG
3 Compliance.

approach of ACD was to illustrate the utility of the framework with a concrete
example in a societally relevant context where a Machine Learning approach to
modelling can help make an otherwise very complicated simulation procedure
to be approximately scaled up massively for policy analysis. Here we discuss
the potentials and the shortcomings of the model and note the issues with this
large-scale black-box approximation that require further investigation.
The existence of categorical variables in the inputs of the BBN limits the
general applicability of ACD as it affects the smoothness and differentiability
of the underlying function that is being approximated. However, for pragmatic
reasons, we have ignored this issue to demonstrate the idea in a large-scale case.

Validation results Compared to the NTA 8800, the Bayesian meta-model

is capable of capturing the most important relationships between inputs and
outputs. However, the difference between the meta-model’s predictions and
the NTA 8800 predictions can be rather high. This means that there is a
large difference between the output of the meta-model and the NTA 8800. The
NRMSEs of BENG 1 (0.82%), and BENG 2 (1.52%) show that the BBN is able
to follow the NTA 8800 to some extent, but is far from accurate as NRMSE
is greater than 0.5%. On the other hand, the model could predict BENG 3
(0.47) relatively accurately. This insufficient accuracy was expected as we have
dictated a particular topology to the BBN while learning the network structure
is an important step in constructing BBNs. This decision was made to enable
the model to function in a bidirectional way: inference and reverse inference
(i.e. evaluation and derivation in the terminology of our ACD framework).

Foreward Inference The trained BBN is now capable of inferring the outputs
of interest given certain input configurations. This inference uses the learned
CPDs to predict the most likely values for the outputs. In this way, we can
predict the energy performance of buildings, in a quick and intuitive way for ex-
ante assesment, based on a certain design configuration. In particular, inference
demonstrates the potentials of a mapping described in Section 3.

23
Backward Inference Since our BBN had only two layers in its network,
it can be reversed. This means that instead of presenting evidence to it, we
can present the desired performance values and ask for the derivation of the
particular configuration of inputs that will produce such an output. This can
be done through the Variable Elimination module of the pgmpy (Ankan and
Panda 2015). The same is also true for a combination of given input-outputs;
meaning that evidence can be given for both inputs and outputs of the BBN.
In such cases, the given inputs can also function as the design constraints.
The reverse inference demonstrates how we can utilize probabilistic models to
navigate a decision space as explained in Section 3.

Augmenting The Bayesian meta-model is capable of representing the input-

output relationships in a bidirectional and probabilistic format, illustrating a
complete example of mapping and navigating processes. However, the use of a
subset of the most influential variables of the NTA 8800 limits the navigation
to decision space made out of the selected variables. Nevertheless, this selection
was necessary to manage the computational resource-intensive task of learning.
Therefore, BBN does not compete with, or mimic the NTA 8800 model, but
rather complements it by increasing its accessibility and providing navigation
capabilities. The result is a model that can augment the designers’ intuition or
experience and enhance the level of accuracy even in otherwise vague processes of
policy formulation, e.g. in assessing the potential efficacy of alternative subsidies
and incentives for building renovation aimed at sustainable energy transition.

Binning The variables are divided into a specified number of bins based on
the frequency of the values. In equal frequency binning, the data is first sorted
in ascending order and then divided into the specified number of bins, with
each bin containing an equal number of observations. Equal frequency binning
can be useful for analyzing data that has a skewed distribution or contains
outliers. By dividing the data into equal-frequency bins, the impact of outliers
may be reduced, and the distribution of the data may be more balanced across
the bins. Considering the prediction robustness of the discretized values, each
bin should contain a fair amount of data points. A fair amount is a bit vague
description because there is no method or rule of thumb for deciding the number
of bins. However, we need to keep in mind that the number of bins corresponds
with the number of states a parameter can have, and there with the number
of parameters that are used to learn the network. Increasing the number of
bins, results in an exponential increase of computational demand. Decreasing
the number of bins, however, results in too few states to gain the desired insight
accuracy in the input-output probabilities. This means that the number of bins
should be chosen carefully, keeping in mind the trade-offs, and satisfying the
research accuracy and the robustness of the Bayesian meta-model.

Discrete, Categorical, and Numerical Variables The outcome of the

inference exercises helped to reveal the relationship between architectural lan-
guage and engineering behaviour and increases the designer’s creative intuition
to guide the design process. The design goals of the research were therefore
(1) to develop a methodology to gain general and explorative knowledge about
the association between spatial and technical building configurations and energy

24
performance in dwellings, and (2) to develop a proposal for an adapted repre-
sentation of the NTA 8800 model tool, embedded in a computational model,
to support intelligent decision making. The research workflow that has been
described is a valid approach to reaching the stated design goals.

Toy problem The results of the toy problem show the implementation pos-
sibilities, and show, of course, simplified, the possibilities for professionals to
utilize the tool in the practical field. As shown in the previous chapter, the tool
works as expected, and is capable of making preliminary predictions with very
few input variables. It appeals to the imagination when the tool is connected
to spatial configurations instead of unmeaning variables in a computer script.

5 Outlook
The ACD framework and its constituent concepts can be best positioned within
the context of performance-driven computational design and generative design.
In particular, the idea of approximating complex and non-linear functions for
estimating measurable performance indicators from configurations of decision
variables, even if referring to non-spatial decision variables, is generalisable to
all areas of computer-aided design. However, such surrogate models are not
to replace simulation models based on first principles as they can not match
their transparency and explainability. Nevertheless, in cases where one needs to
estimate the effects of design decisions on human factors, ergonomics, or com-
binations of many different types of governing equations, an estimation model
trained from actual data can be of utility in that it provides a basis for compar-
isons in the absence of analytical knowledge. In other words, the utility of ANNs
for mapping the associations between decision data points and performance data
points is apparent.
The navigation problem, on the other hand, is much harder, philosophically,
technically, and mathematically for being solved in any sense. The real advan-
tage of an AI framework in dealing with a design space navigation problem can
be attained if the latent space of the model reveals interpretable information
or if it is at least coupled with a sensible low-dimensional space. If the latent
space of e.g. an Auto Encoder (Marin et al. 2021) is understandable as a low-
dimensional vector space (as an endpoint of the mapping and the start point of
the navigating processes) then it can be used not only to guide the navigation
process but also to gain insight as to which design variables are more important
in determining the attainment levels of outcomes of interest. In other words,
even though it appears that in the mapping process, the information content
of the decision data points is reducing gradually, one can think of this process
as a distillation of an elixir from a large data point that makes the information
richer from a human perspective.
In this light, the major advantages of the proposed framework are twofold:
Firstly, providing an elegant framework for applying AI in computational de-
sign in the presence of many complex quality criteria; and secondly, providing an
elegant framework for designing spatial manifolds very much like the method-
ology of electrical engineering in designing electronic circuits and systems for
signal processing. The latter point requires much more space for discussing the
theoretical minimum for such an approach to design from a signal processing

25
standpoint. In short, however, we can briefly mention that the idea of defining
a central representation of a configuration as a discrete manifold provides for
directly modelling the functionality of the spatial manifold w.r.t. the flows of
walks, light rays, or forces not only from the point of view of spatial movement
but also much more elegantly and efficiently in the frequency or spectral domain
(which can be attained using Discrete Fourier Transform or Spectral Mesh Anal-
ysis). One fundamental idea of analogue circuit design from a signal processing
point of view is that of designing passive “filters” whose properties can much
better be understood in the so-called frequency domain analyses put forward by
Fourier and Laplace transforms of the so-called transfer functions of the RLC
(Resistor, Self-Induction Loop, Capacitor) circuits. This approach to circuit
design can be traced back to the ideas and propositions of Oliver Heaviside
(1850-1925), a self-educated pioneer of electrical engineering. Arguably, this
frequency-based outlook (relating to the spectrum of eigen frequencies of vibra-
tion of shapes, thus also identifiable as a spectral approach), has revolutionized
the formation of the field of electronics and thus contributed significantly to the
development of AI as we know it today. Identifying spectral latent spaces and
associating them with low-dimensional performance spaces and latent spaces of
ANNs is a topic that calls for further theoretical research and computational
experimentation.

References
Ankan, Ankur, and Abinash Panda. 2015. “pgmpy: Probabilistic graphical models
using python.” In Proceedings of the 14th python in science conference (scipy
2015), vol. 10. Citeseer.
Ankenman, Bruce, Barry L. Nelson, and Jeremy Staum. 2010. “Stochastic Kriging for
Simulation Metamodeling.” Operations Research 58, no. 2 (April): 371–382. https:
//doi.org/10.1287/opre.1090.0754. https://doi.org/10.1287%2Fopre.1090.0754.
Azadi, Shervin, and Pirouz Nourian. 2021. “GoDesign: A modular generative design
framework for mass-customization and optimization in architectural design.” In
Towards a new, configurable architecture, 1:285–294. Novi Sad, Serbia: CUMIN-
CAD, August.
Bai, Nan, Shervin Azadi, Pirouz Nourian, and Ana Pereira Roders. 2020. “Decision-
Making as a Social Choice Game” [in en]. In Proceedings of the 38th eCAADe
Conference, 2:10.
Barton, R.R. 1998. “Simulation metamodels.” In 1998 Winter Simulation Conference.
Proceedings (Cat. No.98CH36274). IEEE. https://doi.org/10.1109/wsc.1998.
744912. https://doi.org/10.1109%2Fwsc.1998.744912.
Batty, M. 1974. “A Theory of Markovian Design Machines.” Environment and Plan-
ning B: Planning and Design, https://doi.org/10.1068/b010125.
Bui, Van, Nam Tuan Le, Van Hoa Nguyen, Joongheon Kim, and Yeong Min Jang.
2021. “Multi-Behavior with Bottleneck Features LSTM for Load Forecasting in
Building Energy Management System.” Electronics 10, no. 9 (April): 1026. https:
//doi.org/10.3390/electronics10091026. https://doi.org/10.3390%2Felectronics
10091026.

26
Conti, Zack Xuereb, and Sawako Kaijima. 2018. “A Flexible Simulation Metamodel
for Exploring Multiple Design Spaces” [in en]. International Association for Shell
and Spatial Structures (IASS). https://doi.org/10.13140/RG.2.2.23313.53600.
http://rgdoi.net/10.13140/RG.2.2.23313.53600.
. 2021. “Explainable ML: Augmenting the interpretability of numerical simula-
tion using Bayesian networks.” In The Routledge Companion to Artificial Intelli-
gence in Architecture, 315–335. Routledge.
Fang, Kai-Tai, Runze Li, and Agus Sudjianto. 2005. Design and Modeling for Computer
Experiments. Chapman / Hall/CRC, October. https://doi.org/10.1201/9781420
034899. https://doi.org/10.1201%2F9781420034899.
Friedman, Yona. 1975. Toward a scientific architecture [in English]. First American
Edition. Cambridge, Mass: MIT Press, January. isbn: 978-0-262-06058-5.
NTA 8800. 2022. Accessed on Feb 28th, 2023, December. https://www.gebouwenergi
eprestatie.nl/bepalingsmethode.
Gero, John S, and Udo Kannengiesser. 2004. “The situated function–behaviour–
structure framework.” Design studies 25 (4): 373–391.
Giedion, Sigfried. 2009. Space, time and architecture: the growth of a new tradition.
Harvard University Press.
Gigch, J.P. van. 1991. System Design Modeling and Metamodeling. Language of science.
Plenum. isbn: 9780306437403. https://books.google.nl/books?id=M5mD0ZZcw
aEC.
Herman, Jon, and Will Usher. 2017. “SALib: An open-source Python library for Sen-
sitivity Analysis.” The Journal of Open Source Software 2, no. 9 (January): 97.
https://doi.org/10.21105/joss.00097. https://doi.org/10.21105%2Fjoss.00097.
Hicks, Charles Robert. 1964. Fundamental Concepts in the Design of Experiments.
New York: Holt, Rinehart / Winston.
Hillier, Bill. 2007. Space is the machine: a configurational theory of architecture. Space
Syntax.
James, G., D. Witten, T. Hastie, and R. Tibshirani. 2013. An Introduction to Statistical
Learning: with Applications in R. Springer Texts in Statistics. Springer New York.
isbn: 9781461471387. https://books.google.nl/books?id=qcI%5C AAAAQBAJ.
Jia, Zhuoran, Pirouz Nourian, Peter Luscuere, and Cor Wagenaar. 2023. “Spatial de-
cision support systems for hospital layout design: A review.” Journal of Building
Engineering, 106042.
Khan, Anam Nawaz, Naeem Iqbal, Rashid Ahmad, and Do-Hyeun Kim. 2021. “En-
semble Prediction Approach Based on Learning to Statistical Model for Effi-
cient Building Energy Consumption Management.” Symmetry 13, no. 3 (March):
405. https : / / doi . org / 10 . 3390 / sym13030405. https : / / doi . org / 10 . 3390 %
2Fsym13030405.
Kleijnen, Jack P. C. 1975. “A Comment on Blannings Metamodel for Sensitivity Anal-
ysis: The Regression Metamodel in Simulation.” Interfaces 5, no. 3 (May): 21–23.
https://doi.org/10.1287/inte.5.3.21. https://doi.org/10.1287%2Finte.5.3.21.
. 2008. “Low-order polynomial regression metamodels and their designs: basics.”
In Design and Analysis of Simulation Experiments, 15–71. Boston, MA: Springer
US. isbn: 978-0-387-71813-2. https://doi.org/10.1007/978-0-387-71813-2 2.

27
Koller, Daphne, and Nir Friedman. 2009. Probabilistic graphical models: principles and
techniques [in en]. Adaptive computation and machine learning. Cambridge, MA:
MIT Press. isbn: 978-0-262-01319-2.
Kroes, Peter. 2010. “Engineering and the dual nature of technical artefacts.” Cam-
bridge journal of economics 34 (1): 51–62.
Kroes, Peter, and Anthonie Meijers. 2006. “The dual nature of technical artefacts.”
Studies in History and Philosophy of Science 37 (1): 1–4.
Marin, Riccardo, Arianna Rampini, Umberto Castellani, Emanuele Rodolà, Maks Ovs-
janikov, and Simone Melzi. 2021. “Spectral Shape Recovery and Analysis Via
Data-driven Connections” [in en]. Int J Comput Vis 129, no. 10 (October): 2745–
2760. issn: 0920-5691, 1573-1405, accessed July 24, 2022. https://doi.org /10.
1007/s11263-021-01492-6. https://link.springer.com/10.1007/s11263-021-01492-
6.
Martin, Carla D, and Mason A Porter. 2012. “The extraordinary SVD.” The American
Mathematical Monthly 119 (10): 838–851.
Nen 7120+C2:2012/A1:2017 NL. 2017. Accessed on Feb 28th, 2023, June. https :
//www.nen.nl/nen-7120-c2-2012-a1-2017-nl-229670.
Nourian, Pirouz. 2016. Configraphics: Graph Theoretical Methods for Design and Anal-
ysis of Spatial Configurations [in en]. TU Delft Open, September. isbn: 978-
94-6186-720-9, accessed September 23, 2021. https : / / doi . org / 10 . 7480 / isb
n . 9789461867209. https : / / books . bk . tudelft . nl / press / catalog / book / isbn .
9789461867209.
Nourian, Pirouz, Shervin Azadi, and Robin Oval. 2023. “Generative Design in Archi-
tecture: From Mathematical Optimization to Grammatical Customization.” In
Computational Design and Digital Manufacturing, edited by Panagiotis Kyratsis,
Athanasios Manavis, and J. Paulo Davim, 1–43. Cham: Springer International
Publishing. isbn: 978-3-031-21167-6. https://doi.org/10.1007/978-3-031-21167-
6 1.
Nourian, Pirouz, Romulo Gonçalves, Sisi Zlatanova, Ken Arroyo Ohori, and Anh Vu
Vo. 2016. “Voxelization algorithms for geospatial applications: Computational
methods for voxelating spatial datasets of 3D city models containing 3D surface,
curve and point data models” [in en]. MethodsX 3 (January): 69–86. issn: 2215-
0161, accessed April 28, 2020. https : / / doi . org / 10 . 1016 / j . mex . 2016 . 01 . 001.
http://www.sciencedirect.com/science/article/pii/S2215016116000029.
O’Shaughnessy, Connor, Enrico Masoero, and Peter D. Gosling. 2021. Topology Op-
timization using the Discrete Element Method. Part 1: Methodology, Validation,
and Geometric Nonlinearity [in en]. Preprint. engrXiv, July. Accessed Decem-
ber 20, 2021. https://doi.org/10.31224/osf.io/c6ymn. https://osf.io/c6ymn.
Pearl, Judea. 1988. Probabilistic reasoning in intelligent systems: networks of plausible
inference. Morgan kaufmann.
Peng, Yun, Shenyong Zhang, and Rong Pan. 2010. “Bayesian Network Reasoning
with Uncertain Evidences.” International Journal of Uncertainty, Fuzziness and
Knowledge-Based Systems 18, no. 05 (October): 539–564. https://doi.org /10.
1142/s0218488510006696. https://doi.org/10.1142%2Fs0218488510006696.
Regenwetter, Lyle, and Faez Ahmed. 2022. Towards Goal, Feasibility, and Diversity-
Oriented Deep Generative Models in Design. ArXiv:2206.07170 [cs], June. Ac-
cessed July 22, 2022. http://arxiv.org/abs/2206.07170.

28
Regenwetter, Lyle, Amin Heyrani Nobari, and Faez Ahmed. 2022. “Deep Generative
Models in Engineering Design: A Review” [in en]. Journal of Mechanical Design
144, no. 7 (July): 071704. issn: 1050-0472, 1528-9001, accessed July 24, 2022.
https://doi.org /10.1115/1.4053859. https://asmedigitalcollection.asme.org /
mechanicaldesign/article/144/7/071704/1136676/Deep- Generative- Models- in-
Engineering-Design-A.
Sacks, Jerome, William J. Welch, Toby J. Mitchell, and Henry P. Wynn. 1989. “Design
and Analysis of Computer Experiments.” Statistical Science 4, no. 4 (November).
https : / / doi . org / 10 . 1214 / ss / 1177012413. https : / / doi . org / 10 . 1214 % 2Fss %
2F1177012413.
Simon, Herbert A. 1973. “The Structure of Ill Structured P coblems” [in en]. Artificial
Intelligence, 21.
. 2019. The Sciences of the Artificial, reissue of the third edition with a new
introduction by John Laird. MIT press.
Sobol’, Il’ya Meerovich. 1990. “On sensitivity estimation for nonlinear mathematical
models.” Matematicheskoe modelirovanie 2 (1): 112–118.
Soman, Aditya, Azadi, Shervin, and Nourian, Pirouz. 2022. “DeciGenArch: A Gener-
ative Design Methodology for Architectural Configuration via Multi-Criteria De-
cision Analysis” [in en]. In Proceedings of eCAADe 2022, 459–468. Education /
research in Computer Aided Architectural Design in Europe.
Union, Europian. 2021. Proposal for a Directive of the European Parliament and of
the Council on the Energy Performance of Buildings (Recast). Accessed on Feb
28th, 2023, December. https://eur-lex.europa.eu/legal-content/EN/ALL/?uri=
CELEX%3A52021PC0802.
Veloso, Pedro, and Ramesh Krishnamurti. 2020. “An Academy of Spatial Agents: Gen-
erating Spatial Configurations with Deep Reinforcement Learning.” September.
. 2021. “Mapping generative models for architectural design.” In The Routledge
Companion to Artificial Intelligence in Architecture, 29–58. Routledge.
Weng, Lilian. 2021. “What are diffusion models?” lilianweng.github.io (July). https:
//lilianweng.github.io/posts/2021-07-11-diffusion-models/.
Zeng, Mingliang, Xu Tan, Rui Wang, Zeqian Ju, Tao Qin, and Tie-Yan Liu. 2021.
“MusicBERT: Symbolic Music Understanding with Large-Scale Pre-Training.”
In ACL-IJCNLP 2021. June. https : / / www . microsoft . com / en - us / research /
publication / musicbert - symbolic - music - understanding - with - large - scale - pre -
training/.

29
A Acronyms

Table 1: Acronyms

Acronym Term
ACD Augmented Computational Design
AEC Architecture, Engineering, and Construction
AI Artificial Intelligence
ANN Artificial Neural Networks
BBN Bayesian Belief Networks
BEM Building Energy Modelling
BENG Bijna Energie Neutrale Gebouwen: Nearly Zero-Energy Buildings
BIM Building Information Model
CAD Computer-Aided Design
CEN Comité Européen de Normalisation: European Committee of
Normalization
CPD Conditional Probability Distribution
DAG Directed Acyclic Graph
DoE Design of Experiment
EPBD European Energy Performance of Buildings Directive
JPD Joint Probability Distributions
MAPE Mean Absolute Percentage Error
MAGMA Multi-Attribute Gradient-Driven Mass Aggregation
ML Machine Learning
NEN Nederlandse Norm: Royal Dutch Standardization Institute
NRMSE Normalized Root Mean Square Error
NTA 8800 Nederlandse Technische Afspraak (Dutch Technical Agreement)
PGM Probabilistic Graphical Models
SVD Singular Value Decomposition
VAE Variational Auto-Encoders

30
B Notation

Table 2: Notations

Notation Name Definition

x design/decision space x ∈ (0, 1]n ; each xi corresponds to a
single spatial decision variable
o performance space o ∈ (0, 1]q ; each ok corresponds to an
objective or outcome of interest
o = f (x) := map from design to f : (0, 1]n 7→ [0, 1]q ; representing a
[fk (x)]q×1 performance meta-model that approximately maps
the decision space to the performance
space
x = f −1 (o) map from performance to f −1 : [0, 1]q 7→ (0, 1]n ; pseudo-inverse of
design a meta-model that approximately maps
the performance space to the decision
space
J := [Jk,i ]q×n Jacobian matrix of f [Jk,i ]q×n = [ ∂f k
]
∂x q×n
= [∇T fq ]q×1
i
Uq×q := matrix of left singular vectors UUT = UT U = Iq×q ; ordered by
[uk ]1×q importance
Vn×n := matrix of right singular vectors VVT = VT V = In×n ; ordered by
[vi ]1×n importance
Σq×n := matrix of singular value Σ is an q × n rectangular diagonal
[uk ]1×q matrix with non-negative real numbers
on the diagonal ordered by importance,
i.e. singular values σc , c ∈ [0, min{q, n})

31