Redefining the role of

computers in architecture:
from drafting/modelling tools
to knowledge-based design
assistants
Yehuda E Kalay
The use o f computers for automating the processes of of the architectural design process promised a significant
design and manufacture promised significant improvements improvement in architects' productivity which would permit
in designer's productivity and products' quality, neither of the economic design of low quantity, high reliability, com-
which, so far, have been realized in architectural design. plex systems such as buildings.
This paper argues that in order to realize such improve- However, even though a host of computer-aided design
ments the modelling/drafting role computers have been systems have been developed in the past 20 years for use by
assigned in architectural design should be changed, so that architects, their impact on the architectural design process
computers will become intelligent assistants to designers, as a whole has been marginal. In fact, the productivity of
relieving them from the need to perform the more trivial most architectural offices that use computers has improved
design tasks and augmenting their decision making capabili- only slightly, if at all, and the complexity of buildings has
ties. been virtually unaffected by them 1-3. In comparison, the
To support this argument, architectural design is modelled use of computer-aided design systems by electrical engineers
as a search process in a space of alternative solutions, seeking has enabled them to increase the complexity of integrated
one or more solutions that satisfy certain design criteria. circuits by several orders of magnitude, while significantly
Design is shown to be a special case of general problem- reducing their design time 4.
solving processes, and thus comprised of two major com- The failure of CAD to improve architectural design
ponents: design states and the generator/test cycle that practices and products is primarily due to the role com-
facilitates transitions between them. It is then shown that puters have been assigned in the overall design process. Over
the symbolic representation capabilities of computers 90% of the systems that have been installed worldwide, so
qualify them to simulate such design states and the generate/ far, are used for drafting, which is not, in itself, an essential
test cycle, using techniques that were developed independ- step in the progress of a product from concept through
ently in the fields of geometric modelling and artificial design to production, but is rather simply a means of com-
intelligence. A conceptual framework of a knowledge-based municating between various activitiess'6.
computer-aided design system, which brings these tech- This paper sets out to redefine the role computers should
niques to bear on architectural problems, is presented, and play in architectural design from mere drafting/modelling
its potential for increasing the utility of computers in the tools to intelligent, knowledge-based design assistants. It
design of buildings is discussed. does so by identifying where in the architectural design
process computers could be used most effectively and then
artificial intelligence, design assistant, architectural design, by showing how such usage can be realized. The process of
solution states, search architectural design is discussed first, with its inherent
difficulties pointed out in order to identify those most in
need of assistance. The properties of computers are discussed
The rapid technological developments in the last 20 years next, with particular emphasis on their symbolic representa-
have led us to believe that increased productivity and more tion and simulation capabilitie~ Thirdly, it is demonstrated
efficient use of resources in our economy can be achieved that these properties are most suitable for simulating the
through automation based on computer technology. Auto- architectural design process, and a framework for such
mation of the processes by which artifacts are designed and knowledge-based computer-aided architectural design
manufactured has also been assumed to follow this trend. systems is presented.
In particular, because the design of increasingly more com-
plex artifacts requires the application of ever-growing
amounts of physical and informational resources. DESIGN AS A PROBLEM SOLVING PROCESS
Following the example set by electrical engineering in To identify the problem areas of design in general, and of
the design and fabrication of integrated circuits, automation architectural design in particular, it is necessary to describe
School of Architecture and EnvironmentalDesign, State University and to understand the process of design by means of an
of New York at Buffalo, Buffalo, NY 14214, USA abstract model.

volume 17 number 7 september 1985 0010-4485/85/070319-10 $03.00 © 1985 Butterworth & Co (Publishers) Ltd 319
 While many definitions and models of design exist 7-13, after it has been generated. In order to minimize the genera-
most agree that design is a purposeful behaviour which is tion of nonsolution states, and thus focus the search and
directed at devising artifacts or environments that attain guarantee its convergence on an acceptable solution, design
certain goals while abiding by certain constraints. Since no uses these analytical procedures not only to identify
formula exists which can translate goals and constraints solution states but also to guide the state-generating process
into a self-consistent physical form, design is an iterative, itself, by pointing out deficiencies and potentials to be
educated 'trial-and-error' process that relies heavily on developed in the current state.
knowledge and experience. As such, design shares many The second characteristic of architectural design makes
characteristics of general problem-solving processes, as it an example of what Simon termed 'problem-solving in a
defined under a theory which has been formalized by semantically rich task domain '1°. In other words, most of
researchers like Newell and Simon in the past 20 years 14'1s. architectural design's goal-directed behaviour is determined
According to this theory, for every problem we may by information which is external to the particular problem
define a solution space, that is, a domain that includes all environment. Consider, for example, the goal directed
the possible solutions to the problem. Problem-solving can behaviour in solving a maze puzzle or in playing chess, both
then be characterized as a process of searching through considered 'typical' problem solving processes. In chess, the
alternative solutions in this space in order to discover one decision about taking the next move depends entirely on
or several which meet certain goals and are, therefore, con- the current board situation, with extrapolation to subse-
sidered 'solution states'. The word 'search' is used here quent board situations as a consequence, and its objective is
metaphorically to describe a process of seeking and evaluat- to satisfy asingle goal. The design of a window, on the other
ingalternative solutions, either to the problem as a whole or hand, depends on functional and aesthetic considerations
to its subproblems, through analytical, rational, or random that incorporate information from a broad range of disci-
means. plines, such as mechanics, thermodynamics, wave theory,
The solution space of a design problem consists of a set economics, history, sociology and psychology, and its
of states, each representing a specific solution for the objective is to satisfy many goals, some of which are in con-
problem in some degree of detail. One or more of them flict with others.
represent the 'current' state of the design (allowing for These characteristics are aggravated by two additional
multiple alternatives). The process can thus be viewed as a factors. First, that the information that guides the design
sequence of actions that advance the current state from one process is always incomplete and often inaccurate. Second,
state to the next. Typically, the transition process is guided that alternative possible action-sequences may lead to many
by local or heuristic knowJedge, so as to guarantee its con- different yet acceptable solutions. Together, these two
vergence on a recognizable solution state in reasonable time factors render design a nonmonotonic, nondeterministic
(if such a state exists), and thereby bring it to a successful problem-solving process, where trade-offs between different
conclusion (see Figure ]). 'satisficing' solutions must be evaluated in order to choose
onelO, 16.
Features of architectural design Three major problems of the architectural design process
are evident even from this rather abstract description:
Architectural design is distinguished from many other
problem-solving processes by two major characteristics: • it is not known what constitutes an adequate set of para-
meters to describe a state of the design process
• the states representing candidate solutions must be • it is not known how new states can be generated from
generated before they can be evaluated existing ones
• the heuristics that guide the search rely not only on • trade-offs between dissimilar qualities are hard to make.
information internal to the particular problem, but also
on information which is external to it, for example, The first problem is further complicated by the fact that
cultural norms and styles physical artifacts do not exist in isolation. Rather, they are
always embedded in some context, and therefore must
The first characteristic of the architectural design problem-
function in concert with it. A design state thus consists of
solving process forces accomplishment of transitions through two environmental representations: an 'inner' one which
the generation of new states from the current ones, by represents the substance and organization of the artifact
means of a set of actions that assign new values to the para- (such as the structure of a building), and an 'outer' one
meters that define each state. Some of the states which which represents the context in which the artifact operates
have been generated in this manner may constitute solution (such as the socio-economic profile of the building's inten-
states, in that they attain the design goals while complying ded occupants).
with the constraints of the problem. However, since states While representing the states of design is a difficult
do not exist before they are generated, identification of one problem in itself, it is the generation of new states from
as a solution state can only be done 'after the fact', that is, existing ones which is the single most difficult problem of
architectural design. Combined with the richness of the
information which must be represented, with the need to
trade off dissimilar qualities without objective means, the
generation of a new design state is a creative process whose
practitioners often achieve the status reserved for artists.
Unlike art, however, the product of architectural design
must fulfil many physical functions. Foremost, it must
state ~ Oes~gn'stotes support and sustain the human activities for which the
building has been commissioned, and it must do so within
Figure 1. Design as a goal-directed search process the limits set by economic, physical and cultural constraints.

320 computer-aided design

SIMULATING THE DESIGN PROCESS tions by adding semantic information to the syntactic one
Computers could aid in accomplishing this complicated has been the trait of a branch of computer science known
process because, like human beings, they are information as artificial intelligence (AI), which is concerned with
processors that use abstract symbol structures to internally symbolic reasoning. AI is comprised of techniques for
simulate external environments and tasks (in varying degrees designing expert systems, natural language understanding,
of detail). By operating on these symbol structures and knowledge representation, inference, and learning, that are
manipulating them according to certain rules, computers used for achieving intelligent-like behaviour of computers by
could reason about, and even predict, the future behaviour providing them with information about the particular
of the simulated environments themselves1°. domain in which they are employed, and the means to use
One of the things computers are capable of simulating is that information effectively 2°-22. The information which is
the built environment. In order to do so, computers, or relevant to a particular domain is known as its knowledge
more precisely the symbol structures stored by them, must base, and the means to use it are known as search and infer-
be capable of simulating the two major components of the ence. Using systems in which these principles have been
goal directed, problem-solving process of design: the states implemented, computers can now provide expert counselling
of the designed environment and the generatorJtest cycle on certain medical diagnoses, oil drilling, mineral prospect-
that induces transitions from one state to the next. ing, they can play chess and backgammon, and perform an
Much progress has been made by researchers like East- increasing number of tasks that before were the prerogatives
man and Yasky 17, Rasdorf and Kutay is, and Lafue 19 in of humans23'24.
computer-simulating the design state of built environments. These techniques have been formalized under the collec-
Yet no model has been found which encompasses in a single, tive name of knowledge engineering and include simulation
comprehensive manner both the internal relationships of 'real world' environments and their manipulation in an
between the numerous components of a building, and the effective manner. The two techniques of knowledge engi-
conditions that are external to it (eg environmental, social, neering which are of particular interest from CAD's point
psychological, etc). Without such a comprehensive model, of view are:
analytical processesthat rely on it as a source of information
are bound to be limited in their scope, and unable to take • search and analysis that facilitate transition between
into account the many factors and trade-offs that affect states
certain conditions. The lack of adequate analysis, in turn, • object centred representation of data
hinders the generation of new design states that could
advance the process towards successful completion.
Heuristic search
Not surprisingly, therefore, current CAD systems avoid
virtually all design-related simulations, and provide designers Viewed as a collection of action rules, computer programs
with little more than automated rendition tools. They leave can easily execute lists of predefined tasks. If the application
the meaningful interpretation of the symbols they represent, of the rules is simple and well-defined, the program is said
along with the relationships between them, to the designer. to be algorithmic in nature: actions follow some predefined
Consequently, special purpose analysis programs require 'recipe' which, given the same initial conditions, always
many enhancements to the data before it can be processed, yield the same result. Many problems, however, cannot be
and the transition from one design state to the next must solved algorithmically, either because their solution proce-
be done by the designer with little or no assistance at all. dure is ill defined or because not all the information they
In order to employ computers more effectively in the require is available or accurate. Such problems make it
design process, their usefulness in simulating design states necessary to use less specific and more adaptive solution
must be extended beyond mere geometric and nongeometric processes (known as 'weak methods'), which typically rely
descriptive information, which comprises the syntax of on trial-and-error techniques before they arrive at any
design. It must also include the semantics (meaning) of that solution. Such techniques are, by definition, comprised of
information. Additionally, the process of analysing design search and evaluation, which are also the hallmarks of
states and generating new ones must be explicated through architectural design.
a set of rules and procedures that are capable of extracting Systematic trial-and-error, which examines all the possible
relevant information from the simulation model. It must solutions to a problem until it finds one that meets the goal
infer information which is not explicitly modelled, and criteria, is applicable only to very small problems, due to
select actions that will modify the model in the desired the exponential number of alternative solution paths. For
manner. any realistic problems the search process must somehow be
How can the representational and the operative simula- directed, so that the exploration of nonproductive steps is
tions be augmented so that computers become useful aids minimized.
in architectural design? This is the question that will be In some cases it is possible to eliminate exploration of
addressed by the remainder of this paper. To do this, we unproductive alternatives by using local information, which
shall first examine some of the tools that were developed in is inherent to the particular problem. More often, however,
order to facilitate the transition between states, namely, to it is necessary to use information drawn from a broader
direct the search for solution states, and for representing perspective, perhaps even from outside the particular
'meaning' in the symbol structures and the operators that problem domain itself. Such global information, which is
modify them. Later, we shall examine how these techniques used to guide the problem solving process, has been termed
can be used in computer-aided architectural design systems. 'heuristic knowledge' and is often encoded separately from
the specific problem-solving process itself, in terms of a
general set of rules. This information is often experimental,
TECHNIQUES FOR SEARCH AND drawn from knowledge possessedby expert designers and
REPRESENTATION accumulated over long periods of time while solving similar
Augmenting the symbolic simulation of states and transi- problems.

volume 17 number 7 september 1985 321

Meaning representation engineers for the design of integrated circuits (IC). The
Directed search comprises only that part of the design introduction of CAD tools and practices that position each
process which generates candidate solutions and determines element in the circuit, determine the most efficient way to
how well they attain the goals and abide by the constraints of interconnect its parts, simulate its behaviour and assist in its
a given design problem. To fully simulate the design process debugging, has increased the productivity of electrical
in the computer, however, the states of the artifact on which engineers by a factor of more than seven, and has increased
the generation/test processes act must also be explicitly the number of components/chip by five orders of magnitude
simulated. In other words, the objects which comprise the since 1 962 4. This success is attributed to several factors, of
artifact and the links that relate them to each other must be which design automation is only one. It has, nevertheless,
modelled in a manner which is well-formed, complete, set an example for other design and engineering disciplines
general and computationally efficient. to follow.
The state of a physical artifact can naturally be represen- The success of IC design automation is due to, and
ted by the objects it consists of: their form and other attri- largely based on, the work of Mead and Conway in forma-
butes, and the relationships between them. This is particu- lizing the methods and practices for the design and fabrica-
larly true when dealing with artifacts such as buildings, and tion of integrated circuffs 3°. Their work paved the road to
it differs from other domains of knowledge representation the application of AI techniques to IC design. In fact, the
such as natural language or image understanding where most formalization of design-check rules by Mead and Conway
of the knowledge is based in the inferences made from the made the knowledge domain of IC design a popular test
state of many independent units of data, each of which field for applying new AI methodologies. Some of the better
carries a relatively small amount of information on its own. known work in this area was done at MIT, Stanford and
Instead, the knowledge base in the case of physical design Xerox PARC. The systems that were developed there made
is object centered: objects contain both the data and the extensive use of symbolic representation of parts and
operators to modify it, as it applies to themselves. assemblies, and of constraint satisfaction for defining
The objects to be simulated consist of many important relationships between adjacent parts. Specific examples
attributes which describe their form, their composition, and include Alan Borning's THINGLAB 31, Sussman, deKeer,
the information which is relevant to their operation. These Stallman and Steele's work on constraints 32-~, Bobrow
attributes vary with regard to the difficulty of their simula- and Winograd's KRL language3s, Smalltalk 36, and Ivan
tion in computers. In particular, the 3D form attribute Sutherland's famed SKETCHPAD system 37, to name a few.
(shape) is most difficult to represent by means of the linear Borning's THI NGLAB is a system that provides an object-
symbol structures that are used in computers. Composed of oriented environment for developing simulation experiments
topology and geometry, shape information is a complex, in geometry and physics (including electrical circuits). Like
interlinked collection of spaces, surfaces, lines and points Sutherland's SKETCHPAD program, it incorporates part-
which, when combined in certain well defined ways, consti- whole and inheritance hierarchics for describing the struc-
tute a model of volumetric solid objects. The principles ture of a simulated environment, and like Sussman and
which enable such representation of physical artifacts have Steele's work, it employs constraints as a means for describ-
been developed in the past 15 years by a branch of mathe- ing relations among the parts that constitute the simulated
matics known as geometric modelling, yielding a host of environment. It is implemented in the Smalltalk 76 pro-
techniques for their implementation 2s-28. gramming language, which is an object-centered language
The representation of individual objects alone is, however, that facilitates communication between objects through
adequate to represent only the static state of artifacts that messages. SKETCHPAD, the earliest interactive general
are made of many interrelated objects, known as assemblies. purpose system for drawing and editing pictures on a com-
When viewed as an integral part of the dynamic design puter, has, in addition to its many important contributions
process, the representation of the dependencies between to computer graphics and CAD also identified the master-
objects is as important as the representation of the objects instance relationships between parts that make up the whole.
themselves. Interobject relationships, represented as links, This type of relationship has been elaborated by Eastman 29
provide the means to combine objects dynamically into and has been implemented in the VEGA geometric model-
meaningful systems and cause changes that are applied to ling system, developed at Carnegie-Mellon University by
one part of the system to have an effect on its other parts. Woodbury, Kalay and Glass38.
For example, relocating a wall may cause abutting walls to Another feature of SKETCHPAD, which was later used
stretch and shrink accordingly, and windows and doors to by Borning and by Sussman and Steele, is the specification
move along with the wall itself 29. of constraints as the means to guarantee some degree of
In summary, viewed as a problem-solving process, design consistency within the system. Constraint satisfaction
can be simulated by means of heuristic search procedures methods employed by these systems include propagating
and a network of interrelated objects. Such simulation will degrees of freedom (the one path method), and relaxation
facilitate the transition between successive design states, (the iteration method). THINGLAB has extended the
maintain their internal consistency, and guide the process to domain of constraints used by SKETCHPAD to include
a recognizable solution. The obvious question is, then, how nongeometric values, and like Sussman and Steele's CON-
can such simulation be actually implemented to produce a STRAINTS system it uses them both as descriptors and as
useful computer-aided design system? operators (a rule and a set of methods to achieve it).
THINGLAB also made extensive use of Smalltalk's
object-centered knowledge factorization, and it is, in effect,
IMPLEMENTATIONS an extension of Smalltalk. Of particular importance in this
Similar problem-solving models of the design process have context are Smalltalk's message-based constraint satisfaction
been implemented to a limited degree in a variety of CAD procedures, used to generate code that is invoked when an
systems over the last 20 years. Foremost among such imple- object is asked to propagate some change to its parts or sub-
mentations are CAD systems intended for use by electrical parts.

322 computer-aided design

 edge engineering techniques for generating and testing design
User interface
and
states, geometric modelling techniques for object represen-
tation, and highly interactive graphic user interfaces. A
®)
knowledge-
acquisition
conceptual framework for such an integrated system con-
-Database
cludes this exposition.

CONCEPTUAL FRAMEWORK FOR A

Key: o--=-b a controls b KNOWLEDGE-BASED COMPUTER-AIDED
Figure 2. Conceptual frame worh of a hnowledge-bosed
DESIGN SYSTEM
CAD system Given this problem-solving model of the design process and
the methods that can be used to simulate it in the computer,
we can now define a conceptual yet practical framework
In contrast to this AI approach to design, earlier work for a knowledge-based computer-aided design system (see
on automating the architectural design process was based Figure 2).
on applying operations research (OR) optimization tech- The two major components of this system are the plan-
niques to floorplan layout by researchers like Buffa 39, ning component and the design state representation com-
Whitehead and Elders4°, and Shaviv and Gall41. This ponent. The first is responsible for directing the process
approach is based on representing architectural design from its initiation through its successful conclusion, while
knowledge by a matrix of the affinities between various the second component keeps track of the emerging design,
activities in a building and the physical proximity between maintains its integrity, and provides the basis for the
those activities to compute the 'value' of the particular decisions made by the planning component.
spatial arrangement. A binary relocation mechanism is Each of these components is made of subcomponents:
employed to switch activities around in an attempt to planning consists of a heuristic search component and an
derive a layout that minimizes the value of the floorplan, evaluation component, which together perform the 'means-
and is, therefore, considered optimal. end' comparative analysis that simulates transitions between
Shape grammars are another approach to architectural design states. The state representation component is com-
design automation. These employ rule-based knowledge prised of objects and the links between them, and of a
representation coupled with production techniques to means to maintain the integrity of the data by propagating
transform a pattern of lines and labelled points from some
changes to other objects.
initial state to a highly complex geometric representation of
One additional component rounds out the system: it is
certain artifacts or buildings Such systems, developed the user interface, which facilitates the communication
among others by Stiny and Mitchell42'43, Flemming and
between the real world and the symbol structures in the
Downing~'4s, and by March and Steadman~'47, require
computer. This component is also responsible for the
extensive knowledge bases, and are, in effect, algorithmic system's growth through knowledge acquisition, and allows
systems that produce spatial layouts in a narrow domain of training and tuning the system by supplying information
design. that is not currently simulated or by replacing information
An expert system approach to architectural CAD and which is outdated or inadequately simulated.
to the construction industry has been discussed by Gero
et al 4~-s° and Landsdownsl, and to civil engineering by
Fenves, Rehak, and their studentss2-s4. Such systems use Planning
rule-based knowledge representation that is applied to The planning component is responsible for determining the
certain facts which describe the state of the designed artifact. strategy for solving the design problem, and ensuring that
Rules are made of premises and conjunctions of clauses, this strategy will be followed by the search process in terms
that indicate an action to be performed if the conditions of of accomplishing a sequence of goals and subgoals. The
the premises are fulfilled. Application of the rules, in a planning component is thus made of three major subcom-
certain order, replaces one fact set with another one, there- ponents:
by advancing the state of the design towards a terminal
state. The process of applying the rules may be 'opportunis- • a sequence of goals, each of which consists of a hierarchy
tic', in that the states are tested for compliance with the of subgoals, representing the 'milestones' which the
design criteria after each rule has been applied, or it may be design process should reach
'planned', in that an optimal sequence of rules is determined • a set of evaluation procedures, which determine how
first, and is then applied algorithmically to the first design well (if at all) a given design state accomplishes the
state. The methods practiced by this approach lend them- criteria and abides by the constraints prescribed by a
selves particularly to well defined and precisely represent- particular goal or subgoal
able processes, such as codes of practice and building regu- • the design process 'controller', which determines the
lation consultancy. Where judgemental and experiential goal or subgoal to be achieved next, and the method for
knowledge is required, probability and fuzzy set theory are doing so
used to support the decision-making process.
These and similar systems have pioneered the knowledge- Goals and subgoals
based approach to CAD. Yet none has attained wide The goals and the subgoals constitute the design plan and are
acceptance and they remain largely experimental. To reach comprised of both a sequential and a hierarchical ordering.
maturity, as will be proven by commercialization, a more Goals represent the accomplishment of phases that follow
comprehensive approach to CAD is needed, an approach each other (eg schematic design, followed by design develop-
that will integrate in one system features that so far have ment, followed by detailed design, etc), while subgoals
been dispersed among many. Such features include knowl- elaborate the achievements that comprise each goal (see

volume 17 number 7 september 1985 323

 Gools to determine the overall quality of the evaluated design
state. Trading off dissimilar qualities is a task that relies on
preferences that are often subjective and therefore project-
specific. Project independence of the evaluators can be
maintained by assigning the task for trading off the various
merits and disadvantages reported by them to the goals that
Subgools
the design process controller attempts to achieve, and whose
Figure 3. Goal sequence and subgoal hierarchy modular design could reflect subjective preferences by
assigning different weights to each result. Moreover, since
not all trade-offs and actions are currently known, the user
Figure 3). Goals and subgoals are formulated in terms of of the system may be called upon to exercise his own judge-
weighted criteria that must be satisfied by a design state ment in many cases.
that is said to accomplish them. Such formulation does not The hierarchical nature of the design process requires
prescribe the composition of the state itself, only its relative that analysis be performed at many different levels of
merits and tolerable drawbacks. The sequence of goals abstraction; that is, with different amounts of information.
ensures that adequate information exists at each phase of This can be facilitated by augmenting the information that
design, by virtue of achieving (completing) preceeding is available at a given level with information stored in the
phases, and ultimately reaching the solution state of design system's knowledge base, a process which is akin to using
as a whole. defaults in more conventional systems. This augmentation
The hierarchical structuring of subgoals, on the other reduces, however, the accuracy of specific evaluations, as it
hand, reduces the differences that must be bridged between uses 'assumptions' rather than actual facts. A weight des-
successive design states. By reducing these differences the cribing the accuracy of the analysis may be reported, so that
overall design problem may be reduced to a series of manage- the goals will know the degree of reliability of the data.
able tasks that can be solved by processes guided by infor-
mation which is specific to the problem (eg square footage
for a given room), and by project independent information Design process control
which is part of the system's knowledge base (eg typical Design plans, as discussed above, include both goals and
layout of given rooms). subgoals, and the means to accomplish them. It is important
Subgoals are not sequential, and not all subgoals must be to distinguish here between goals and the progresses for
achieved in order to satisfy a given goal. Since their purpose accomplishing them: a goal is an interrelated set of criteria
is to facilitate the completion of design phases rather than against which states of the design are evaluated. The proces-
prescribe the flow of the design process as a whole, different ses for accomplishing the goals comprise the design states
combinations of subgoals that have been achieved may be generator; that is, a set of specific actions that performs the
considered sufficient to satisfy the goal itself. The achieve- transitions between successive design states. When this set
ment of particular subgoals depends, therefore, on the of actions is combined with a heuristic knowledge base that
characteristics of the emerging design states Since subgoals enforces its convergence on the solution state, it becomes
may require further decomposition to bring their complexity the design process controller.
down to a manageable level, their structure resembles a tree Given the characteristics of a design state and the current
hierarchy of nested subgoals. phase of the design process, the design process controller
Different design problems require different sequences of determines which goal (or subgoal) should be achieved next,
design phases for their completion, therefore one set of goals and prescribes the means to achieve it. These means will
cannot fit all needs. Instead, modular goal-sets must be include one or a combination of the following strategies,
developed that can be substituted as called for by a parti- depending on how well the task can be defined and whether
cular problem. It is conceivable, though, that the hierarchy the information it requires is available to the system:
of subgoals is project independent, and therefore the same
subgoal hierarchy can be used for different projects. The set • algorithmic processes, as demonstrated by space alloca-
of goals, then, becomes a modular, project-specific skeletal tion programs, if the task is very well understood
plan which is determined by the designer or by means of • nonalgorithmic processes, which rely on rule-based
'metadesign' procedures, and which is elaborated by the knowledge representation (as demonstrated by shape
hierarchical, project independent subgoals. grammars) if the task is not well understood but its
possible products are
• user interaction, when the task is ill understood
Evaluation
As discussed earlier, search, by definition, is always accom- Employing evaluators for guiding the state generating process
panied by evaluative procedures that determine how well a promotes the importance of communication between the
state meets certain goals or subgoals and abides by their two processes, and is the reason for grouping state generation
accompanying constraints. The information provided by and evaluation into one subsystem. The knowledge-base
these evaluators can be used, in turn, by the design process which is associated with each goal and subgoal may invoke
controller to determine what is the next goal or subgoal certain evaluators to advise on the degree of adaptability of
that should be accomplished. the design state being considered a candidate solution state.
Since typical architectural design problems involve many By evaluating the merits and drawbacks of that state against
different goals and constraints, specific evaluators that goal criteria, the design process controller can effectively
specialize in one narrow domain each could be employed. employ a means-end analysis, where the end prescribes the
Such are energy analysis, cost estimation, structural analysis, means, that is, the process that will minimize the differences
etc. Because each evaluator is a 'specialist' in one area of between the drawbacks of the candidate state and the goals,
interest, their results must be traded off against each other and maximize its merits.

324 computer-aided design

 Design state representation of relationship more general, instances may be able to
The state representation component is responsible for selectively 'inherit' the attributes of the master, thereby
storing the symbol-structures that constitute design states, introducing a limited degree of autonomy in individual cases.
Group relationship is the most general kind of interobject
and maintaining their internal consistency. It is thus made
of two major subcomponents: link, as it binds together any number of objects that share
some common property. Such a grouping provides the
means to link objects for the purposes of change propagation
• the database itself, which stores the objects and the links
between them in both directions, a feature which the other two kinds of
relationship lack. For example, group relationship provides
• a consistency maintenance subsystem that ensures the
a means to express the perpendicularity constraint between
integrity of the database by propagating changes that
two walls in a symmetrical way, so that when either one of
have been applied to one object over the entire network
the two walls is rotated the other one will be adjusted
of objects
accordingly. Group relationship also differs from the other
The database two relation types in that it requires explicit definition of
the nature of the link. This explicit definition makes group
The database comprises the symbolic representation of
relation more general than part-whole and master-instance
objects and assemblies, and the relationships between them.
relations, where the nature of the relation is implied by the
It stores not only the objects and the links that represent
type itself. By adding a conditional component to the
design states, but also most of the knowledge that pertains to
definition (in the form of a rule set), the group relation can
them. Every object includes both descriptive and functional
exercise the link selectively, depending on the nature of the
knowledge: the descriptive knowledge is comprised of the
form and the other properties of the object, such as geo- change and the nature of the affected objects. For example,
metrical information, location, material, cost, etc. The the perpendicularity constraint may be waived for external
functional knowledge describes how the object should be walls, but enforced for internal partitions.
manipulated and used, and what the relationships are
between its various parameters. Together, these two kinds Maintaining database consistency
of knowledge constitute an 'abstract data type', a term When an object in the database is affected by some externally
coined by programming language developers to describe a imposed change, such as those initiated by the design pro-
consistent unit of information. Links between individual cess controller, the consistency maintenance subsystem will
objects transform them into a network of interrelated parts determine which other objects (if any) should also be
of one whole. affected, and how. The knowledge base of this subsystem is
The links between objects of the database can be classi- thus capable of traversing the relationship network, as well
fied according to the types of relationships they simulate as resolving such problems as circularity and priority.
(as is shown in Figure 4): Circularity is a problem that may arise when a change
returns to the originating object after having been propa-
• part-whole relationship gated through the network, due to cyclic links. Priority is a
• master-instance relationship problem that may arise when one object receives two (or
• group relationship more) conflicting change messageswithin a single propaga-
tion cycle, due to membership in more than one relation
Part-whole relationship links objects in a hierarchical struc- group. The solution of both problems is similar: the number
ture. A window, for example, is part of a wall, which is part of changes an object may undergo in one propagation cycle
of a room, etc. Part-whole relationship provides a natural may be limited to at most one, or a recursive 'relaxation'
means for propagating Iocational change information; when technique could be employed which applies changes until
the parent object is relocated, so are all its children. some termination criteria has been satisfied. Of course, it
Master-instance relationship defines a type or variable may also happen that a conflict cannot be resolved by the
kind of association, where every object belongs to a class of consistency maintenance subsystem, in which case the user
objects that defines its generic parametric characteristics in will have to be consulted.
terms of both data and operations. While individual objects As an example of the use of abstract data types as a
may have different values assigned to their parameters such means to automatically maintain the internal consistency of
as size, location and orientation, they still share some a complex symbol structure, consider the boundary repre-
common characteristics and the internal, nonparametric sentation models that were developed by Eastman, Weiler,
structure of the class (eg they are all parallelepiped shapes). and Baumgart, for representing and manipulating the form
Such a relationship enhances the integrity of the database of physical artifacts when simulated by computersss-s~.
by eliminating redundancy and by introducing some degree Such structures consist of many objects, which include
of uniformity. It also provides another means for change- faces, edges and vertices, and many links that interconnect
propagation: by changing some generic attributes of the them into one structure. The high degree of interconnect-
master all its instances are also changed. To make this kind ivity of such structures causes every change to have a
'rippling' effect on many adjacent objects. For example, the
deletion of an edge may cause subsequent deletion of an
adjacent face, or the creation of a 'hole' in another face
EZZ3E3 (depending upon the particular circumstances), in a manner
E 3p,_,- 1 that will guarantee the compliance of the symbolic structure
with a modified version of Euler's Law (number of faces
minus number of edges plus number of vertices equals two).
Master objects Instances Relation- The maintenance of this complex symbol structure is per-
groups
formed by a set of operators known as the Euler Operators,
Figure 4. Objects and/inks

volume 17 number 7 september 1985 325

which propagate the implications of high level data manipu- in the planning component, by observing the sequence
lation transactions over the entire network of objects. users prefer in progressing through goals and subgoals.
In the case of an open ended database which is comprised In addition to fulfilling the important role of keeping the
of many different objects that do not abide by one pre- system current, the knowledge acquisition mechanism also
defined set of rules such as Euler's Law, the means for reflects (simulates) the nature of design as an evolving
propagating change information must reside with the objects process which constantly requires modification of knowl-
themselves, rather than with an external set of operators. edge and revision of methods.
More specifically, each object must encode externally signi-
ficant change parameters as 'messages' that are communica-
SUMMARY
ted to the objects it is directly linked to. These objects, in
turn, will decode the message and, if relevant to their own We have examined the architectural design process from a
state, will act upon it. Both relevancy and action can be problem-solving point of view, and seen that it can be
determined by the functional knowledge which is specific modelled as a search process in a space of alternative
to each object, thereby further enhancing the importance solutions, seeking one or more states that satisfy certain
and usefulness of object-centred data representation. design criteria. Since the solution states do not yet exist,
the search process involves a solution generator and a
solution evaluator that determine the degree of adaptability
User interface and knowledge acquisition
of the generated solutions to the design criteria, as well as
The discussion presented so far in this paper may lead to the guide the subsequent solution generation cycle.
mistaken perception that through implementing knowledge- We have also seen that computers, through the symbol
based systems for design the role of the human designer in structures stored in them, can simulate both states and
the process will diminish, perhaps to the extent that he will transitions and thereby the design process as a whole, The
become unnecessary altogether. This, however, is not the complexity of the task-domain which is of interest to us,
case: the complexity of the design process practically that is, the built environment, makes the representation of
precludes its complete formal simulation, because many states and transitions an extremely complicated task.
important parts remain unknown. For example, though The concepts that have been discussed in this paper, and
algorithms for translating a building program into a bubble the conceptual framework that was developed to support
diagram, then into a spatial arrangement of rooms, have them, demonstrate two things:
been known since the late 1960's, none comes even close to
an actual architectural floor plan (let alone a fully 3D • a model for the design process can be constructed
building), and thus they provide only an aid to human • it can be simulated in a computer system
designers who must take the extra step(s) for completing
the design sa'sg. Furthermore, many trade-off decisions, These discussions led to the development of a conceptual
which are characteristic of architectural design, are based framework for knowledge-based computer-aided design
on subjective preferences that are dependent on particular systems. Such systems will embody techniques of search as
situations, and thus become evident only as the design developed by AI, and the principles of representing physical
process unfolds. For example, designers capitalize on artifacts as developed by geometric modellin& Augmented
particularly attractive merits of the emerging designed build- with specific skeletal plans and evaluations, such systems
ing, and therefore trade away qualities that otherwise might will be capable of assisting designers at a much higher level
have been preserved 6°. of the design process than current CAD systems do.
More specifically, the very complexity of the design Knowledge-based CAD systems can assist designers in
process itself renders CAD systems mere assistants to human two ways:
designers, whose role is to relieve the designer from perform-
ing the more trivial tasks in the process and maintain its • they could fully model the artifact that is being designed
integrity. An effective and friendly user interface between and automatically maintain the semantic integrity of the
the computer and the human designer is, therefore, manda- model
tory. This interface allows the designer to monitor (graphi- • they could guide the designer through the planning
cally and analytically) the progression of the emerging process and provide him with informative feedback on
design and affect it so that it reflects his preferences. design decisions
Furthermore, the user interface enables inclusion of the
designer in the generate/test cycle, a feature which is essen- These capabilities will increase the utility of time spent by
tial since the system is incapable of performing all the designers (and thereby improve their productivity) by
necessary transitions between design states on its own. relieving them from the need to explicitly propagate the
The user interface fulfils another important role: it implications of primary design decisions over related parts
provides the conduit through which the knowledge base of of the project, and the need to attend to and manage many
the system can be augmented, modified and updated. The secondary and tertiary design tasks of supportive rather
ability to acquire new knowledge and update existing than primary design nature.
knowledge is a necessary feature to prevent early obsoles- Furthermore, since the designer will have more time to
cence of the system, and to improve its utility. By monitor- concentrate on defining the specifications for the design of
ing the transactions initiated by the user, new knowledge artifacts and environments, their suitability for intended
can be incorporated in the system's knowledge base, a usage will be improved, and the delegation of responsibility
process which is akin to learning by humans. For example, for maintaining the integrity of the data (through the auto-
a novel type of furniture or room layout that has been matic or user-interactive resolution of conflicts and uncoor-
devised by the user could be incorporated into the system dinated decisions) will minimize costly errors that otherwise
as an object class, if it conforms to certain functional might escape notice until it is too late or too expensive to
constraints. Similar improvements can be incorporated correct them.

326 computer-aided design

 The higher productivity of designers and the freedom 17 Eastman, C M and Yasky, Y 'The integrated building
from spending much time on detailed solutions may allow model and database scheme for the 2nd phase of
them to explore more alternative design solutions in the integrated CAEADS' Technical Report Institute of
time they used to spend on fewer ones, thereby increasing Building Sciences, Carnegie-Mellon University
the possibility of finding a better solution to the design (February 1981)
problem as a whole.
18 Rasdorf, W J and Kutay, A R 'Maintenance of integrity
during concurrent access in a building design database'
ACKNOWLEDGEMENTS Comput.~aided Des. Vol 16 No 4 (July 1982)
The author thanks Anton Harfmann, Henry Konigsberger pp 201-207
and Lucian Swerdloff for the helpful discussions leading to 19 Lafue, G M E 'Integrating language and database for
the writing of this paper, and the referees for their critical CAD applications' Comput.-aided Des. Vol 11 No 3
comments on its earlier draft. (May 1979) pp 127-130
20 Sowa, J F Conceptual structures: information process-
REFERENCES ing in mind and machine Addison-Wesley Publishing Co,
1 Raker, D S 'Architectural design with computer Reading, MA, USA (1978)
graphics: linking automation with intuition - how to 21 Nilsson, N J Principals of artificial intelligence, Tioga
apply logic to creative work' Computer Graphic News Publishing Co, Palo Alto, CA, USA (1980)
(September/October 1982) pp 6-7 22 Winston, P H Artificial intelligence, Addison-Wesley
2 Mileaf, H (moderator) 'How far have we come, how far Publishing Co, Reading, MA, USA (second edition)
are we going, and who will benefit from the revolution? (1 984)
Round Table in Architectural Record (May 1983) 23 Buchanan, B G and Shortliffe, E H Rule-based expert
pp 39-53 systems: the M YCIN experiments of the Stanford
3 Mitchell, W J 'What was computer-aided design?' heuristic programming project Addison-Wesley
Progressive Architecture (May 1984) Publishing Co, Reading, MA, USA (1984)
4 Mayo, J S 'Design automation: key to future techno- 24 Feigenbaum, E A and McCorduck, P The fifth genera-
logy' High Technology (November 1983) pp 17-30 tion: artificial intelligence and Japan's computer
5 Orr, I N 'The merits of design automation' Computer challenge to the world Addison-Wesley Publishing Co,
Graphics World (January 1985) Reading, MA, USA (1983)
6 Bloom, G L 'Justifying the use of CAD in architecture:
25 Baer, A, Eastman, C M and Henrion, M 'Geometric
taking advantage of repetition requirements' Computer
modelling: a survey' Comput.-aided De~ Vol 11 No 5
Graphics World (March 1982) pp 79-80
(September 1972) pp 253-272
7 Mitchell, W l Computer-aided architectural design
26 Requicha, A A G 'Representations for rigid solids:
Petrocelli/Charter, NY, USA (1977)
theory, methods and systems' Comput. Surveys Vol 12
8 Akin, O 'Representation and architecture' in Akin and No 4 (December 1980) pp 437-464
Weinel (ads) Representation and Architecture Informa-
27 Eastman, C M and Preiss, K 'A review of solid shape
tion Dynamics Inc, Silver Spring MD, USA (1982)
modelling based on integrity verification' Comput.-
9 Eastman, C M 'Recent developments in representation aided Des. Vol 16 No 2 (March 1984) pp 66-80
in the science of design' Research Report No 83,
28 Braid, I C 'The synthesis of solids bounded by many
Institute of Physical Planning, Carnegie-Mellon
University (April 1981) faces' Communication, ACM Vol 18 No 4 (April 1975)
pp 209-216
10 Simon, H A The sciences of the artificial, MIT Press,
Cambridge, MA, USA (1969) 29 Eastman, C M 'The design of assemblies' Technical
Report 11 Institute of Building Sciences, Carnegie-
11 Akin, O 'How do architects design?' in Latombe (ed) Mellon University (October 1980)
Artificial intelligence and pattern recognition in
computer-aided design I FIPS, North Holland Publishing 30 Mead, C and Conway, L Introduction to VLSI systems
Co (1978) Addison-Wesley Publishing Co, Reading, MA, USA
(1980)
12 Freeman, P and Newell, A 'A model for functional
reasoning in design' in Proc. Int. Joint Conf, Artificial 31 Borning, A 'THINGLAB - a constraint-oriented
Intelligence (1971 ) simulation laboratory' Technical Report, STAN-CS-79-
746, Stanford University (July 1979)
13 Akin, O Models of architectural knowledge: an inform-
32 Stallman, R M and Sussman, G J 'Forward referencing
ation processing model of design, PhD Thesis, Carnegie-
Mellon University (1979) and dependency-directed backtracking in a system for
computer-aided circuit analysis' A I Journal (September
14 Simon, H A Models of thought Yale University Press, 1977) pp 135-196
New Haven, CT, USA (1979)
33 Sussman, G J and Steele, G 'CONSTRAINTS - a
15 Newell, A and Simon, H Human problem solving language for expressing almost-hierarchical descriptions'
Prentice-Hall, Engelwood Cliffs, NI, USA (1972) Artificial intelligence (August 1980) pp 1-40
16 Mackinder, M and Marvin, H 'Design decision-making 34 deKeer, J and Sussman, G J 'Propagating constraints
in architectural practice' Building Research Establish- applied to circuit synthesis' MIT AI Lab Memo 485
ment, UK Department of Environment (July 1982) (September 1978)

volume 17 number 7 september 1985 327

35 Bobrow, D G and Winograd, T 'An overview of KRL, a architecture' in Proc. CAD 84 Butterworths, UK
knowledge representation language' Cognition Science (1984) pp 529--546
Vol 1 No 1 (1977) pp 3-46 50 Gero, J S 'Expert systems in design and analysis'
36 I ngalls, O H H 'The Smalltalk-76 programming language: Technical Report Department of Arch. Science,
design and implementation' in Proc. 5th AnnualACM University of Sydney, Australia (1985)
Symposium on Principles of Programming Languages 51 Landsdown, J Expert systems: their impact on the
Tucson AZ, USA (January 1978) pp 9-16 construction industry RIBA, London (1982)
37 Sutherland, I E SKETCHPAD: aman-machinegraphical 52 Sriram, D, Maher, M k and Fenves, S J 'Knowledge-
communication system PhD Thesis, M IT, Cambridge,
based expert systems in structural design' Carnegie-
MA, USA (I963) Mellon University, Pittsburgh, PA, USA (December
38 Woodburry, R F and Glass, G J 'VEGA: a geometric 1984)
modelling system' in Proc. Graphic Interfaces Edmon- 53 Maher, M L 'HI-RISE: a knowledge-based expert
ton, Alberta, Canada (May 1983) pp 103-109
system for the preliminary structural design of high
39 Buffa, E S, Amour, G C and Vollman, T E 'Allocating rise buildings' PhD Dissertation Carnegie-Mellon
facilities with CRAFT' Harvard Business Review University, Pittsburgh, PA, USA (December 1984)
Vol 42 (November 1969) 54 Rehak, D R, Howard, H C and Sriram, D 'An integrated
40 Whitehead, B and Eldars, M Z 'An approach to the knowledge-based structural engineering environment in
optimum layout of single-storey buildings' Architects Proc. IFIPS WG5.2 Working Conference on Knowledge
Journal (I963) Engineering in CAD (1984)
41 Shaviv, E and Gali, D 'A model of space allocation in 55 Eastman, C M and Weiler, K 'Geometric modelling
complex buildings' Build International (1 une 1979) using the Euler operators' in Proc. Conf. on Computer
42 Stiny, G 'Introduction to shapes and shape grammars' Graphics in CAD/CAM Systems (May 1979)
Environment and Planning B (July 1980) pp 343-351 pp 248-259
43 Stiny, G and Mitchell, W ] 'The palladian grammar' 56 Weiler, K 'Edge-based data structures for solid modelling
Environment and Planning B (May 1978) pp 5-18 in curved surface environments' IEEE Computer
Graphics and Applications (January 1985) pp 21-40
44 Flemming, U 'Structure in bungalow plans' Environ-
ment and Planning B (August 1981) pp 393-404 57 Baumgart, B 'Winged-edge polyhedron representation'
Stanford Artificial Intelligence Report No CS-320
45 Downing, F and Flemming, U 'The bungalows of (October 1972)
buffalo' Environment and Planning B (August 1981 )
58 Hashinshoni, R, Shaviv, E and Wachman, A 'Elaboration
46 March, L and Steadman, P Thegeometr2 of the of a graph into planar layout' Technion Research and
environment MIT Press, Cambridge, MA, USA (1974) Development Foundation Ltd Report No 020-176
47 March, L (ed) The architecture of form Cambridge (may 1977)
University Press, Cambridge, MA, USA (1976) 59 Baybars, I and Eastman, C M 'Enumerating architec-
48 Gero, J S, Radford, A D, Coyne, R and Akiner, V T tural arrangements by generating their underlying
'Knowledge-based computer-aided architectural design' graphs' Environment and Planning B Vol 7 (1980)
in Proc. IFIPS Working Conference on Knowledge pp 289-310
Engineering in Computer-aided Design, Bud apest, 60 Schon, D A The reflective practitioner - how profes-
Hungary (1984) sionals think in action Basic Books Inc, New York
49 Gero, J S and Coyne, R The place of expert systems in (1983)

328 computer-aided design