Research in Design Thinking, 1992

Huts, ships and bottleracks: Design by analogy for architects and/or machines

Alexander Tzonis, Faculty of Architecture
Delft University of Technology Berlageweg Delft
The Netherlands

A first version of this paper appeared in: Tzonis, A. 1990.
"Hütten, Schiffe, und Flaschengestelle" Archithese 20 (3), pp 16- 27

Elementarist allocations: the analytical paradigm.
It has been almost a quarter of a century since the computer made its glamorous entrance into architecture with the publication of two seminal works, Serge Chermayeff's and Christopher Alexander's Community and Privacy (1963), followed a year later by Christopher Alexander's Notes on the Synthesis of Form (1964).
In spite of this enthusiastic beginning, the relation between architecture and the computer has not been a steady one. There has been a period of excitement, intensive efforts and experimentation which took place mainly during the decade of the nineteen sixties, which was succeeded by a decade of neglect to be finally followed by the current era of renewed interest.
Why have the relations between architecture and computer gone through this full cycle? Is the computer here to stay, and in what capacity? Can it help improve design? Should the failures of the past be left behind or are they something to examine seriously for the valuable information they can provide for the future of the computer in architecture? It seems to me that a brief retrospective look will be useful before responding to these questions.
We will start with Chermayeff's and Alexander's books. They are significant not only because they were first in introducing the computer, but also because they had a lasting paradigmatic impact on architecture: their views shaped the framework within which people thought about what and how computers could contribute in design over a period of the last thirty years and they continue to do so today. We will call this framework the analytical paradigm.
Here is how Chermayeff-Alexander looked at architectural design: First of all they criticized architecture for following received truths. Designers should not rely on the "memory" of "obsolete" design products "obscured by semantic misapprehension". They should start instead with the "redefinition" of a new, "modern", "progressive vocabulary", with a "list" of elementary functional "requirements", expressed in the form of requirements that a building had to fulfil. Chermayeff and Alexander identified, in their book, thirty three such requirements. The list by itself "only half designed" a building. To complete the process, Chermayeff and Alexander continued, "we must look at the links", the "interactions" between these elementary requirements. This "pattern" of interactions had to be processed further
also. In order to explicitly identify the most tightly interlinked requirements it was "split apart at its interstices". (Chermayeff and Alexander, 1963, pp. 155-167).
Placing the method in a broader context, Alexander (1964) claimed that in primitive, "unselfconscious" societies the handling of "pattern of interactions" is easy because people deal with only a handful of requirements. But today, even with thirty three requirements to consider, one has to examine "some 10.000.000.000 different cleavages". And this cannot be grasped by the eye. Fortunately, Chermayeff and Alexander asserted, we have electronic computers that, while unable to invent, can
identify patterns. A computer, an IBM 704, had come in.
The first program for grouping or "decomposing" these interacting requirements was HIDECS, developed in 1962 by Christopher Alexander and Marvin Manheim at the Massachusetts Institute of Technology. The interaction matrix was used as "input" for a "decomposition process" which through step by step division and partition resulted in a number of subsets of requirements expressed in the form of what is called a "tree" graph, a schematic association of elementary assertions which, in a tree like manner, branches out hierarchically.
Subsequently the designer was left to interpret intuitively this tree graph pattern into a topological space pattern that linked places and groups of places with each other. This spatial pattern looked very much like a rough drawing of a building and as a result an instant folk myth emerged that the computer could design correct buildings by itself as long as there was an input of correct requirements. In fact, the computer had carried out an important but extremely limited task. It had only sorted out interactions between requirements and clustered them in a logical graph representing compatibilities between these requirements. When in the next step this graph was translated into a space topological diagram the process was carried out "by hand", that is by traditional intuitive architectural thinking and not by the machine. Still, it was a major breakthrough. The computer had finally entered architecture.
The analytical heritage
Chermayeff's and Alexander's paradigm of the 1960's can be inscribed within a long tradition of design optimization efforts that extends from the military architecture of the sixteenth century to the World War II Operations Research space planning. This architectural tradition belongs, in turn, to a broader movement in history, the movement to rationalize thinking through language reform. It goes back to Raymond lull's Ars Combinatoria and to leibniz's project for an "alphabet of human thinking" and, closer to our time, to the work of Frege, the Vienna Circle and the logical positivist movement to "replace" "low", everyday as well as "high" intuitive, metaphysical expressions by a "newly constructed more exact" language (Carnap, 1947) "built it up in a consistent manner from uniform simple elements", (von Mieses, 1951). We will refer to this approach as the analytical heritage.
The highest point of triumph of the analytical heritage arrived with the appearance of the computer. Without it the Chermayeff-Alexander break through, developing a new computer-based design would not have been possible.
But together with all the benefits the Chermayeff-Alexander paradigm acquired from the analytical heritage, precision, coherence, control, all of an unprecedented degree in the history of architecture, there were also basic weakness:
1. The paradigm was limited by lack of effectiveness of its products. 2. It led to an inefficient design process.
Unfortunately, most attempts to improve effectiveness only succeeded in making efficiency worse and vice versa. Thus in search of more robustness and more freedom from historical prejudice architects were trapped powerless in a true Scylla and Charybdis dilemma. How did that happen?

Scylla and Charybdis of analytical design odyssey
The work which followed the Chermayeff-Alexander paradigm was directed towards a more genuinely computer based design. Intuitive thinking was permitted only in the beginning of the design process, and in a very explicit way, the collection of data to be entered in the list of requirements. The matrix of interaction between requirements was replaced by a matrix that represented anticipated circulation between spaces of a building. The resulting patterns of groups of spaces expressed a pattern of spatial associations, that the program implemented further into spatial patterns.1
"Space Allocation" (S.A.) techniques, as they came to be known, were more pragmatic than the method proposed by Chermayeff and Alexander. Gone were the critical references to "middle class" architectural culture that Chermayeff and Alexander had made. S.A. techniques were developed to handle "real" data such as the size, shape and cost of accessibility.2
Yet, whatever the disagreements between the Chermayeff Alexander method and the S.A. techniques both shared the same analytical paradigm. It can be summarised as follows:
1. Design techniques should be conceived as a general-purpose method, independent of the particularities of a given project.
2. Design should depart from elementary requirements, their relations tabulated in a matrix form.
3. In addition to this matrix only a small number of restricting rules are needed to define a design solution.
4. Generation of a design is a quantitative, combinatorial problem defined by the matrix of interactions plus a few restricting rules, a problem which can always be solved with the aid of the computer.
Alas the products of Space Allocation techniques, truly "automation generated plans", proved ineffective. They looked more like patchwork than real buildings. They were of little use for designers.To overcome this lack of realism, Space Allocation techniques were soon modified incrementally through the introduction of more rules expressing zoning orientation, room alignment, proportions.3
But while these new facts promised more effective results. The
complexity of computer based inference begun to be sensed.4 And it was a complexity that no heuristics and powerful algorithms could cut down. In a seemingly simple Space Allocation problem scaled down to the design of an eight locations facility, considering only three objectives, and only the topological level of the solution, Robert Berwick (1972) studied the "combinatorial explosion", the counter to intuition intractability of the problem. It was a phenomenon well known to Artificiallntelligenc~ engineers, but not to many other engineers and certainly not to architects, at that time.
By the middle of the 1970's the still young field of design computer applications was faced with a paradox, one that computer scientists had already discovered by the end of the nineteen fifties,that analytical computer techniques confronted with a single task and a well-defined set of data could optimise highly sophisticated lay-out problems such as the plan of a warehouse, but was no help in creatively arranging an ordinary living room, a task which even illiterate people could easily carry out. Unfortunately, most designers paid attention to the apparent incompetence of Space Allocation techniques blaming for that "the machine" disregarding the investigation of the causes.
Thus the nineteen-seventies were characterised by benign neglect of computer based design with one exception, the so called Computer Aided Design (C.A.D.), which came to stand for a family of techniques to carry out drafting automatically. What was said about Space Allocation techniques can be repeated about C.A.D. .C.A.D. drafting techniques could execute feats of drawing so tedious that no human being could endure them, but could not carry out the most simple design act such as adapting a simple load bearing structure to a plain piping network because C.A.D., techniques like those of S.A., performed well numerically, parametrically but had no architectural knowledge incorporated in them. They both used the computer as a kind of large scale calculator. Was there any life for the computer beyond that? Was there an alternative?
Turning Julius on his head?
Any attempts to redeem the weakness of the analytical paradigm confronted the dilemma of having to choose between improving effectiveness at the expense of efficiency or vice versa. The problem was not one of weak computers, of feeble mathematics. There was enough evidence to suspect that the limitations of the computer-based architectural techniques resulted from the very foundations of the analytical paradigm. But what was wrong with it? Which part of these foundations caused the trouble? Could it be that the paradigm:
1. had assumed very little constraints in advance, in the design process? 2. more specifically, that it employed minima! architectural knowledge? 3. the process relied on the computer as a large scale calculator? 4. that it ignored aspects of time and memory?
Could one say in this case that what was needed was an altogether different paradigm, and could this be perhaps one that turned Lullus on his head, one that permitted:
1. strong constraints to be employed in advance to generate architectural plans?
2. architectural knowledge to playa significant role in the design process as a means to introduce such constraints?
3. the computer to be used as an instrument for reasoning, using, perhaps, symbolic systems to realize intelligent inferences.
At the beginning of the mid nineteen-eighties the spotlights rivetted once more onto the computer. There were many good reasons for this. Computers had become more potent, easier and cheaper. Pressures to make design firms more productive were mounting. It was asked: Why cannot computers help design as they do in other areas? Numerical, parametric automated design is not enough. Decisions on the pre-parametric level of design determining configurational aspects of an artifact -what in fact Chermayeff and Alexander were initially concerned- is too significant to be left unchallenged by the machine. There appears a real, pragmatic need to go beyond the initial computer based analytical techniques. And urgently.
Once more the question emerged: how should we use computers in architecture? And more specifically, how can we go beyond analytical
design techniques, how can we develop a new paradigm for computer based architecture?
Our critical review of the analytical paradigm can help us perhaps find out what a new design paradigm might be like, a paradigm competent to
produce plans nearing the intelligence of architects who, for example can
make a building scheme to fit the context of its neighbouring buildings externally while internally fits the life style of its users, or how to produce an intelligent program which can recognise lines as representations of artifacts, that can draw pipes around construction without piercing it as a smart robot walk around a landscape without colliding.

The new paradigm: Artificial intelligence for the Intelligent architect
To talk about intelligent machines and programs we should first talk about architectural intelligence:
1. What is the behaviour of an intelligent designer like?
2. What creates this behaviour, what is the competence intelligent architects are endowed with?
We will start, jotting down a few observations about intelligent architectural behaviour. First is a list of things good architects, "experts" do. Compare it to what analytical techniques do:
1. Architects create, fast, a rough global organisation of a plan, a sketch, a preparametric spatial scheme that captures essential characteristics without loosing themselves in details of local problems. After that they put the sketch to test and then they proceed with specific details. Analytical techniques cannot generate global schemes. They can only solve local
problems.
2. Architects achieve a good balance between efficient design process and effective product. Analytical computer based techniques either lead to unrealistic computation processes -computation explosions -or to project solutions that are unrealistic and thus undesirable.
3. Architects generate total project solutions that simultaneously satisfy many requirements, many functional points of view. They use data that might change in time. Partial, good only for limited requirements and restricted static conditions are the design products of analytical computer based techniques.
4. Architects produce satisfactory project solutions under very
unsatisfactory conditions, incomplete data, implicit instructions, missing specifics. On the contrary analytical computer based techniques break down and cannot give any result under such onerous circumstances.
5. Architects give answers to common design problems easily and to rare ones with difficulty. That seems reasonable but the reverse is true for analytical computer based techniques. They specialize on highly specialized problems but fail to respond to the most frequently encountered everyday ones.
The behaviour of good expert architects differs from that of analytical techniques not only in terms of the kinds of tasks they fulfil, but also how they go about fulfilling them.
1. Architects appear to use experience extensively. They exploit precedents, of theirs or others, solutions developed as answers to previous problems. They use knowledge, rules of thumb, general rules, theories and principles. The knowledge they use is transferred from one kind of problem to another, from one design domain, or even world domain, to another.
2. Architects infer their design solutions reasoning qualitatively, combining information of many levels of abstraction and of different degrees of precision few of which are numerically expressed.
3. Architects keep track of what they produce and change their minds about how to design. They are informed about their own successes and failures.
Analytical techniques seem not to follow such methods. They do not
apply prior experience, knowledge. They do not work with heterogeneous and fuzzy data. They do not reason qualitatively. They are rigid in the way they solve problems. They do not make leaps in their reasoning. They do not judge their own products. They do not learn.
Take the case of Le Corbusier's Unite d' Habitation, the"pre-parametric" way he conceived the building's spatial concept, grasping from the outset, seemingly effortlessly and spontaneously, its fundamental aspects, inventing a truly complex multi-functional unprecedented form synthesised out of and in analogy to, a multitude of precedents: the savage hut, the liner, the winebottlerack, the Greek temple, and more, recalling these precedents from memory, examining them, dissecting them, trying them and recombining them, putting old tools to new uses and old ones in new compositions.

Although, in many respects, extreme, the case of Le Corbusier's invention of the Unite d'Habitation is actually far from being unique in history. In an other occasion (Tzonis, 1989) we had the opportunity to look in depth, document and analyse the creation of the triangular bastion in the Renaissance by Leonardo da Vinci. In both cases we find the same kind of characteristic intelligence. We find it also in the work of the majority of designers although to a lesser degree. How can the machine be intelligent in a similar manner, a fact that will make the whole effort of using computers in architectural offices more worth wile. Doesn't the fact that architects succeed using precedents, rules and principles rather than departing from a tabula rasa, imply an approach? Doesn't their using qualitative as well as quantitative computation, and analogical, preparametric conception rather than analytical, parametric calculation, suggest a direction?
From the computational point of view it is obvious that to mobilize architectural knowledge, either in the form of architectural precedents,
rules, or principles, at the very beginning of the design process, implies constraints which if appropriately integrated can limit the search for the appropriate form of a product. Architectural knowledge could create a more efficient design process. The introduction in computational design of architectural knowledge, also guarantees that the products of this process will be realistic, effective. In other words the form of the proposed artifact will satisfy many requirements simultaneously and not just a reductive abstract point of view.
What kind of computer based design system can generate such behaviour? How can it possess recall, and process the information such precedents, principles and rules contain as incoming programmatic design demands for new projects arrive? How can it develop design concepts, plan sketches, preparametric schemes, out of a mixed bag of qualitative, fuzzy,
incomplete data and even contradictory facts? What are the tools we have at our disposal to mobilise towards such an intelligent architecture?
A frame for representation of architectural knowledge
We will start responding to these questions with the problem of representation of architectural knowledge, the basic concepts and
structures which capture information contained in precedents, principles, rules of architecture. We will then try to identify how these fit into a reasoning mechanism that departs from a so called "programme of architectural needs", exploits knowledge and leads to design products. Like good choice of paradigm, good choice of representation can save time and memory more than algorithms and hardware can (Amarel, 1968).
What is the bare minimum of design domain knowledge? Which are the basics of representations of the architectural world that govern design intelligence? Which are the categories through which designers reason about artifacts schematically on the preparametric, "sketch" level? (Ulrich, 1988)
A core of an intelligent design system should represent significant aspects of how artifacts are made out, how they work, what they do in respect to what has to be done, how they fit into the environment,and how all these aspects relate to each other. It must be able to develop descriptions of facets of these aspects that can enter into reasoning patterns.
The most obvious place to start putting together such a system is how a design product is made, the artifact's form, morphology of a building, its characteristic attributes, its spatial composition and its material structure. We will use the word form here to denote configurational attributes as well as the physical constitution of design products.
Seen from a different perspective, that of the physiology so to speak, of architectural artifacts, buildings contain operations. The form of buildings controls, holds or channels, people, objects, equipment associated with activities.
From yet another point, the teleological one, we look at buildings in terms of their anticipated as well as their actual performance.
We finally consider all the above, form, operation, performance of a design product, in reference to the context within which the artifact is to be realised.
The reader must have noticed that the concept of function does not appear in our system here. The reason for that is its ambiguity among architects. Architects refer to the "function" of a corridor, for example, as being circulation -an operation according to our terminology. They also talk about the "functional" merits of a scheme of a corridor in providing privacy- an anticipated performance according to our terminology. (Zandi-Nia, 1992).
Form, operation, performance and context are interrelated. This interrelationship can be expressed in constraints that state which performance of a building may result from which operation and, in turn, which operation may result form which form, a rule chain whose links are neither deterministic nor closed. The performance of an artifact may depend on external conditions, conditions that apply to its operation, as the operation itself may depend on external conditions attached to the artifact's
form.
What kind of reasoning do designers use in practice expressing constraints linking form, operation, performance and context to each other?
First, given the performance of a building they try to explain the aspects of operation and form -that may have influenced it, or given the operation of a building, which aspects of its form may have affected its operation. ( Davis and Hamscher, 1990)
From such an explanation one can arrive at what can be called design diagnostics, which identify causes -operational and formal -for building pathology unanticipated performance of artifacts. ( Davis and Hamscher, 1990, Davis, 1984)
Diagnostics answer questions such as: "why this apartment complex is unsafe" or "why this office layout makes people come face to face so rarely".
A second type of reasoning processes information the opposite direction. From form it predicts operation, and from operation performance. In design practice predictions are used in the evaluation of artifacts. That means, given an artifact's form and operation, to forecast how close the expected performance of the artifact is to the normative one, as specified by the design programme. Or, how an artifact ranks in relation to that of an other artifact. in respect to an exptected performance. Evaluation answers questions such as: "Is this sequence of rooms making people wasting time going from task to task?" or "Which of these different locations of the entrance gives to users more information about events in the complex?"
Design generation, morphogenesis starts with performance and terminates with form. It searches to identify the form that the artifact has to take to bring about one or more normative performances. It uses the facts of design explanations and predictions about the way constraints of form- operation-performance are interconnected and inserts them into a deontic reasoning framework. It starts with prescriptions of performance and terminates with prescriptions form. It searches to identify the form that artifacts have to take to bring about one or more performances that have to come about. Morphogenetis answers questions like: "If a building has to be highly safe (performance-norm), what pattern of circulation of people has to occur (operation), and if this circulation pattern has to take place what the configuration corridor has to be have (form)?"
Context enters in design reasoning by attaching conditions within which the principles or rules about relations between form-operation-performance apply. Context aspects usually lead to plausible explanations, uncertain predictions and probable architectural diagnostics. (A state of affairs not different from medical diagnostics and electrical engineering interpretations of bad circuit behaviour). Here is how context applies: "If a corridor has a specific shape X then people can safely evacuate a building unless the lighting conditions are of type Y."
Contextual aspects that involve plausibility uncertainty, probability and use beliefs, such as the above give design thinking richness and realism. They can also be complicated to describe and very tricky to program. But this does not mean architectural thinking is not representable computationally or that its complexity is intractable by definition.
Frames and facts
When architects use knowledge to describe, explain, evaluate, generate building schemes they do not use isolated facts, about form, operation, performance, context. They mobilize highly organized information, which generalizes classes of situations or combines data into interdependent complexes.
In order to represent this very significant characteric of design thinking we will introduce frames, one of the classic ways that Artificial Intelligence (Minsky, 1975) uses to represent knowledge. Frames are a powerful data- structure to capture standard cases, and exploit law-like facts of architectural knowledge. They can embody architectural rules and principles qualitatively employing nominal values. They can carry out symbolic reasoning without excluding the possibility of using non-verbal, "sub- symbolic", numerical computations. Visually the idea of a frame can be displayed through a graph whose nodes and links, stand for objects and relations between them. These nodes and links make up a kernel of design thinking which gets hold of constants, facts. Nodes and links spread out of this kernel to account for particular fact- instances through slots, terminals which can receive specific values and keep track of differences and
changes.
The design frame, graphically expressed, looks like that:

Our first encounter with frames was almost two decades ago to analyze architectural discourses, the texts people produce to communicate verbally about architecture. (Tzonis, Berwick, et.al. 1975.) At that time, we used frames to match text parts with frame elements. The elements of the frame and their relations represented prototypical structures of architectural thinking.
Here is how the above information in the diagram looks translated into LISP, -an A.I. language oriented towards handling symbols rather than numbers only -representation:
To express more complex information about design, frames can be attached to other frames on the same level or in subordination (sub-frames). Such nested structures can picture intricate associations between sets of design facts, whole-part form relations, hierarchies of operation or performance.
Design frames and the facts they embody about form, operation, performance can be processed through productions, condition-action rules These represent states of information in design thinking. Rules are made
out of two components: an IF, or condition part and a THEN, or action part, as the above illustration shows. Rules are chained together in lines, which map possible sequences of reasoning. But we can also think of them as chained into multidimensional contraptions capable of holding intricate thinking textures.
Our brief discussion of frame methodology suggests that we can do three things:
1. Organize design facts in a way that reflects essentials of domain knowledge.
2. Condense such knowledge into structures which allow us to focus explicitly on facts needed and when needed according to the state of evolution of a design problem.
3.Reason computationally through highly structured knowledge and as a result design intelligently.
A strictly rule based design inference system however may still easily end up into a highly inefficient methodology. Imagine hundreds of rules in a
maze of interactions. They may easily defeat the very idea of computational economy in bringing knowledge in the automated design process. We have to be, on one hand, aware of this issue as we try to assemble a design reasoning system, on the other, seek ways of avoiding a computationally undesirable outcome. Since architecture is a late comer in the use of knowledge based systems it should not repeat the same errors of other domains where such systems were applied. How can we safeguard the system from running out of time and memory?
Some of these questions will become evident in the following part where the form-operation-performance frame, an abstract formula, will be applied to a specific design case.
Representation of design explanation
Let us see how we can turn this very abstract (1) form-(2) operation-(3) performance frame of architectural representation into a realistic description. We will use once more the example of Le Corbusier's Unite d'Habitation (U.H.) (Offermans and Cohen, 1989). Not to clutter our discussion we will try to be extremely concise. We will start with a verbal description of the building:
What is immediately distinguishable about the (1) form of the U.H. it is its three part division into:
(1.1) "pilotf' base, the way the building rises on stilts. Equally unique is: (1.2) its slab body structure, inside which modular units of individual apartments fit. Intriguing also is its:
(1.3) "deck" roof, the way the building comes up and terminates onto a public terrace.
From the point of view of (2) operation U.H. has singular characteristics too: (2.1) Air can circulate without obstruction under the building. It is also special how:
(2.2) the overall bearing structural framework of the building is independent from the structure enclosing each apartment. (2.3) People can look around from high up.
Considering (3) the performance of U.H., the building is unique in terms of: (3.1) it is environmentally good.
(3.2)lt provides a flexible choice of life styles, and also offers (3.3) enjoyable views of its surrounding.
Let us now describe verbally the way performance -operation- form in U.H. are associated with each other:
The building is: (3.1) environmentally good because (2.1) air can circulate under the building thanks to (1.1) the "piloti".
U.H. offers: (3.2) large choice of life styles due to (2.2) the way bearing and enclosing structures of the building are kept independent of each other, thanks to its (1.2) rack slab body.
The (3.3) views of U.H. are enjoyable because (2.3) people can look from high up thanks to (1.3) its "piazza" roof.
What I did was to split the nodes of performance, operation, form into constituent subnodes. I now recombine these subnodes, in a new subframe where each performance-part corresponds to its operation-part and form-
part.

Here is, graphically expressed, how the original design object general frame is applied, instantiated, into the case of the Unite d'Habitation:

and how it looks recombined:

Representation of the morphogenesis
After design explanation, follows, in the same abbreviated way, a discussion on design or morphogenesis. Design generation looks like the mirror image of design explanation. Here we start from normative performance, move to operation and through that arrive to the form the building.
We will try to reconstruct Le Corbusier's groping with the conception of the Unite d'habitation. It appears to have taken the following design reasoning path.
Assuming the performance description of the project Le Corbusier searches for precedents. He asks: "Do I know of any products which": 1. do not disrupt the natural continuity of the terrain? (slot 1)
2. have independent bearing framework and subdivision structures? (slot 2) 3. have public spaces with commanding horizon vistas? (slot 3)
L.C.searches in his memory to find if there are any existing artifacts which respond to the above question. This search involves checking for descriptions of products, which have identical or similar operation slots with those of the project to be designed.
As it is well known L.C. had a very endowed memory. He observed and noted down in numerous notebooks but also in his memory a very large number of artifacts. He possessed also in his mind a true library of objects, a Thesaurus of Precedent Plans, properly tagged and sorted out. Indeed he comes out of his memory library triumphant carrying three precedent products:
1. The "peasant hut", which matches operation slot (1): The hut does not disrupt terrain continuity.
2. The "winebottle-rack", which matches operation slot (2): The bottle-rack has bearing framework which is relatively independent from the shell of the bottles.
3. The "ocean liner", which matches operation slot (3): The ocean liner has deck with commanding vistas.
Next L.C. identifies which form aspect in the precedent artifacts relates to the specific operation. He isolates and extracts out of the precedents the specific part of their spatial form that supports an operation corresponding to the operation demanded:
As we can see from the above diagram in each of these form descriptions the precedent part has two characteristic slots. One specifies the form part itself. the other the syntactic relation of the form partto the total form of the precedent. The "syntactic slot" restricts the possible connection of a part of an artifact to the rest of its parts. It constrains the way parts from precedents may be put together to form a new design product. It helps designing a new whole through recomposing precedents. Here is the result:

Design knowledge -design theory
Certainly, this is a very rough preview of how design thinking can be represented computationally.
We showed how form-operation-performance descriptions are implemented in a knowledge system, here frames, which capture essentials of architectural "intuitive" and experienced thinking. We have also showed how old solutions relate by analogy to new problems.
In the analysis of the case we have represented creative design by analogy using a Winston (1984) style reasoning by analogy. Other ways of deriving new design plans analogically have been tried in domains outside of architecture ( Mostow, 1990). Winston's style is the one we found most close to representing architectural thinking.
Here are the key points of creative design process as we talked about it up to now:
Presuppositions and input:
1. Presupposed principles linking form-operation-performance. (a causal theory of design)
2. Presupposed syntactic principles about form specifying how design objects can be decomposed and recomposed into parts. (a topological theory of design)
3. Architectural program: Performance norms of planned project. 4. Precedent projects. ( Design Thesaurus).
Process:
5. Identification of operations in planned project.
6. Search for precedents descriptions whose operation attributes correspond to those of the planned project. Matching.
7. Extraction out of the precedent of matched component.
8. Integration of the extracted components into the new product.
Examining this outline raises immediately many questions:
Where can we find these rules linking form-operation-performance or decomposing and recomposing the form of artifacts syntactically? We obviously need a theory for that and principles to activate this theory into design acts.
-How can we parse ordinary architectural programs and extract out of them explicit performance norms? How can we identify intentions, obligations, goals in documents which are basically a flow of words?
-How can we develop a Thesaurus of precedent solutions? How are such precedents represented in such a Thesaurus without creating memory or search problems? And certainly, how are they selected?
-How can we reason through images simulating potential operations occurring in a given forma- spatial organisation of precedents?
-How can we extract emergent image-parts out of precedent images and put them together into new wholes as the case of Unite d'Habitation indicated? Further on what if the pieces don't fit together so neatly as our example showed?
Once more it is clear that we cannot make a step towards answering these questions without a robust design theory.
We have to confess that the U.H. story we have chosen, appears to be leading smoothly from principles over the bridge of design by analogy and
precedents to the happy end of a creative composition without many adventures. This idealisation might raise some scepticism about the degree our suggested representation of creative design is generalizable under less favourable conditions. One could easily think of other design plots which lead to more difficulties. Here are two typical cases:
1. Precedent conflict A performance norm might propose a candidate precedent component to be integrated in the new project whose formal characteristics conflict with those of an other performance's candidate precedent component. Here is a crude topological example of a case of conflict:
2. Precedent adaptation: Problems arise if the precedent's form does not conform with some constraints of the new problem. Suppose for example the U.H. had to be erected in an area with a zoning regulation that only thatched roofs were permitted. Then the precedent had to be adapted to the new situation by modifying its form characteristics without worsening significantly its performance.
New solutions naturally lead to new problems. Yet there is no reason to suspect that they demand a paradigm of design different from the analogical preceden- based one we just proposed. They do suggest however the indispensability of design computational theory to bring an automated method of design endowed with some intelligence. ( "Computational" theory is used here in the sense of Marr's (1982) computational theory definition.) It is evident by now that to be able to put together a computer based design system with a "Le Corbusier-like" creative behaviour it is necessary to have in hand knowledge about form-operation-performance relations, as well as how to carry out the kinds of "Le Corbusier-like" inferences encountered in the Unite d'Habitation case. A mere collection of explicit rules of thumb for associating architectural form with operation and performance extracted from specialists, wether casually or even through the use of protocols, stitched together in a so called "expert system", is far from being the kind of domain knowledge we need to be able to put together a computer based intelligent design system.
The rules one extracts from piecemeal ad hoc observations tend to be too "shallow". They have a very low capacity of generalisation. A consequence of their "shallowness" is that they lead to a system which depends on too many design rules and too many interactions to elicit solutions. In the absence of "deeper", more universal principles they takes us to new untractable problems. (Berwick, 1990) Thus, the few current efforts so far to create architecture expert systems have come up to the familiar dilemma, familiar from the period of Space Allocation analytical techniques, to have to choose between effectiveness and efficiency of the system. Without doubt there is a wave of trials to introduce computer knowledge based systems to design the last five years ( Gero, 1990, Gero, Maher, Zhang 1988, Coyne et al. 1990) yet there is a conspicuous absence of deeper computational analysis of the problems at hand as well as a cognitive analysis of the knowledge involved. This suggests that we have not yet drawn the proper conclusions from the past. We still do not perceive that the questions of the "economy" so to speak of a computer-based intelligent design system are fundamental and cannot be postponed to the very end of the development of the system. It is from this perspective, the perspective of limits of computability, of resources, such as time and memory, of "minimal rationality" ( Cherniak, 1986) that design knowledge should be discussed in reference to the system we want to put together.
For domain knowledge to play the key role we sketched about in the beginning of our paper, following the critique of Space Allocation analytical techniques, it has to be assigned the task to cut down search. To do so design domain knowledge has to be based on solid computational design theory capturing both specifics of architecture, sufficiently generalized canonical conventions constraining architectural form, constraints of form on operation etc. as well as intrinsic, sufficiently "deep", cognitive constraints. It is out of such computational theory of design that the rich variety and intricacy of apparent rules -which empirical studies have
frequently reported in design practice- can be described, explained, predicted and planned. And it is through such constraints that design
domain knowledge can help overcome inimical problems of complexity and intractability so typical to architecture.
If the methodology of symbolic, frame-based representation is the ultimate weapon to conquer the problem or other, 'deep reasoning' or sub- symbolic tools have to be brought in it is of less importance. One way or another one has to be prepared to face the straights of Scylla and Charybdis of design, and knowledge based stratagems are the only way through.
The Unite d'Habitation case has helped us see that contemporary computer systems can undertake truly intelligent tasks which involve complex qualitative design reasoning at the level of project preparametric conception rather than long but plain quantitative calculations at the level of parametric design and machine drafting. Parametric design, numerical calculations, techniques such as Space Allocation, machine drafting, C.A.D. will always be in demand. After all the primary task for machines is to free people from non intelligent work. But in order to efficiently and effectively
put into action such low intelligence tools they have to be integrated inside intelligent cognitive frameworks. Increasingly, the importance of design decisions on this conceptual level is recognized and the lack of rigorous methodology focussed on such decisions is acknowledged. The problem is not one of 'upgrading' quantitative techniques. It is exactly the reverse. To develop machine-based methods of high intelligence that can be extended
to practice tedious assignments, such as calculating dimensions of details to the millimetre, specifying materials, estimating costs to the penny and visualising perspectives with hundreds of shades and tones.
Given the importance design knowledge plays in the development of intelligent computer based design systems we are oblidged to investigate very seriously cultural constructs, architectural theory, as well as natural products, the human design cognitive apparatus. In doing so mechanisation achieves what people usually tend to ignore that machines can do: it offers a better understanding of human nature, thus making humans more human.
Notes
1 These techniques became known under many names: "Facility layout" (Block and lewis, 1975, Francis and White, 1974), "Automated Spatial Synthesis" (Eastman, 1975), and
"Space Planning" (Gero, 1977, Miller, 1911). We will use the term "Space Allocation" as a more general one. (For an overview see Grant, 1983 and Fourcade, 1975).
2 In fact S.A. techniques to a great extent were reinventing computerised plant layout techniques developed by industry engineers which anteceded the Chermayeff and Alexander program (Koopmans and Beckman, 1957) unnoticed before by architects.
3 (Weinszapfel, Johnson, Perkins, 1971, Eastman, 1971).
1. Relational rules, expressing interactions between spaces.
2. locational rules, specifying orientation and zoning, restrictions.
3. Configurational rules, controlling alignment, visual access and proportions of spaces. An other strategy followed, one I was very much part of, was the introduction of multiple criteria into the process to improve the effectiveness of the product.
That was very much the focus of my collaboration with Serge Chermayeff at Yale trying to develop a hierarchical, multi-objective design model (Chermayeff and Tzonis, 1967). But it was only in 1969 that we started developing a true multicriteria design technique with the collaboration of Oradiah Salama introducing the ElECTRE method by METRA (Benayonn,
R., Roy, B. and Sussman, B., May and June 1966, Tzonis and Salama, 1972, 1974 and 1975).
4 Heuristic techniques were introduced as well as more powerful al90rithms to cut down vast search problems. Various decision rules were devised to define the order in which operators of allocation were placed, that is the order in which locations were placed on the available site like molecules of space. Interesting procedure strategies were developed by Eastman (1971), Pfeffercorn (1975), Mitchell and Dillon (1972). In general these strategies fell under the following categories:
1. Analytical solution, linear, quadratic and integer programming.
2. Exhaustive enumeration, ALDEP program (1965).
3. Hill climbing, CRAFT (1964), GSP (1970) and IMAGE (1970) programs. 4. Constructive methods, CORELAP (1967) and MAGIC (1975) programs.
In addition even more radical reforms were tried in order to improve the effectiveness of Space Allocation techniques. They took Space Allocation process not as a tool to generate definitive plans through optimization (Simmons, 1969, Armour, Buffa, 1963), or satisfaction of a set of constraints (Eastman, 1971, Grason, 1972) but as an aid for the architect limiting their usefulness to:
1. Improving an existing plan (Armour, Buffa, 1963, Weinzapfel, (Johnson, Perkins, 1971),
2. Aiding the designer who was ultimately controlling the design process, Space Allocation plans playing the role of positive intervener only (Negroponte, 1969),
3. Optimising a plan only from one point of view and letting the architect take care, by intuition and traditional methods of the rest (Bernholz).
4. Testing only solutions provided by architects intuitively. (Harper, 1968).
5 Our AliA group in Delft has been investigating such general principals laws and intrinsic constraints. More specifically interactions between form-operation-performance have been studied on the level of topological organization of form. These relations have been identified as "deep thinking' systems, systems working through rigorous numerical mathematical methods embodying the wary experts read and interpret rough drawings, such as the ones we showed Le Corbusier used, combining the same time symbolicity represented domain principles (Zandi Nia, 1992, Koppelaar and Zandi Nia, 1992, Tzonis, 1987). In an other study architectural formal-spatial constraints have been researched towards a system of visual recognition of architectural precedents (Koutamanis, 1990). Once more the system
employed explicitly represented general principles (Tzonis, 1986) which architects use implicitly in identifying and discriminating architectural plans.
References
Alexander, C. (1964);
Notes on the Synthesis of Form, Harvard University Press. Argan, G.C. (1965);
'On the Typology of Architecture'. Architectural Design, December. Armour, G.C. and Buffa, E. (1963);
'A Heuristic Algorithm and Simulations Approach to the Relative Location of Facilities'. Management Science, Vol. 9, no. 2, January, pp. 294-309.
Arrow, K. (1951);
Social Choice and Individual Values.
Benayonn,R., Roy, B., and Sussman, B., (1966);
'Manuael de reference du programme ELECTRE'. Note de
Synthese et Formation, no. 25 de la Direction Scientifique de la SEMA, May.
Bernholz, A. (1969);
'LOKAT'. Laboratory for Computer Graphics and Spatial Analysis IV. 4 G.S.D., Harvard University, Cambridge, March. Berwick, R.C., Fong, S. (1990);
'Pr,inciple-Base Parsing: Natural Language Processing for the 1990s'. In: Winston, P., and Shellard, S., Artificial Intelligence at MIT.
Block, T.E. and Lewis, W.P. (1975);
A Review of Plant Layout: an Analytical Approach. Carnap, R. (1947);
Meaning and Necessity, Chicago.
Chermayeff, S. and Alexander, C. (1963); Community and Privacy.
Chermayeff, S. and Tzonis, A. (1967);
Advanced Studies in Urban Models. Chermayeff, S. and Tzonis, A. (1970); Shape and Community.
Chermayeff, S. and Tzonis, A. (1970); Towards an Urban Model. Cherniac, C. (1986);
Minimal Rationality. Cohen, M. (1989);
Master thesis, Faculty of Architecture, Delft University of Technology.
Coyne, R.D., Rosenman, M.A., Radford, A.D., Balachandran, M., Gero,
J.S. (1990);
Knowledge-Based Design Systems. Condorcet, Marquis de;
'Memoire inedite sur les hospitaux' (1786). Presentation par Alexander Tzonis, Dix-huitieme Sciecle. 'Le sain et Ie malsain', no. 9, Paris, 1977, pp. 109-115.
Cross, N. (1989);
Engineering Design Methods, Chichester. Eastman, C.M. (1971);
'G.S.P. a System for Computer Assisted Space Planning', 8th Design Automation Workshop, Atlantic City.
Eastman, C.M. (1975);
Spacial Synthesis in Computer Aided Building Design. Randall, D. and Hamscher, W.C. (1990);
'Model-based reasoning: Troubleshooting'. In: Winston, P. and Shellard, S.A., Artificial Intelligence at MIT.
Davis, R. (1984);
'Diagnostic Reasoning Based on Structure and Behavior'. Artificial Intelligence, Vol. 24, no. 3, pp. 347-410.
Fourcade, A.M. (1975;,
Architecture and Automized Methods. (Thesis) M.I.T., Department of Architecture. Foz, M.A. (1972);
Some Observations on Designer Behavior in the Parti. Unpublished masters thesis, M.I.T., Department of Architecture and Urban Studies and Planning.
Francis, R.L., and White, J.A. (1974);
Facility Layout and Location: an Analytical Approach.
Gero, J.S. (1990);
'Design Prototypes: a knowledge representation schema for design'. AI Magazine, pp. 26-36.
Gero, J.S., Maher, M.L. and Zhang, W. (1988);
'Chuncking structural design knowledge as prototypes'. In: Gero, J.S. (ed.), AI in Engineering Design.
Gero, J.S. (ed.) (1977);
Computer Applications in Architecture. Grant, D.P. (1983);
'Space Planning Methods, Part One and Two'. DMG-DRS Journal, vol. 17, no. 1, pp. 4-36 and no. 2, pp. 57-98.
Grason, J. (1971);
'An Approach to Computerized Space planning Using Graph Theory'. Proceedings of the 8th Annual Design Automation
Workshop.
Guilbaud, G.Th. (1952);
Les theories de I'interet general. Economique Applique. (tr. in Lazersfeld, Readings in Mathematical Social Science, 1966).
Koppelaar, H. and Zandi-Nia, A. (forthcoming);
'Solving a class of architectural design problems by a neural network'. International Journal of Intelligent Systems.
Harper, G.H. (1968);
'B.O.P.: An Approach to Building Optimised'. Proceedings Assoclation of Computer Machinery, 23rd National Conference, november, pp. 75-83.
Koopmans, T.C. and Beckman, M. (1957);
'Assigment Problems and the Location of Economic Activities'. Econometrica, vol. 25, no. 1.
Koutamanis, A. (1990);
Development of a computerized handbook for architectural plans. Doctoral dissertation, Delft University of Technology.
Marr, D. (1982); Vision.
Mc.Carthy, J. (1980);
'Circumscription'. Artificial Intelligence, vol. 13, nos. 1,2. Minsky, M. (1975);
A framework for representing knowledge. In: Winston,P. (1975) Op. Cit.
Minsky, M. (1988);
The Society of Mind.
Mitchel, W.J. and Dillon, R.L. (1972);
'A Polyomino Assembly Procedure for Architectural Floor Planning'. In: W.J. Mitchel (ed.), Environmental Design, Los Angeles. Published also in Eastman, C.M. (1975).
Mostow, J. (1990);
'Design by Derivational Analogy'. Carbonell, J. (ed.), Machine Learning.
Negroponte, N. (1969);
'The Architecture Machine'. Architectural Design, september, pp.
39-51.
Newell, A. (1980);
'Physical Symbol Systems'. Cognitive Science, vol. 4.
Offermans, E. (1988);
Master Thesis, Faculty of Architecture, Delft University of Technology.
Pfeffercorn, C. (1975);
'A Heuristic Problem Solving Design System for Equipment of Furniture Layout'. CACM, vol. 18, no. 5.
Protzen, J.P. (1974);
'Editorial'. DMG-DRS Journal, vol. 8, no. 3. Simmons, D.M. (1969);
'One Dimensional Space Allocation: an Ordering Algorithm' Simon, H. (1969);
The Sciences of the Artificial. Stiny, G. (1980);
'Introduction to Shape and Shape Grammaers'. Environment and Planning B, vol. 7, pp.343-51.
Tzonis, A. and Lefaivre, L. (1986);
Classical Architecture, Cambridge.
Tzonis, A., Freeman, M. and Berwick R. (1978);
'Discourse Analysis and the Logic of Design'. Harvard G.S.D. Publication Series in Architecture, A-7817.
Tzonis, A., Freeman, M., Lefaivre, L., Salama, 0., De Cointet, E. and Berwick R. (1975);
Les Systemes Conceptuels de I' Architecture. Tzonis, A.and Salama, O. (1972);
'Selected Exercises in Programmatic Analysis'. Harvard G.S.D. Publication Series in Architecture.
Tzonis, A.and Salama, O. (1974);
'Problems of judgement in Programmetic Analysis in Architecture; the Synthesis of Partial Evaluation'. DMG-DRS Journal, july- september. Also in Journal of Architectural Research, vol. 4, no. 2, 1975.
Tzonis, A. (1990);
Hermes. Cambridge. Ulrich, K.T. (1988);
'Computation and Pre-Parametric Design'. MIT Artificial Intelligence Laboratory, Technical Report 1043.
Von Mieses, R. (1951);
Positivism. New York.
Weinszapfel, G., Johnson, T.E. and Perkins, G. (1971);
'Image'. 8th Design Automation Workshop. Published also in Eastman, C.M. (1975).
Winston, P.H. (1975);
The Psychology of Computer Vision. Winston, P.H. (1978);
'Learning by Creating and Justifying Transfer frames'. Artificial Intelligence, vol. 10, no. 2, 1978.
Winston, P.H. (1979);
'Learning by Understanding Analogies'. AIM-520. Winston, P.H. (1980);
'Learning and Reasoning by Analogy'. Communications of the Association for Computing Machinery, vol. 23, no.12.
Winston, P.H. (1982);
'Learning new Principles from Precedents and Exercises'. Artificial Intelligence, vol. 19, no. 3.
Zandi-Nia, A. (1992);
Topogene: an Intelligent System for generating Topological Patterns of Buildings in Respect to Social Norms. Doctoral dissertation, Delft University of Technology.