Valeria Bottelli, Christian Fogh

Politecnico di Milano, Facoltà di Architettura, Dipartimento di Programmazione Progettazione e Produzione Edilizia, Via Bonardi 3, 20133 Milano, Italy
tel. -39-2-23995125 fax. -39-2-23995150 email: arlati@cdc8g5.cdc.polimi.it

Abstract
The paper presents an on-going Ph.D. research aimed at the investigation of the nature of process knowledge in architectural design and the development of a case-based navigation tool for its support.
We assume design as an incrementally defined intentional planning activity, leading from an incompletely specified state, comprising the space of acceptable solutions, toward a final configuration, comprising adopted decisions. We consider the essence of design as the continual recursive transformation of the initial solution model, in order to map the desired state onto the enacted one.
On the basis of this concept of design we describe the model of Galathea, a case-based tool, aimed at progressively representing the enlarging space of acquired knowledge, and at supporting the designer's central role in the management of the design process.
Galathea is a navigator aimed at guiding designers in the description, mapping and management of the complete decision-making path of projects by means of the dynamical representation of the relationship between goals, constraints and the decisions/actions adopted at specific nodes. These abilities also develop as a case-based instrument aimed at enabling the storage, retirieval and adaptation of relevant micro-processes in similar project contexts.

Key words: knowledge-based design; case-based reasoning; design process control, planning

Introduction

We report on an on-going Ph.D. research investigating the nature of process knowledge in architectural design and developing a case-based tool for its support.
Today, technical, planning, management and environmental issues have created a scenario of such complexity that there is a demand for new, integrated instruments to handle the different kinds of information necessary to handle the decision-making process underlying architectural design with the requested qualitative result.
In order to acquire the adequate cultural insight into this complex problem, we started out on this research project investigating the architectural design process under two points of view. On one side the historical evolution of the nature of decisions and consequently the development of manuals and treaties utilised throughout the centuries for their management; and on the other side the history of cognitive models aimed at the formalisation of the design process.
We thus followed the development of the function of architectural technical manuals from the outset, through the 18th and 19th centuries, which marked the moment of their largest fortune, up to their crisis, occurred from the turn of this century onward.
The rationalist era of the 18th century, with its central affirmation of a scientific method, marked the beginning of a "modern" relationship between information, planning and project, with a strongly modified attitude towards the worlds of artistic and technical issues.
The 19th century saw the rise of a positivistic and taxonomic attitude toward architecture, which is expressed in manuals in their progressive transformation into catalogues of exemplary procedures and solutions, a precursor of the concept of cases.
At the turn of the century, the dramatic acceleration of industrialisation and urban development, with the consequent rise of new building techniques, new social needs and organisational issues, progressively led to a separation between the conception and realisation of architecture. This considerably weakened the weight of traditionally accepted building techniques.
With the Modern Movement from the 1920's onwards, a deep change occurred in the relationship with technical issues and the organisation of work, with the search for the utopia of standardisation and industrialisation of architecture. This led to a definitive loss of attention for consolidated building procedures and a growing attention for a typological and abstract conception of architecture.
The last thirty years in Italy, as in the rest of Europe, have been characterised by an exponential development of a number of complex factors influencing the architectural process, such as the proliferation of new building products, the growth in complexity of building regulations and legal issues, the differentiation of professional figures involved in project management etc.
All this has progressive weakened the legitimisation of the architect's central role and has created the demand for new control tools, apt to aid designers in representing and following dynamically the knowledge and expertise-intensive decision-making process underlying architectural design.
On the other hand, we followed the development of formal methods utilised in the past years to analyse process knowledge in architecture. This has meant studying the path leading from the first essentially deterministic conceptions, along the increasingly central role of flexible and recursive models, right down to the present knowledge and memory intensive theories, aimed at reconstructing a unitary idea of architecture from the very first heuristic phases of the process. It has also meant dealing with the area of project-management and decision support systems, following the evolution of alternative models along the progressive weakening of algorithms.
The results of this introductory research has been a valuable foundation for the formulation of the present work.

The architectural design process
In this work, architectural design is viewed as a complex, ill-defined decision-making process, characterised by a number of different related information domains, most of which vague and underspecified in constant change. Therefore, the essence of designers' skill is represented by the ability to continually and recursively modify the decision-making strategy.
In other words, design is a planning activity in which both the rules governing the process and the required properties of the various steps of solution are subject to continual refinement, substitution and review. A decision taken in one domain will reverberate its consequences on the other domains building up the configuration of the process as a whole; therefore any sequential solution approach is methodologically inadequate.
In the problem context we have just outlined, the decision-making activity requires the adoption of an incrementally refined model of reality, apt to progressively incorporate all possibly relevant issues as they arise. That is to say that the decision-making process should abandon all sectorial approaches and be founded upon a method incorporating the highest possible number of issues from the very first heuristic phases of the project, so as to reach a preliminary solution model. The design activity thus becomes the planning of a path between increasingly accomplished configurations of the initial model, which evolves as variables reach higher specification levels.
Moreover, decision-making in design is an information and expertise intensive activity largely relying on the availability of information sources and on the experience and reasoning skills of the decision-maker. In fact, qualified solutions to complex problems are seldom developed from scratch and decision-makers constantly build on knowledge and expertise acquired in the solution of previous similar problem contexts, or cases. Decision making thus requires refinement, modification and intuitive combination of solutions to previous problems, leading to an incrementally configured hybrid solution model.
Therefore, this view of architectural design activity has led us to the formulation of a cognitive proposal for the modelling of the decision-making process based upon the central role of information, memory and experience expressed through the nature and quality of interrelations.

Case-based reasoning
In the introductory remarks, we have defined memory and experience skills of paramount importance in a successful decision-making process in design.
Memory and experience in complex, ill-defined problems rely on the efficient manipulation of relevant knowledge acquired in previous similar situations in order to enhance and ameliorate the task of making decisions in underconstrained and underdefined problem contexts. In fact, the importance of the knowledge and memory base in relation to intelligence is now a wholly accepted concept in AI and cognitive psychology. (Kolodner, 1991)
In such problem contexts, an increasingly central role belongs to those aspects of design referring to the knowledge-base, and in particular to memory. A well organised knowledge and memory base is essential for an effective management of the complex network of interrelations amongst different and non homogenous information domains.
In fact, when tackling a new design problem, it is known that we strongly base our search path upon previous professional experience: we search our memory of precedents for a relevant example (or case), recall the solution process, and thus adapt the previous solution procedure to the new situation.
Case-based reasoning (CBR) is a 'weak' paradigm based upon the modelling of memory and experience. It is characterised as a problem-solving approach inferencing from previous solutions which are adapted to current situations. It has proved useful in domains where experience is strong but domain model is weak, as is generally true of architectural design.
After a first generation of knowledge-based studies utilising CBR concentrated research efforts into trying to replicate human knowledge, current work is primarily directed at a partnership between the human and machine agent. Thus, the main aim of such systems is to "enhance human decision-making by suggesting alternatives, predicting consequences and pulling together the information that goes into decision making", the underlying principle being that of external cognition.
In our view, a case-memory is thus seen as a case-container in which designers themselves manage the nature and semantic connections of cases, according to their personal professional method. This is necessary in a domain like design, in which the entire management of the process, and consequently the search strategies underlying it, do not follow any predetermined scheme and gain effectiveness with the possibility of free exploration.
A case in design KBS literature is generally viewed as a design solution, however well-defined. In our work, primarily focused on the nature of process knowledge in design, a different perspective is utilised, which owes much to a relatively new theoretical approach coupling CBR, planning and general problem solving issues.
This is to say that the matter we have assumed as most relevant in the design process, which is the resource of our case-base, is not the progressive development of project solutions, but the incremental planning and decision-making activity underlying these solutions.
In other words, since we have assumed design as an incrementally defined knowledge-intensive planning and decision-making activity, we have focused precisely on these aspects when building the case organisation model.
CBR in this view allows to retrieve, evaluate and adapt pertinent sub-processes obtained from previous design experiences, rather than tackle each new decision-making problem from scratch. In our view, CBR is useful both in adapting microprocessors stored from past projects, and in helping designers to plan out a general framework for projects, tailored to personal skills and habits.
A case in this work is thus viewed as the process trace of a choice, however complex, adopted within the process. Conceptually, this bears some resemblance to the "foot-printing" model proposed by Carbonell (Carbonell, 1986), although the recursive and progressively defined nature of design implies that the foot-printing approach cannot but be extremely loose and 'weak'. In fact, unlike the problem solving domains studied by Veloso (Veloso, 1994), architectural design, especially in early phases, is not subject to any "strong" planning formalisation method, for there is no consolidated sequential path leading to the final solution, different architects adopting extremely different courses of action.
Retrieval in such a CBR planning model is based upon the evaluation of the current problem configuration compared to similar process or sub-process traces of previous design contexts. The degree of similarity reflects some resemblance to the concept of "derivational analogy" proposed by Cabonell, and consists in the comparison of new and previous sub-process traces of solved problems, where the relevant characteristics which allow comparison are the mapping of goals and constraints and the general problem configuration.
Thus, sub-processes, i.e. decision traces, may be stored, retrieved and adapted in new project contexts as cases, on the basis of process analogy. In this view, CBR allows the application of incrementally detailed relevant sub-processes from previous projects, and helps designers to clarify the terms of a project and to plan out a framework of reference tailored to personal skills and habits.

Galathea: a model for a case-based knowledge navigator for architectural design
On the basis of the view we have just outlined of the nature of knowledge in the architectural design process and of case-based reasoning, we here propose a model of the design process and of a tool for its support.
We consider design as a form of intentional planning (Cohen, Levesque, 1990), in which the architect is viewed as an intention-driven agent, exhibiting the following basic set of cognitive abilities:
_ the ability to analyse at any stage of the process, the current world situation in terms of a certain number of characteristics (goals, constraints and their relationships). This ability reflects the subjective expertise-intensive structure of knowledge pertaining to each specific architect;
_ the ability to define a more satisfactory project step configuration, by means of the evaluation of the causal role of the characteristics, and therefore the nature of their modification, in order to bring about the desired changes in the overall situation;
_ the ability to deliberate about the actions which are likely to transform the current project situation into a more satisfactory one, by means of the evaluation of alternative courses of action.
Given these abilities, the decision process underlying design involves the ability to transform a given project situation, by modifying one or more of its defining characteristics in the direction of a higher level of satisfaction.
This is to say that, given a project context, to plan means to select, among the possible preferable configurations ("chosen" in an intentional model such as that of Cohen and Levesque) the one exhibiting the most satisfactory expected performance, and therefore executing it ("intended", again in an intentional model). In case the adopted course of action does not achieve the given goals, then architects generally return to previous decision nodes, committing to a different solution.
Given the incremental specification of design, a plan in any given moment should be neither more nor less detailed than is needed at the current stage of development. This idea is known in the planning community as least commitment.(Weld, 1994)
A plan can be viewed as a complex network of interdependent nodes of knowledge and goals, large areas of which are initially blank and will be progressively specified as designers proceed in deliberation. Parts of the network will still be underspecified when other domain areas will have already been totally solved and executed. A typical example of this is that structural issues in a building plan are designed, verified and executed long before details referring to windows and doors may not have even been designed.
To fully specify a project means to develop all the nodes down to the desired level of detail. In this process, each piece of knowledge and intentions may reverberate its effects to the others. Starting from scratch, and from an analysis of the initial goals and constraints, the designer's task is to transform an initial situation into a more satisfactory one, by recursively modifying the appropriate characteristics of the progressively evolving situation.
This conception of planning is crucial in our analysis of design and, coupled with our conception of case-based reasoning, dealt with in the previous paragraph, constitutes the cognitive framework for the design process model we propose for Galathea.
Summarising, we represent the architectural design process on three levels of abstraction:
_ a higher level, comprising all potentially acceptable solutions This level regards the mapping of goals and constraints (where goals are viewed as a subjective subset of constraints, i.e. they are not environment-driven but intention-driven constraints). At the outset, a certain number of goals and constraints are specified by the designer. These allow the description of the initial configuration of the model, which is incomplete and underdefined.
_ a lower level comprising the set of decisions/objects actually adopted, which taken together make up the actual solution;
_ an intermediate level comprising the space of all decision nodes making up the path leading from the space of possible solutions to the final executed solution. Decision nodes are viewed as the building blocks of the process. They are elementary units flexible enough to govern the complexity of each single step, at whatever level of abstraction, in the process.
This level is constituted by the decisions/objects needed to combine the elementary units forming the adopted solution (lower level) coherently with the evolving map of goals and constraints (higher level). The executed solution represents the actual intention taken out of all chosen possibilities.
Based upon this overall concept of design, the structure of Galathea, a tool aimed at supporting design activity, is currently under development. Such model may be envisaged as a partner to the architect; the underlying principle being that of external cognition (Winograd & Flores, 1986). This is to say that we see it as a valid knowledge navigating consultant, aiding designers in making decisions quicker and with a broader and clearer knowledge and experience base. Such a tool is conceptually viewed as a part of an integrated, open, flexible environment, devised to adapt to, and grow with, the architect.
Our cognitive model analyses architectural design from a process viewpoint; therefore our tool model is aimed at the description and representation of the architectural design process as a globally coherent evolving activity, and of its constitutive elements. Such constitutive elements include the declaration of evolving goals, constraints and interrelations, directionality (time), progressively restricting space of adoptable decisions/objects, variable weight changing over time, precise and complete mapping of all decision nodes.
Galathea is aimed at the satisfaction of the following three functions:
_ map-making function: that is to say the ability to explicitly describe the evolving project situation, through the declaration of characteristics (goals, constraints and their relationships) defining any specific project step
_ navigation function: that is to say the ability to explicitly represent the path of decisions making up the design process through the establishment of links between decision nodes, enabling the transfer of consequences of any decision/action on the connected mapped nodes
_ case-based function: that is to say the ability to save, retrieve, adapt and use relevant micro-processes so as to enable re-use in similar project contexts.

At the outset of a new project, Galathea should first guide the designer in the mapping out of the context, by means of the explicit declaration of the known constraints and goals, some of which will be context-dependent and some others context-independent, i.e. subjective, as well as of their interactions.
This initial map of the higher level of abstraction (the space of possible, or chosen, decisions) will necessarily be incomplete.
In the course of the design activity, with the growth of relevant knowledge acquisition, the map will progressively evolve, and the nature, weight and origin of constraints will change continually and recursively, as new decisions are taken, starting to configure the lower (or intended) level of abstraction.
During the project, Galathea will moreover report the results of temporary candidate solution evaluation on the basis of the map of constraints, testing the effect of the candidate solution on the whole configuration of the process.
Again, whenever designers feel they need it, it will be possible to refer to previous cases, on the basis of the degree of similarity between the process configuration of the new problem to previously solved ones.
The design support tool may therefore be viewed as a planner, capable to map out the complete process path by means of the dynamical representation of the relationship between goals, constraints and the decisions taken at specific nodes.
We also view it with the aid of the metaphor of a navigator, or map-maker, (hence the name Galathea), since it enables designers to explore their own personal route which, in the space of possible, or chosen, solutions, leads to the temporary adoption, evaluation, adaptation and final execution of the intended one. In this view, the tool is a navigator skilled in the exploration and construction of a route (characterised by its being unique and unrepeatable), rather than a driver skilled at choosing the most appropriate previously defined road.
This model is currently under feasibility analysis. The next steps we are currently planning are the transfer of this conceptual model into a formalised structure, and subsequently the choice of an architecture.
For the moment we conceive an environment comprising Galathea in association with a number of individually chosen, application tools, according to specific needs of an individual/project. (Arlati et al., 1995)
In other words, the model of Galathea has been developed within a general conception of PatriArch, an environment comprising an object-oriented CAD system and a number of individually-chosen external software products devoted to specific technical domains, such as structural verification, lighting and noise simulation, etc.
This of course means that the environment is not seen as an integrated one, but as a "light" solution, in which domain experts are "external" and independent; therefore the only point of contact between Galathea and the agents is the assimilation and treatment of results. This may be done through an extension of the domain of variables pertaining to the object-oriented CAD system, so as to enable communication with external domain experts.
For the moment, we are therefore planning a prototype of Galathea, in which domain expert software products will not be integrated in the system, but the results of their elaboration techniques will be input and used in the navigation tool, in order to test consequences of actions/decisions on the general configuration. The links between situation configurations, desired configurations and actions effected are stored as procedural cases, which may then be indexed, retrieved and adapted in new problem contexts. The organisation of the resulting case-base, i.e. the principles of indexing, retrieval and adaptation have not yet been analysed.

Summary
We have outlined a cognitive approach to architectural design based on four main knowledge areas:
_ a general cultural background defining the complex, ill-defined context presently characterising design activity
_ our approach to design as a recursive, incrementally defined process
_ our approach to design as an intentional planning activity
_ our approach to design as a knowledge intensive activity strongly based on memory and experience, hence the case-based reasoning paradigm.
We have thus proposed a model for design activity based on three levels of abstraction applied to the process.
These three levels of abstraction develop into the model of Galathea, a tool aimed at supporting the design process in its globality and constitutive elements.
This tool may be viewed as a navigator, supporting designers to construct an original decision-making path amongst incrementally defined process configurations, up to the final solution.
This path is constructed with the aid of reference to previous similar project configurations on the basis of analogy between the current and past mapping of evolving goals, constraints and interrelations.

Conclusions and further research
Current work in this research is centred upon further conceptual analysis of the thresholds of abstraction in the design process, so as to enact a smooth and free possibility of exploration of decision nodes from general down to very detailed levels.
The feasibility analysis is also beginning to focus on the simplifications necessary in order to plan a first implementation scheme.

References

Akin, O. (1986), Psychology of Architectural Design. London: Pion

Arlati, E., Bottelli, V., Fogh, C., Tirassa, M., (1995), Modelling process knowledge in architectural design - a case-based approach, Intersymp 95, Baden Baden

Bartsch-Spoerl Brigitte, (1994) Representing and using cases in visualization-oriented design, in: Proceedings of the 2nd European Workshop on Case-Based Reasoning, Chantilly

Bottelli, V., Drago, P., Fogh, C. (1994) Can preconditions of architectural projects be considered as cases? Development of a case-based tool in: Pohl J. (ed.), Advances in knowledge-based building design systems, Windsor: Windsor School of Computer Science

Carbonell, J.G. (1983) Learning by analogy: Formulating and generalising plans from past experience., in: Machine Learning vol. 1.

Carbonell, J.G. (1986) Derivational analogy: A theory of reconstructive problem solving and expertise acquisition. , in: Machine Learning vol. 2.

Cohen, P.R., Levesque, H.J. (1990) Intention is choice with commitment., in: Artificial Intelligence 42: 213-261

Colajanni B., De Grassi M., Di Manzo M., Naticchia B. (1991) Can Planning be a Research Paradigm in Architectural design?, in: Proceedings of the 1st International Conference on Artificial Intelligence in Design, Edinburgh

Kolodner, J. (1993) Case-based Reasoning. San Mateo, CA: Morgan Kaufmann

Pohl, J., Myers L. (1994) A distributed cooperative model for Architectural Design, in: Carrara G. and Kalay Y. (ed.), Knowledge-based Computer-aided Architectural Design, Amsterdam: Elsevier

Veloso, M.M. (1994) Planning and Learning by Analogical Reasoning, Berlin: Springer-Verlag

Weld, D.S. (1994) An introduction to least commitment planning. In: AI Magazine Winter 1994: 27-61.

Winograd, T., Flores, F. (1986) Understanding computers and cognition. Norwood, NJ: Ablex.