ROOSI - Version One of a Generative Expert System for the Design of Building Layouts

Resume. ROOS1 est Ia premiere version dun systeme expert generateur de plans de batiments, pouvant s'appliquer a differences sortes de problemes. Le systeme no reproduit pas la clemarche des designers, mais plutot cherche a completer leur action grace a ses moyens de recherche systematique d'alternatives et a sa capacite a prendre en compte un nombre important de parametres de design. Le developpement du systeme vise d fournir on apercu de I'applicabilite des techniques de l'Intelligence Artificielle a l'amenagement de l'espace et a la conception de batiments en general. Le systeme est base sur one strategic de generation puis test. Ses composantes principales sont on generateur d'alternatives, on module de test et une structure de controle (qui pourra titre amelioree afin de devenir on veritable module de planification). Les elements clue le generateur juxtapose sont des rectangles. Les relations spatiales, au-dessus, au-dessous, a-gauche, a-droite, s'appliquent aux paires d'elements intervenant dans le plan et sort les variables de base permettant de difterencier les solutions et de produire les differentes alternatives. Le generateur agit independamment de toute contexte , et engendre toutes les facons possibles de disposer on nombre donne d'elements. Le module de test, au contraire, est lie au type de probleme rencontre et contient des informations relatives a la forme de plan demandee. Le systeme peut s'appliquer a plusieurs domaines en ulilisant le module de test adequat et eventuellement une structure de controle specifique. La structure de controle guide la recherche de solutions, en activant alternativement le g'en'erateur et le module de test. Plusieurs versions du systeme, de complexite croissante, vont titre developpees. Chacune sera concue selon one architecture claire et precise, et noun _ comptons evaluer chaque architecture en termes do possibilites potentielles et limitations relativement a divers domaines. Cet article decrit la premiere version du systeme.

ROOS1 is the first version of a generative expert system for the design of building layouts that can be adapted to various problem domains. The system does not reproduce the behavior of human designers; rather, it intends to complement their performance through (a) its ability to .systematically search for alternative solutions with promising trade-offs; and (b) its ability to take it broad range of design concerns into account. Work on the system also aims at providing insights into the applicability of Artificial Intelligence techniques to space planning and building design in general. The system is based on a general generate-and-test paradigm. Its main components are a generator, a tester and a control strategy (which is to be expanded later into a genuine planner). The generator is restricted to the allocation of rectangles. The spatial relations above below, to the left and to the right are defined for pairs of objects in a layout and serve as basic design variables which define differences between solutions and govern the enumeration of alternatives. Within the class of layouts it is able to produce, the generator is completely general and able to generate all realizable sets of spatial relations for a given number of objects. In contrast, the tester is domain-specific and incorporates knowledge about the quality of layouts in a specific domain. The system can be applied to various domains by running it With the appropriate tester and, possibly, the appropriate control strategy. The control strategy itself mediates between planner and tester and, when expanded into a planner, is able to streamline the search for alternatives. The system will go through a sequence of versions with increasing complexity. Each version will have a conceptually clean and clear architecture, and it is our intention to evaluate each architecture explicitly in terms of its promises and limitations with respect to various domains. The first of these versions is described in the present paper.