RE-THINKING OPTIMIZATION AS A COMPUTATIONAL DESIGN TOOL: A SITUATED AGENT APPROACH SOMWRITA SARKAR, JOHN S. GERO AND ROB SAUNDERS Key Centre of Design Computing and Cognition University of Sydney Email address:{somwrita, john, rob} @ arch.usyd.edu.au Abstract. This paper presents a situated agent-based tool for design optimization. A situated agent captures, learns from and re-uses the interactions which it has with its external environment, forming the basis for experience-based knowledge building in an agent. An agent is developed for design modeling, reformulation and algorithm selection – a class of tasks in design optimization traditionally performed by humans based on their experience, and hard to automate. 1. Motivation Optimization has been a predominant approach supporting automated design in architecture and engineering. Space layout problems, for example, have been automated using a number of optimization approaches (Liggett, 1985; Jo and Gero, 1995; Gero and Kazakov, 1997; Michalek and Papalambros, 2002). Optimization models designing as search. It finds the “best” design for some expected performance from a well structured, fixed solution space (Radford and Gero, 1988; Parmee, 1998; Papalambros and Wilde, 2000). For most automated algorithms, the structure of the model remains unchanged throughout search, as does the behavior of the search process. Activities like design modeling, reformulation and algorithm selection, which involve interactive exploration and a dialogue with the problem representation, remain primarily human endeavors. In general, tools fail to support or assist the designer during conceptual design which is when modeling and reformulation are the most pronounced activities. Designers can often produce better designs using heuristics learnt through personal experience. They transfer design knowledge from past experiences into current ones, and treat modeling, reformulation and search as mutually interacting concurrent activities. Suwa et al. (1998) observed from studies conducted on designers that they use sketches not just as external representations, but also as a mode of interaction with the developing design leading to unexpected discoveries and inventions of new design issues. This dynamic, interaction based experiential learning enables a way of designing that Schon and Wiggins (1992) refer to as the “interaction of making and seeing”. This is in contrast to the behavior of static, non-interactive optimization tools.