An inductive approach to learning and reusing design strategies

Abstract.

We describe a computer system called LearnIT II, which can observe an iterative solution to a parametric design problem and learn the design strategy employed. The program uses identification tree learning to determine which attributes of the design best indicate which design modifications the designer will make. The program re-represents the identification trees as design rules which it can then use to create new designs automatically. Because the rules are learned from the designer, the new designs reflect the designer's engineering judgment, knowledge of implicit constraints, and overall familiarity with the problem. The program has a number of intended uses. As a design automation tool, it provides a convenient means for exploring a design space. As a design documentation tool, it implicitly captures the original designer's insights about a problem and brings them to bear when it is necessary to modify the design. As a tool for preserving institutional memory, it captures design strategies and preserves them for future use.