Abstraction and Decomposition in Hillclimbing Design Optimization

ion and Decomposition in Hillclimbing Design Optimization Thomas Ellman Mark Schwabacher Department of Computer Science Hill Center for Mathematical Sciences Rutgers University New Brunswick, New Jersey 08903 (908) 932 4184 fellman,[email protected] CAP-TR-14 Abstract The performance of hillclimbing design optimization can be improved by abstraction and decomposition of the design space. Methods for automatically nding and exploiting such abstractions and decompositions are presented in this paper. A technique called \Operator Importance Analysis" nds useful abstractions. It does so by determining which of a given set of operators are the most important for a given class of design problems. Hillclimbing search runs faster when performed using this this smaller set of operators. A technique called \Operator Interaction Analysis" nds useful decompositions. It does so by measuring the pairwise interaction between operators. It uses such measurements to form an ordered partition of the operator set. This partition can then be used in a \hierarchic" hillclimbing algorithm which runs faster than ordinary hillclimbing with an unstructured operator set. We have implemented both techniques and tested them in the domain of racing yacht hull design. Our experimental results show that these two methods can produce substantial speedups with little or no loss in quality of the resulting designs. Content Area: Automated Reasoning, Search, Reasoning about physical systems, Knowledge Acquisition, Machine Learning.