An incremental-approximate-clustering approach for developingdynamic reduced models for design optimization

摘要

This paper we describe a method for improving genetic algorithmbased optimization using reduced models. The main idea is to maintain alarge sample of the points encountered in the course of the optimizationdivided into clusters. Least squares quadratic approximations areperiodically formed of the entire sample as well as the big enoughclusters. These approximations are used as a reduced model to computecheap approximations of the fitness function through a two phaseapproach in which the point is first classified (into potentiallyfeasible, infeasible or unevaluable) and then its fitness is computedaccordingly. We then use the reduced models to speedup the GAoptimization by making the genetic operators such as mutation andcrossover more informed. The proposed approach is particularly suitablefor search spaces with expensive evaluation functions, such as thosethat arise in engineering design. Empirical results in severalengineering design domains demonstrate that the proposed method cansignificantly speed up the GA optimizer