A divide & conquer strategy for improving efficiency and probability of success in genetic programming

Abstract

A common method for improving a genetic programming search on difficult problems is either multiplying the number of runs or increasing the population size. In this paper we propose a new search strategy which attempts to obtain a higher probability of success with smaller amounts of computational resources. We call this model Divide & Conquer since our algorithm initially partitions the search space in smaller regions that are explored independently of each other. Then, our algorithm collects the most competitive individuals found in each partition and exploits them in order to get a solution. We benchmarked our proposal on three problem domains widely used in the literature. Our results show a significant improvement of the likelihood of success while requiring less computational resources than the standard algorithm. © Springer-Verlag Berlin Heidelberg 2006.