Adaptation of Representation in GP

This paper discusses our initial work on automatically adapting Genetic Programming (GP) representation. We present here two independent techniques: AMS and ACE. Both techniques are based on Constrained GP (CGP), which uses mutation set methodology to prune the representation space according to some context-specific constraints. The ASM technique monitors the performance of local context heuristics when used in mutation/crossover, during GP evolution, and dynamically modifies the heuristics. The ACE technique iterates complete CGP runs and then uses the distribution information from the best solutions to adjust the heuristics for the next iteration. As the results indicate, GP is able to gain substantial performance improvements as well as learn qualitative heuristics.