Comparison of semantic-based local search methods for multiobjective genetic programming

We report a series of experiments that use semantic-based local search within a multiobjective genetic programming (GP) framework. We compare various ways of selecting target subtrees for local search as well as different methods for performing that search; we have also made comparison with the random desired operator of Pawlak et al. using statistical hypothesis testing. We find that a standard steady state or generational GP followed by a carefully-designed single-objective GP implementing semantic-based local search produces models that are mode accurate and with statistically smaller (or equal) tree size than those generated by the corresponding baseline GP algorithms. The depth fair selection strategy of Ito et al. is found to perform best compared with other subtree selection methods in the model refinement.