Semantics Based Mutation in Genetic Programming: the Case for Real-valued Symbolic Regression

In this paper we propose two new methods for implementing the mutation operator in Genetic Programming called Semantic Aware Mutation (SAM) and Semantic Similarity based Mutation (SSM). SAM is inspired by our previous work on a semantics based crossover called Semantic Aware Crossover (SAC) [19] and SSM is an extension of SAM by adding more control on the change of semantics of the subtrees involved in mutation operation. We apply these two new mutation operators to a class of real-valued symbolic regression problems and compare them with the Standard Mutation (SM) of Koza [13]. The results from the experiments show that while SAM does not help to improve the performance of Genetic Programming, SSM helps to significantly enhance Genetic Programming performance on the problems tried. The experiment results also show that the change of the semantics (fitness) in SSM is smoother than ones of both SAM and SM. This, we argue that is the main reason to the significant performance improvement of SSM over SAM and SC.