Space Partition based Gene Expression Programming for Symbolic Regression

For the problem of symbolic regression, we propose a novel space partition based gene expression programming (GEP) algorithm named SP-GEP, which helps GEP escape from local optimum and improves the search accuracy of GEP by letting individuals jump efficiently between segmented subspaces and preserving population diversity. It firstly partitions the space of mathematical expressions into k subspaces that are mutually exclusive. Then, in order for individuals to jump efficiently between these subspaces, it uses a subspace selection method, which combines multi-armed bandit and epsilon-greedy strategy. Through experiments on a set of standard SR benchmarks, the results show that the proposed SP-GEP always keeps higher population diversity, and can find more accurate results than canonical GEPs.