Adapting Learning Classifier Systems to Symbolic Regression

Genetic programming (GP) approaches have been widely studied for symbolic regression problems and have achieved substantial progress. This work investigates the effectiveness of niching property and multiple learned solutions of a Learning Classifier System (LCS) to symbolic regression benchmark problems. Specifically, an XCS with real-valued interval based conditions and code fragmented action termed as XCS-SR is proposed for tackling symbolic regression problem. This is the first LCS ever to address the problem of symbolic regression. The results on nine standard symbolic regression benchmarks show that the proposed XCS-SR method consistently obtains statistically-better results on a majority of the benchmarks, in terms of average absolute error together with an increased number of exact solutions as compared with the GP benchmark.