A Novel Symbolic Regressor Enhancer Using Genetic Programming

This paper proposes a framework combining genetic programming (GP) with other symbolic regression (SR) methods, called the symbolic regressor enhancer (SRE). The basic idea is to use the syntax tree of the expression obtained from other SR methods to improve both the efficiency and the quality of the evolutionary procedure. Specifically, this paper investigates on the different ways of hybridization, selection, and crossover to assemble the proposed SRE. The effectiveness of SRE is demonstrated with the Taylor polynomial, the fast function extraction, and the GP-based SR methods, including Operon, the GP variant of gene-pool optimal mixing evolutionary algorithm, the epsilon-Iexicase selection, and gplearn. Out of 28 benchmarks from the SR benchmark and the Feynman SR database, the statistical test indicates that SRE applied to each selected SR method significantly outperforms the respective SR method in at least 8 and at most 24 benchmarks.