Effects of Mutation Before and After Offspring Selection in Genetic Programming for Symbolic Regression

In evolutionary algorithms mutation operators increase the genetic diversity in the population. Mutations are undirected and have only a low probability to improve the quality of the manipulated solution. Offspring selection determines if a newly created solution is added to the next generation of the population. By definition, offspring selection is applied after mutation and the effects of mutation are directed and quality-driven. In this paper we propose an alternative variant of genetic programming with offspring selection where mutation is applied to increase genetic diversity after offspring selection. We compare the solution quality achieved by the original algorithm and the new algorithm when applied to a symbolic regression problem. We observe that solutions produced by the new variant have a smaller generalization error and conclude that the proposed variant is better for symbolic regression with linear scaling.