Characterizing Diversity in Genetic Programming

In many evolutionary algorithms candidate solutions run the risk of getting stuck in local optima after a few generations of optimization. In this paper two improved approaches to measure population diversity are proposed and validated using two traditional test problems in genetic programming literature. Code growth gave rise to improve pseudo-isomorph measures by eliminating non-functional code using an expression simplifier. Also, Rosca's entropy to measure behavioral diversity is updated to cope with problems producing a more continuous fitness value. Results show a relevant improvement with regard to the original diversity measures.