A New Mutation Paradigm for Genetic Programming

Abstract

Levy flights are a class of random walks directlyinspired by observing animal foraging habits, where apower-law distribution of the stride length can beoften observed. This implies that, while the vastmajority of the strides will be short, on rareoccasions, the strides are gigantic. We propose amutation mechanism in Linear Genetic Programminginspired by this ethological behaviour, thus obtaininga self-adaptive mutation rate. We experimentally testthis original approach on three different classes ofproblems: Boolean regression, quadratic polynomialregression, and surface reconstruction. We find that inall cases, our method outperforms the generic, commonlyused constant mutation rate of one over the size of thegenotype. Moreover, we compare different common valuesof the power-law exponent to the another self-adaptivemutation mechanism directly inspired by SimulatedAnnealing. We conclude that our novel method is aviable alternative to constant and self-adaptivemutation rates, especially because it tends to reducethe number of parameters of genetic programming.