Mutagenesis as a Diversity Enhancer and Preserver in Evolution Strategies

Abstract

Mutagenesis is a process which forces the coverage of certain zones of the search space during the generations of an evolution strategy, by keeping track of the covered ranges for the different variables in the so called gene matrix. Originally introduced as an artifact to control the automated stopping criterion in a memetic algorithm, ESLAT, it also improved the exploration capabilities of the algorithm, even though this was considered a secondary matter and not properly analyzed or tested. This work focuses on this diversity enhancement, redefining mutagenesis to increase this characteristic, measuring this improvement over a set of twenty-seven unconstrained optimization functions to provide statistically significant results.

A Guided Mutation Operator for Dynamic Diversity Enhancement in Evolutionary Strategies

“…This can be seen according to two different reasonings: theoretically, when the exploitation is more efficient, the algorithm will favor canonical cycles. Empirically, the results obtained are more consistent than the ones available in (Guerrero, Gómez-Jordana, Berlanga, & Molina, 2012), even though the populations sizes are greater (reducing the need for enhanced exploration) and the stopping criterion has been removed.…”

“…Table 1, which represents the worst case for the strategy, since it shows the easiest problems, presents five results where mutagenesis is better according to mean value, but only one is considered statistically significant (even tough in some cases the median is equal or even lower), while four out of the five cases where canonical technique is better are considered statistically significant. This can lead to a discussion of the representativeness of non-parametric tests such as the one presented in Guerrero, Gómez-Jordana, Berlanga, and Molina (2012).…”

“…In (Guerrero, Gómez-Jordana, Berlanga, & Molina, 2012) gene matrix and mutagenesis artifacts were extracted from the technique and modified to enhance their diversity management capabilities. The gene matrix is an artifact which tracks the zones of the search space covered and forces the exploration of those which had not been reached by the main search cycle, also determining the generation at which the algorithm should be stopped according to that search space coverage.…”