Towards explaining metaheuristic solution quality by data mining surrogate fitness models for importance of variables

Abstract

Metaheuristics are randomised search algorithms that are effective at finding "good enough" solutions to optimisation problems. However, they present no justification for generated solutions and these solutions are non-trivial to analyse in most cases. We propose that identifying the combinations of variables that strongly influence solution quality, and the nature of this relationship, represents a step towards explaining the choices made by a metaheuristic. Using three benchmark problems, we present an approach to mining this information by using a "surrogate fitness function" within a metaheuristic. For each problem, rankings of the importance of each variable with respect to fitness are determined through sampling of the surrogate model. We show that two of the three surrogate models tested were able to generate variable rank-ings that agree with our understanding of variable importance rankings within the three common binary benchmark problems trialled.