The optimal filtering set problem with application to surrogate evaluation in genetic programming

Abstract

Surrogate evaluation is common in population-based evolutionary algorithms where exact fitness calculation may be extremely time consuming. We consider a Genetic Program (GP) that evolves scheduling rules, which have to be evaluated on a training set of instances of a scheduling problem, and propose exploiting a small set of low size instances, called filter, so that the evaluation of a rule in a filter estimates the actual evaluation of the rule on the training set. The calculation of filters is modelled as an optimal subset problem and solved by a genetic algorithm. As case study, we consider the problem of scheduling jobs in a machine with time-varying capacity and show that the combination of the surrogate model with the GP termed SM-GP, outperforms the original GP.