Multi-noisy-objective optimization based on prediction of worst-case performance

Abstract

This paper proposes a new approach to cope with multiobjective optimization problems in presence of noise. In the first place, since considering the worst-case performance is important in many realworld optimization problems, a solution is evaluated based on the upper bounds of respective noisy objective functions predicted statistically by multiple sampling. Secondary, a rational way to decide the maximum sample size for the solution is shown. Thirdly, to allocate the computing budget of a proposed evolutionary algorithm only to promising solutions, two pruning techniques are contrived to judge hopeless solutions only by a few sampling and skip the evaluation of the upper bounds for them.