Kriging-based multiobjective optimization using sequential reduction of the entropy of the predicted Pareto front

Abstract

In this paper, a novel Kriging-based algorithm for multiobjective optimization of expensive-to-evaluate black-box functions is proposed. The algorithm is based on sequential reduction of the entropy of the predicted Pareto front. The associated infill criterion selects a candidate design with the highest informational entropy among a set of predicted Pareto designs. The algorithm is tested on three engineering benchmark problems: the Nowacki cantilever beam, a car-side impact problem and a water management problem. The algorithm is also used to find the Pareto front for a snap-fit design. In the benchmark problems, the proposed algorithm outperformed traditional ones (parEGO and EHVI) when three different performance indicators were considered. The results also suggest that the algorithm is robust and produces a wide, representative Pareto front.