Ensemble Selection in Expensive Optimization Problems

Abstract: Computationally intensive simulations are being extensively used across engineering and science in various design optimization problems. To alleviate the high computational load associated with each simulation run metamodels are used, as they provide predicted objective values at a lower computational cost. However, the optimal metamodel variant is typically unknown and is problem-dependant. In an attempt to alleviate this ensembles use multiple metamodels concurrently and aggregate their predictions. However the optimal ensemble configuration is also problem-dependant and typically unknown. To address this issue, this paper proposes an approach in an optimal ensemble configuration is selected during the search out of a family of candidate ensembles, without a need for user intervention or a-priori domain knowledge. Performance analysis shows that the proposed approach improved the search effectiveness over a range of test problems.