Ensemble Topology Selection in Simulation-Driven Optimization

Simulation-driven optimization problems are often computationally expensive and therefore metamodels are used to obtain predicted values at a lower computational cost. To further improve the prediction accuracy ensembles combine the prediction from multiple metamodels into a single output. However, the optimal ensemble topology, namely, which metamodel variants it should incorporate, is typically not known a-priori whereas using an unsuitable topology can degrade the prediction accuracy. To address these challenges this paper proposes an algorithm which continuously adapts the ensemble topology during the search such that an optimal topology is continuously being used. Numerical tests based on a variety of test problems show the effectiveness of the proposed algorithm. Index Terms - Black-box functions, design optimization, ensembles, metamodels.