Online Ensemble Topology Selection in Expensive Optimization Problems

Simulation-driven optimization problems are often computationally-expensive, an aspect which has motivated the use of metamodels as they provide approximate function values more economically. To further improve the prediction accuracy the use of ensembles has been explored in which predictions from multiple metamodels are combined. However, the optimal ensemble topology, namely, which types of metamodels it includes, is typically not known, while using a fixed topology may degrade the prediction accuracy and search effectiveness. To address this issue this paper proposes a metamodel-assisted algorithm which autonomously adapts the ensemble topology online during the search such that an optimal topology is used throughout. An extensive performance analysis shows the effectiveness of the proposed algorithm and approach.