Leveraging indicator-based ensemble selection in evolutionary multiobjective optimization algorithms

Abstract

Various evolutionary multiobjective optimization algorithms (EMOAs) have replaced or augmented the notion of dominance with quality indicators and leveraged them in selection operators. Recent studies show that indicator-based EMOAs outperform traditional dominance-based EMOAs. This paper proposes and evaluates an ensemble learning method that constructs an ensemble of existing indicators with a novel boosting algorithm called Pdi-Boosting. The proposed method is carried out with a training problem in which Pareto-optimal solutions are known. It can work with a simple training problem, and an ensemble of indicators can effectively aid parent selection and environmental selection in order to solve harder problems. Experimental results show that the proposed method is efficient thanks to its dynamic adjustment of training data. An ensemble of indicators outperforms existing individual indicators in optimality, diversity and robustness. The proposed ensemble-based evolutionary algorithm outperforms a well-known dominance-based EMOA and existing indicator-based EMOAs.