Weak Convergence Detection-based Dynamic Reference Point Specification in SMS-EMOA

Abstract

In the evolutionary multi-objective optimization (EMO) field, the hypervolume (HV) indicator is one of the most popular performance indicators. It is not only used for performance evaluation of EMO algorithms (EMOAs) but also adopted in EMOAs for selection (e.g., SMS-EMOA). The specification of a reference point for HV calculation has a large effect on the performance of SMS-EMOA. Thus, the reference point specification should be carefully treated in SMS-EMOA. In this paper, the importance of the dynamic reference point specification in SMS-EMOA is explained first. Then a new dynamic reference point specification mechanism, which is based on weak convergence detection is introduced for SMS-EMOA. Experimental comparisons are conducted for different specification methods in SMS-EMOA: a linearly decreasing mechanism and two static mechanisms. Our results demonstrate the effectiveness of the proposed dynamic reference point specification mechanism.