Near Feasibility Driven Adaptive Penalty Functions Embedded MOEA/D

Abstract

This work extends a near feasibility threshold (NFT) based adaptive penalty function for constrained multiobjective optimization. The NFT zone adjoining the feasible region is considered as good one, where infeasible solutions are relatively less penalized. The modified penalty function, denoted by TAP with five different settings of NFT is embedded in a renowned multiobjective evolutionary algorithm based on decomposition, MOEA/D. This offers five constrained variants, namely CMOEA/D-TAP1 to CMOEA/D-TAP5, of the base algorithm. These variants are tested on well-known constrained multiobjective benchmark test suits, the CTP series and the CF series. The proposed variants are compared with the four best performing algorithms through HV (hyper volume) metric statistics for CTP series, and with seven state-of-the-art algorithms through Wilcoxon rank sum test employed to mean values of both HV and IGD (inverted generational distance) metrics for CF series. Simulation results reflect that overall performance of the newly introduced variants is better than the competitors for the taken benchmark test suits.

Keywords

• Constrained optimization
• penalty function methods
• near feasibility threshold
• multiobjective optimization based on decomposition