Collaborative and Adaptive Strategies of Different Scalarizing Functions in MOEA/D

Abstract

In recent years, the use of decomposition-based multi-objective evolutionary algorithms has been very successful in solving both multi- and many-objective optimization problems. In these algorithms, the adopted Scalarizing Functions (SFs) play a crucial role in their performance. Methods such as the Modified Weighted Chebyshev (MCHE), Penalty Boundary Intersection (PBI) and Augmented Achievement Scalarizing Function (AASF) have been found to be very effective for achieving both convergence to the true Pareto front and a uniform distribution of solutions along it. However, the choice of an appropriate model parameter is required for these SFs. Some studies have analyzed the impact of these parameter values on the performance of the best-known decomposition multi-objective evolutionary algorithm (MOEA/D). In this paper, we propose a strategy based on collaborative populations combining different SFs and model parameter values via an adaptive operator selection based on the multi-armed bandit technique. Our preliminary results give rise to some interesting observations regarding the way in which different SFs are combined and adapted during the evolutionary process of MOEA/D.