Integrating region preferences in multiobjective evolutionary algorithms based on decomposition

Abstract

User preference is of great importance when dealing with many objective optimization. Using the preference information to obtain preferred parts of the Pareto set has become prevalent in the research domain of Evolutionary Multiobjective Optimization (EMO). In this paper, a target region provided by the decision maker (DM), defined by the preferred range of every objective, is utilized to articulate the preference in-formation. This information is integrated with two well-known multiobjective evolutionary algorithms based on decomposition: MOEA/D and NSGA-III. The newly proposed preference-based algorithms, called T-MOEA/D and T-NSGA-III, can be used both a-priori and interactively. Experiments have demonstrated the benefit of applying them interactively. The DM can easily and quickly adjust the preferences according to the current results, and the proposed algorithms can successfully find non-dominated solutions complying with the preferences. Comparative experiments show that the proposed algorithms outperform the dominance-based algorithm T-NSGA-II on many-objective benchmark problems.