Interactive MOEA/D with multiple types of preference information
Giomara Lárraga, Kaisa Miettinen · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2022
Multiobjective optimization has the ultimate aim of helping a decision-maker to find a satisfying solution for a problem with multiple conflicting objectives. Usually, multiobjective evolutionary algorithms have been utilized to approximate the complete Pareto optimal set. However, some such algorithms incorporate preference information to direct the search toward a region of interest. Interactive methods allow decision-makers to learn more about the problem and update their preference information iteratively. Although preference information can be represented in multiple ways, most multiobjective evolutionary algorithms restrict the decision-maker to a single type of preference information. This article proposes an interactive version of MOEA/D that incorporates three types of preference information: reference points, preferred solutions, and preferred ranges. The reference vectors assigned to the solutions that met the preferences are kept in the population for the next iteration, while the remaining ones are re-arranged. The proposed method is applied to solve a river pollution problem to show its potential to support decision-makers in finding a satisfying solution without limiting them to a single type of preference. This gives more flexibility for the decision-maker to direct the search for the most preferred solution. In addition, we compared our method with interactive RVEA utilizing some benchmark problems.