Progressive Preference Learning: Proof-of-Principle Results in MOEA/D

Most existing studies on evolutionary multi-objective optimisation (EMO) focus on approximating the whole Pareto-optimal front. Nevertheless, rather than the whole front, which demands for too many points (especially when having many objectives), a decision maker (DM) might only be interested in a partial region, called the region of interest (ROI). Solutions outside this ROI can be noisy to the decision making procedure. Even worse, there is no guarantee that we can find DM preferred solutions when tackling problems with complicated properties or a large number of objectives. In this paper, we use the state-of-the-art MOEA/D as the baseline and develop its interactive version that is able to find solutions preferred by the DM in a progressive manner. Specifically, after every several generations, the DM is asked to score a limited number of candidates. Then, an approximated value function, which models the DM’s preference information, is learned from the scoring results. Thereafter, the learned preference information is used to obtain a set of weight vectors biased towards the ROI. Note that these weight vectors are thus used in the baseline MOEA/D to search for DM preferred solutions. Proof-of-principle results on 3- to 10-objective test problems demonstrate the effectiveness of our proposed method.