An Unbounded Archive-Based Inverse Model in Evolutionary Multi-objective Optimization

Rongguang Ye, Longcan Chen, Jinyuan Zhang, Hisao Ishibuchi

Abstract: Inverse model (IM) is a method for tailoring solutions to decision-makers based on their preferences. Existing approaches are often trained by the final solution set (as training data) obtained from a multi-objective evolutionary algorithm (MOEA). The final solution set obtained by MOEA usually has a limited number of samples. However, model training will perform poorly when there are few samples in the final solution set. To further improve the performance of the model, we propose an unbounded archive-based inverse model (UAIM) to enhance the quality of the trained inverse model. We first create an unbounded archive to collect all non-dominated solutions during the execution of MOEA. Unlike IM, UAIM is trained using all solutions in the archive. Moreover, for a decision maker’s preference, an alternative solution from the archive is considered if the suggested solution is inferior to the alternative solution in the archive. UAIM thus may provide more reliable suggested solutions for decision-makers. To better evaluate algorithms, we propose two indicators that can measure the matching degree between the suggested solution and the decision maker’s preference. We demonstrate that the proposed UAIM is superior to IM on ten problems.