Using preference learning for multi-objective optimization with applications in supply chain

Choosing the right weight is a challenging task in solving a multi-criteria decision making (MCDM) problems. We utilize the learning-to-rank machine learning approach, Rank SVM, to learn the criteria weights in MCDM. As the training data, Rank SVM needs the pairwise preferences of the alternatives, as revealed by the decision maker (DM). We develop three strategies in offering alternative pairs to DMs. The first strategy is offering pairs from the Pareto frontier which represents a set of optimal solutions, the second strategy is offering pairs from the feasible region meaning dominated and non-dominated solutions that are possible given the constraints and the third one is offering pairs from the utopian space that covers both feasible and infeasible solutions. The main objective of this study is to evaluate the impact of offering pairs from different regions on the learning process of Rank SVM and utilizing information learned in data generation strategies. To evaluate the performance and effectiveness of our strategies, we chose a three-echelon supply chain network problem as our test case. Experimental results obtained from three different settings provide a practical evaluation. We observe distinct impacts between strategies in offering alternative pairs; some strategies yield more accurate or consistent results than others. This highlights the importance of the source of alternative pairs in the effectiveness of preference learning algorithms. In addition, the use of learning information in the generation of training data provided a significant improvement except the Utopian region strategy.