A Reinforcement Learning-Based Framework for Multi-Criteria Decision-Making

Abstract

Multi-Criteria Decision-Making (MCDM) methods have been applied to various ranking problems in finance, engineering, and management; however, many traditional methods rely on data normalization and fixed weights assigned to the criteria. Often, such data normalization has led to distortion of ordinal information and inconsistent results. In this study, we propose an RL-based framework for solving MCDM problems when the decision data are ranked. The ranking task was then reformulated as a learning-to-rank optimization problem, where the preference scores are assigned to alternatives and learned by maximizing agreement with a benchmark ranking. The proposed approach was first illustrated with a simple example and then applied to the real-world evaluation of banks using the CAMELS criteria. The results showed that the RL-based approach achieved greater consistency with benchmark rankings than normalization-based MCDM methods, thereby extending the applicability of rank-preserving approaches.