SQ-FMFO: A Novel Scalarized Multi-objective Q-Learning Approach for Fuzzy Membership Function Optimization

Association Rule mining plays a crucial role in data mining by extracting relationships among the data items in the form of association rules. Fuzzy association rule mining is one of the major techniques in rule mining, and it often provides better rules and flexibility over quantitative rule mining by utilizing fuzzy set theory. The outcome of fuzzy association rule mining is greatly affected by the appropriateness of its membership functions. In this research, we have provided a comprehensive description of fuzzy membership function optimization as a Markov decision process (MDP) as well as introduced two novel algorithms- one for membership function optimization and the other for choosing the optimal number of membership functions for any given problem. This study proposes a novel multi-objective reinforcement learning technique that incorporates scalarization of different value functions while determining optimal policy and is able to construct optimized membership functions with better suitability and coverage. The proposed multi-objective reinforcement learning algorithm outperforms well-known techniques in fuzzy membership function optimization by utilizing the exploration and exploitation mechanism. The performance of the proposed techniques is clearly validated by the experimental analysis incorporated in this study and is the first study that concisely treats fuzzy membership function optimization in terms of an MDP environment it opens up enumerable opportunities for future reinforcement learning endeavors.