A Self-Organizing Neuro-Fuzzy Q-Network: Systematic Design with Offline Hybrid Learning

In this paper, we propose a systematic design process for automatically generating self-organizing neuro-fuzzy Q-networks by leveraging unsupervised learning and an offline, model-free fuzzy reinforcement learning algorithm called Fuzzy Conservative Qlearning (FCQL). Our FCQL offers more effective and interpretable policies than deep neural networks, facilitating human-in-the-loop design and explainability. The effectiveness of FCQL is empirically demonstrated in Cart Pole and in an Intelligent Tutoring System that teaches probability principles to real humans.