NEFRL: A New Neuro-Fuzzy System for Episodic Reinforcement Learning Tasks

In this paper, we propose a new neuro-fuzzy system for episodic reinforcement learning tasks, NEFRL. While NEFRL has all benefits of a neuro-fuzzy architecture, it has the additional advantage that it can learn with a numerical evaluation of performance and there is no need for training input-output pairs. Also, we show that the learning algorithm of this system converges with probability one to a local maximum of the average numerical performance signal. Our experimental results for the pole-balancing task show the power of this system even without any prior domain knowledge.