Building A Fuzzy Inference System By An Extended Rule Based Q-Learning

Building a Fuzzy Inference System (FIS) generally requires experts’ knowledge. However, experts’ knowledge is not always available. When there is few experts’ knowledge, it becomes hard to build a FIS using one of supervised learning methods. Meanwhile, Q-learning is a kind of reinforcement learning where an agent can acquire knowledge from its experiences even without the model of the environment and experts’ knowledge. The Q-learning, however, has weakness that the original algorithm cannot deal with the continuous states and continuous actions. In this paper, we proposed a FIS that can do Q-learning. The proposed FIS structure is made up of several extended rules. Based on these extended rules, Q-learning algorithm for the proposed structure is developed. It is shown that this combination results in a FIS that can learn through its experience without experts’ knowledge. Also the proposed structure can resolve the continuous state/action problem in Q-learning by virtue of a FIS. The effectiveness of the proposed structure is shown through simulation on the cart-pole system