Coarse Tuning of Fuzzy Reinforcement Learning Architecture using Value Iteration Method

This research presents a new method of coarse-tuning the fuzzy reinforcement learning architecture using agent-environment interactive data. This method solves two main challenges the fuzzy reinforcement learning: the low-speed training process and determining the input membership functions of the fuzzy structure. First, the agent interacts with the environment to gather training data. Then, the transition probability matrix and the expected return matrix are calculated by applying a clustering algorithm due to the continuous space. Each cluster is a state of the environment, and an approximation of the transition probability from one cluster to another is calculated using the gathered data. Finally, the parameters of the fuzzy system are adjusted using the modified value iteration method of dynamic programming for the continuous space. The proposed method is fully described with an example. This method increases the learning speed and adjusts the input membership functions of the fuzzy system.