Reinforcement Point and Fuzzy Input Design of Fuzzy Q-Learning for Mobile Robot Navigation System

This paper proposes a fuzzy method to minimize state in reinforcement learning for obstacle avoidance mobile robot. The problem using reinforcement learning to solve obstacle avoidance is how to define the number of states. When the various conditions that might not be predicted by the mobile robot occur, the reinforcement learning is difficult to apply because the number of states become unlimited. The greater the number of states will cause the process requires more memory and performance of the processor. One of the popular methods for reinforcement learning is Q-learning. The Q-learning method is simple and feasible for observing surrounding conditions of mobile robot. The fuzzy used to eliminate the number of state problem, generalize the condition and reduce processor performance is needed. In this paper the comparison between reinforcement point using the change of state (S}{t-1}\rightarrow {\boldsymbol{S}}{\boldsymbol{t}}) and the current state ({S}-{t}) is conducted also the comparison of angle region sensor between 2, 3, 4 and 5. Simulation result show that reinforcement point using the change of state and 5 angle regions of sensor produce the best performance.