The Research Of Mobile Robot Path Planning Based On Reinforcement Learning And ART2 Neural Network

Based on the reinforcement learning and ART2 neural network, the paper does the research on the path planning problem of mobile robot and analyzes the problem of collision avoidance of the path planning.The main work and creative points of the paper are as follows:(1) Firstly, the paper presents an action selecting policy named gradual Soft-Max to solve the problem of selecting policy in reinforcement learning. By using the gradual Soft-Max, the agent can explore to acquire more action experience in the beginning of learning and exploit the actions in the anaphase of learning due to accumulating enough action experience. It also can make a break to accelerate or slower the learning speed. Meanwhile, this paper presents a reinforcement learning method named S-Learning for continuing task and a learning method based on foremost-policy named FPRL (Foremost-Policy Reinforcement Learning).(2) The paper uses ART2 to store abundant classify patterns and state space in order to solve the problem that the traditional lookup-table doesn't adapt to store great amounts of state and action values. We combine the mechanism of selection and evaluation in reinforcement learning with the ART2 as RL-ART2(Reinforcement Learning based ART2 Neural Network)to solve the difficulties of evaluating and selecting the stored classify patterns in ART2 manually.(3) The paper finally presents a RL-ART2 based collision avoidance system of mobile robot (RLART2-CAS). In order to solve the difficult problem of acquiring the collision avoidance rules by hand, the paper utilizes RL-ART2 to make system acquire the collision avoidance rules automatically. We also utilizes RL-ART2 neural network to solve the problem of rules storing and evaluate the collision avoidance result automatically. The RLART2-CAS can perfectly accomplish the path planning by acquiring the ideal collision avoidance actions through reinforcement learning.