Mobile Robot Path Planning Based On Simulated Annealing-q Learning

Among the technical study of mobile robot, navigation is a key technology of intelligence and autonomy, and also one of the current research focus. Path planning is the basic issue of navigation, therefore it is of great significance for intelligence and autonomy to research on mobile robot path planning and improve the adaptability of the unknown environment.After analysing methods of mobile robot path planning, the thesis focuses on the reinforcement learning of Q learning. However, there are many issues in path planning based on reinforcement learning, such as reward function designing, tradeoff of exploration and exploitation, generalization of continuous state and action, etc. According to the above issues, some solutions are correspondingly presented, and the algorithm for mobile robot path planning in unknown environment is proposed. The specific work is as follows:To solve the impact on the convergence rate and tradeoff of exploration and exploitation, a SA-Q learning based on behavior-based decomposition of reward function mobile robot path planning method is proposed. While an uneven reward function is designed to minimize the impact on the convergence rate, simulated annealing(SA) approach is used to select action to solve tradeoff of exploration and exploitation. Simulation results show that the method has improved the convergence rate, solved tradeoff of exploration and exploitation, and could make mobile robot find the sub-optimal path.A SA-Q learning algorithm based on dynamic programming is presented to enhance the convergence rate of SA-Q learning and improve the performance of Q learning based on dynamic programming. Dynamic programming is used to speed up the convergence rate, and the improvement of performance is achieved by SA. The simulation results show that the algorithm has faster convergence rate, higher performance, and could make mobile robot find a collision-free path.For generalization of continuous state and action, a SA-Q learning based on fuzzy inference system (FIS) is proposed. FIS is used to generalize continuous state and action and to determine the output of the system as the action of mobile robot. Simulation results show that the algorithm has strong ability of generalization, and has effectively solved the mobile robot path planning in complex environment.