多机器人追捕问题的分析-计算机应用技术专业论文.docxVIP

哈尔滨工业大学工学硕士学位论文 哈尔滨工业大学工学硕士学位论文 - - II - Abstract Multi-robot pursuit problem provides an ideal platform for the study of coordination and collaboration among the robots. The application of reinforcement learning algorithm to the pursuit problem can make multi-robot systems become initiative to explore the environment, adapt to the environment and improve its own performance and stabiltiy. Since the direct application of standard reinforcement learning algorithm to multi-robot system always results in the exponential growth of the state space, which reduces convegence speed of the algorithm, it attracts less use in practice. This dissertation is devoted to reducing the size of the state space system and improving the convergence speed of algorithm. The main content are as follows. Firstly, the basic framework of reinforcement learning algorithm with the mathematical model is presented and the common reinforcement learning algorithms and flows are listed. Then we give a description of the multi-robot pursuit problem, state and action abstraction and definition of reward function. Considering the same state existes in the traditional method of state abstraction, we come up with a dynamic ID state abstraction method, which reduces the size of the state space and some comparisons with traditional method based on the standard Q -learning algorithm are also presented. Secondly, basic principles of hierarchical reinforcement learnin g are given. The state space decomposition method is proposed to divide the original state space into multiple parts, that is, the application of OPTION- Q learning algorithm makes optimal strategy dispersed into each sub-space so that it can reduce the amount of space policy to accelerate the convergence speed. The comparison between the OPTION- Q and the standard Q based on dynamic ID state abstraction method are made. Thirdly, value function decomposition is developed to improve the OPTION- Q learning algorithm. The sub-tasks state value