Research On Pursuit-Evasion Strategy Of Multi-Robot Based On Reinforcement Learning

As the development of computer technology and wireless communication technology, the research on Multi-Robot System (MRS) is receiving more and more attention. Compared with single robot, MRS has a flexible, efficient and fault-tolerant capability and other advantages. MRS can be widely applied in industries such as large ship manufacturing, space exploring, and unmanned combat system, as well as private and public service industry. The intelligent capture among the MRS is a testing platform for the effectiveness of robot learning strategy, it is of great significance to put intelligent mobile robot into industrial utility. Taking the multi-robot cooperation and coordination based on pursuit-evasion problem as a research object, this thesis designs a new capturing method. Details are as follows:First of all, this thesis summarizes the history of development and the present status of research on MRS, reinforcement learning and pursuit-evasion problem.Secondly, a dynamic model of the robot and the part hardware circuit of the robot are designed. The PC software and lower computer software are written. The basic principle and methods of reinforcement learning are briefly introduced and compared.Thirdly, a pursuit-evasion model for multi-robot is designed and some certain conditions are set in pursuit game. The escape robot takes an intelligent way to escape while pursuit robots take a reinforcement learning algorithm based on state prediction to round up. For the reinforcement learning used in MRS not meet the MDP, a state prediction method is taken; for Q-learning algorithm being hard to converge and being easy to converge to local optimum, an improved Q-learning algorithm is given. Dynamically adjusting three parameters a y T can make the Q-learning algorithm converge quickly; setting the variable interval can avoid the Q-learning algorithm converging to local optimum.Finally, reinforcement learning based on state prediction is applied in the pursuit game of multi-robot successfully, which contrasts with the contraction encirclement method in the absence of obstacles, obstructions environment and when robots are in failure. Also an optimum speed of the pursuit robots is analyzed and simulated in detail. In conclusion, the thesis is summarized, and future research is viewed.