Research On Robocup Simulation System And Program Design Based On Reinforcement Learning

RoboCup simulation motivates distributed artificial intelligence, the technology of intelligent robot and the research and development of the field concerned by providing a standard task. RoboCup simulation game provides an environment of absolutely distributed control and real-time asynchronous multi-agent, and tests various theories, algorithm and client system architecture via this platform, and research the confrontation problem of multi-agent in the real-time asynchronous noisy environment. The core technology of RoboCup is the technology of artifical intelligence, the purpose is to make machine have the wisdom, perceive the environment like people and the learning capabilities to the environment.This paper mainly research RoboCup simulation game. RoboCup simulation game is running in a standard computer environment. The game mode is provided by RoboCup committee with standard Soccer Server and each team programs their own Client programs which simulate the actual soccer team member to take part in the game.First, design and implement of RoboCup simulation system is analyzed in this paper, and then client system structure and program flow of UvA-Trilearn are primarily studied; at last program is designed on the basis of it. It mainly includes methods of increasing scenery strategy and evaluation function and so on. It also researches maximin Q learning algorithm based on Markov decision process. Simulation results show that this method can resolve the confrontation problem between multi-agents well.The main work of this paper is as follows:(1) Research system structure and operating priciple of the whole RoboCup simulation platform, this is the base of designing the RoboCup simulation team.(2) Research the basic action of RoboCup simulation,analyze the features of the actions. On this basis, action evaluation function is designed and scenery strategy is increased.(3) Programe the client with C++ language in linux OS and debug.(4) Research Markov decision process and reinforcement learning algorithm, and design learning algorithm based on the maximin Q, and the application in the problem of shortest path and in the RoboCup. Simulation results show that the algorithm can better solve the confrontations problem between Multi-Agents.