Algorithm Of Learning And Control Based On Robocup 3D Simulation

Multi-agent system (MAS) and distributed artificial intelligence research have been the focus of recent academic research. Robot World Cup (Robocup) which has a dynamic environment is a typical MAS. As each player of the team can be treated as an Agent, it is an ideal research platform of MAS. We can experiment different algorithms on the platform then use them to other system.Robocup3D is a new type of game, it is closer to real life than 2D competition. It is becoming a new hotspot.First, the paper explained the structures, referee system and working principles of the simulation server. This paper also describes the environmental information of the game and the way in which players perceive information and execute command.Second, the paper introduced the classification and conception of Agent. Based on the feature of the game, a kind of new model which is based on mixing layer structure has been put forward. In order to improve the expansibility of the program, the program was divided into 6 modules. Each module has its own function. There is also a redefinition of the individual strategy level, we introduce a new tactics which is called"Mid position"into defensive strategy. Kalman filter is used for improve the precision accuracy of location. A kind of new algorithm which is used to judge the standing state of player is presented. Furthermore, there is a new type of interception algorithm. Both of them are used in the game and get good effect. By way of reasonable gait planning, the player has been realized stable walking.At last, we explained the theory of reinforce learning. Q learning which is a kind of reinforce learning was used for the training of vanguard's shooting. After several cycles of training, the success rate of player's shot has been improved effectively.AFU2010 is developed by the above algorithms. The system platform is OpenSuSE11.2, IDE which is built by VIM is used for the coding of the program.