The Agent Intelligent Decision System In RoboCup2D Simulation League

In recent decades, with the rapid development of computer technology, the research and applications of distributed multi-agent systems have become the focus of a number of related disciplines.RoboCup,namely robot soccer World Cup,is generally considered the major event of studying the artificial intelligence of multi-agent system.RoboCup includes two categories of projects,namely simulation and physical robot.This paper is based the RoboCup simulation competitions,and study the decision making of a single agent including the teamwork of multi-agent in the real-time environment.In the RoboCup simulation competitions,the decision making of agent can be divided into high level decision making and low level decision making.High level decision making mainly work for teamwork of multi-agent,and can be divided into strategic level decision making and tactical decision making.High level decision making is for multi-agent system,specific to a single agent is for its behavior decision making,namely action decision making in RoboCup. The action decision making of a single agent means according to the current state of the environment and the cooperation agreement of our team before the match,to decide the aciton commands which should be sent to the server,and at the same time to update agent’s own state.Low level decision making usually refers to the action analysis of agent’s advanced behavior for decsions of some factors,such as the choice of the ball player when passing the ball and the choice of shoot point when shooting the ball.Machine learning can be applied into the decision system of agent in RoboCup to help it make more reasonable and effective decision-making in the real-time match. In the actual development of RoboCup simulation2D team,most of the designers and developers apply decision tree, artificial neural network or reinforcement learning in it.This paper introduces the machine learning in the study of agent’s advanced actions in order to make the implementation of agent’s advanced actions be better and further improve the success rate,so it can be very good service to the agent’s high level decision making.Applying decision tree learning algorithm to the agent’s passing training make the agent’s passing action more accurate.And applying the artificial neural network to the study of the agent’s shoot action greatly improve the success rate of the agent’s shoot action.At last,this paper applies the Q-Learning to local attacking decision of RoboCup simulation2D match, through the introduction of zoning, incentive functions based zoning and decision making for real soccer game action simulation, after training a large number of cycles of learning, making the agent do the independent action decision,thereby strengthening the multi-agent attacking strength.