Research On Multi-Agent Cooperation In RoboCup

The multi-agent cooperation is an important research hot spot in the field of distributed artificial intelligent. The RoboCup simulation game is a standard ideal platform for studying multi-agent cooperation problem, which simulates the human football game. The cooperation between the agents plays the most important role in winning the match in the complex multi-agent environment. In this paper, learning and prediction techniques are used to design the multi-agent cooperation strategies and models. The main research work is as follows.First of all, the multi-agent Q-learning algorithm based on the multi-dimensional CMAC network is proposed to solve the problem of cooperation strategy between the agents. This algorithm is aimed at increasing cooperation between the agents by learning the traditional action as well as storing Q value table of each move using CMAC network on the basis of traditional Q-learning. This method not only improves the generalization ability of Q-learning but also raises the learning speed and accuracy. The method is now successfully applied to the off-line learning of multi-agent cooperation in RoboCup, which solve2VS1, a typical sub-problem in RoboCup.Secondly, given the situation of more agents, a multi-agent cooperation model based on the behavior prediction is put forward to improve the overall multi-agent cooperation ability, which also helps make cooperation model design simpler, faster, more adaptive and higher intelligent. Through a combination of behavior prediction and motion prediction, multi-agent cooperation ability is increased in a large scale. The team’s cooperation decision-making can also be realized with applying the cooperation model based on the behavior prediction to RoboCup simulation game.Finally, due to the complexity of manual implementation of overall design of the multi-agent and the big influence to the formation of cooperation strategy that prediction accuracy causes, the learning module is added in the prediction model, the Q-learning method based on prediction is proposed. It combines Q-learning algorithm and prediction techniques, thus improves the online learning ability of the agent. And the method is used in the3VS4test to verify its effectiveness.The practical application in the simulation team has verified the effectiveness of the multi-agent cooperation strategy and model raised in this paper.