Research On Multi-agent Cooperation Mechanism In "Predator-prey" Problem

The thesis summing up the domestic and foreign multi- Agent system theory and multi-Agent cooperation problem solving approach, has primary studied of multi-Agent system cooperation mechanisms which based on the Contract Net Protocol, used " Predator - prey "problem, which under the same conditions, has the unknown environment and single "prey", to make simulation and analysis. As well as make a study of the improvement of the efficiency of cooperative capture which in the "predator - prey" system. The main work of thesis includes:1. Proposed a Contract Net Protocol cooperation model which based on k-means clustering algorithmAnalyzed the basic principles, characteristics, the general steps of the distribution of tasks, as well as the deficiencies of Contract Net Protocol. In view of task announcement and biding which easy for the communication load, and after profited from k-means clustering algorithm, designed a Contract Net Protocol cooperation model which based on k-means clustering algorithm, which can be effectively solve the communication load caused by participants increase in the traditional contract net protocol. Then established and simulated "Predator - prey "system in the Swarm, the simulation results show the theoretical analysis results2. imported Q-learning to improve the efficiency of cooperative capture of "predator - prey" systemanalyzed the cooperative capture efficiency problems of "predator - prey" system. and had considered premise conditions that a completely cooperative multi-Agent systems, named "predator - prey" system which can be directly applied with a similar single-Agent reinforcement learning, as well as the advantage of Q-learning algorithm which belongs to single-Agent reinforcement learning that take no account of environment model. Then chose Q-learning algorithm as a learning algorithm for the Agent. designed the steps of "Predator" Agent using Q-learning algorithm to learn. the simulation results indicates "Predator" Agent can obtain optimal control strategies after using Q-learning algorithm, and this approach is effective on improving efficiency of cooperation capture.