The Study On Multi-Agent Cooperative Algorithm Based On Game Theory

Using multi-agent system (MAS) Negotiation Mechanism, the agents can independently negotiate to assign and finish the tasks under unsupervised conditions and adaptively learn a new stategy to improve the efficiency in changing enviroment.During the process of MAS cooperating to complete complex task, each agent's best action not only depends on environment and task but also other agents' actions. The equilibrium situations of game theory are extensions of different individual optimality in MAS in which the strategy of any agent is the best-response strategy of it to the strategies of the others.In the past decade, algorithms based on game theory that regard the equilibriums as the optimal solution for the cooperation in multi-agent system (MAS) have been studied intensively. In this paper, we build a modified hawk-dove game model to simulate the interaction between agents based on evolutionary game theory, and then proposed a ponder-replicator algorithm to find certain consistent maximal reward equilibrium for the group. The algorithm based on evolutionary stable equilibriums (ESE) gained more focused not only because they can give a consistent optimal solution for the MAS, but also because it partly solved the equilibrium selection problem of game theory. But since ESE is a dynamic stable process, the strategies of every single agent keep on changing. As the consequence, the degree of the convergence is lower and the process of convergence is long as showed in the simulation.For these disadvantages in the ESE based on game theory algorithms, a reinforcement factor was introduced to strength the influence of the utilities of strategies, and then a ponder-reinforcement algorithm and a ponder-replicator-reinforcement algorithm were proposed to accelerate the process of convergence and improve the stability of convergent ESE. The multi-agent foraging simulations verified the efficiency of the proposed algorithms.