多任务联盟形成中的Agent行为策略研究

Agent联盟是多Agent系统中一种重要的合作方式,联盟形成是其研究的关键问题.本文提出一种串行多任务联盟形成中的Agent行为策略,首先论证了Agent合作求解多任务的过程是一个Markov决策过程,然后基于Q-学习求解单个Agent的最优行为策略.实例表明该策略在面向多任务的领域中可以快速、有效地串行形成多个任务求解联盟.

Agent-behavior strategy in serial multi-task coalition formation

Agent-coalition is an important approach to agent-coordination and cooperation, in which the coalition formation is a key topic. Existing researches are restricted in single-task environments, and the results are not applied to multi-task environments. In this paper, a new agent behavior strategy in serial multi-task coalition formation for problemsolving is presented. The conclusion shows that the agent-task selection is a Markov Decision Process. The Q-learning is used to optimize the behavior strategy for a single agent, and the cooperative multi-agent reinforcement learning improves the learning rate. Experiments prove that the strategy can effectively and serially form coalitions for multi-task.