多智能体协作策略的研究及在RoboCup中的应用

【作者】 刘苗；

【导师】 彭军；

【作者基本信息】 中南大学 ， 计算机应用技术， 2008， 硕士

【摘要】 多智能体协作是多智能体系统MAS研究的核心问题。在复杂、动态、不确定的多智能体环境中,为了满足多智能体协作中对局部配合和通信受限的要求,本文对智能体策略寻优、行为协调和动作规划问题进行研究,来构建适用于不同情况下的多智能体协作策略,并在典型的MAS——RoboCup机器人足球仿真系统平台下进行检验。首先,为了实现智能体行为选择的优化和多智能体的局部协作,提出基于行为协同优化的多智能体协作策略。智能体通过模块化模糊Q学习对其他智能体的行为进行评估,考虑它们的行为对自身行为选择的约束,来优化自身的行为决策,再采用共享联合意图的协调方法解决智能体之间的行为冲突,得到其最优行为策略。其次,在通信受限的情况下,提出基于多智能体行为图的分层规划协作策略。根据智能体感知到的局部环境信息,利用行为图对其行为过程进行预测规划,再结合模块化模糊Q学习中获得的行为选择的先验知识,逐层调整其初始行动计划,获得智能体协调一致的动作规划序列,使其针对当前环境快速做出有效决策来实现与其他智能体的协作。论文提出的多智能体协作策略应用到中南大学CSU_YunLu机器人足球仿真球队中,在实际训练和对抗比赛中验证了其有效性。更多还原

【Abstract】 Multi-agent cooperation is an important research focus of multi-agent system (MAS). In complex, dynamic and uncertain multi-agent environment, this dissertation studies these problems, such as strategy optimization of single agent, behavior coordination and action planning, to satisfy the requirements of local collaboration and communication limitation in the process of multi-agent cooperation. Then multi-agent cooperation strategies are conducted to be applicable in different cases and examined in RoboCup soccer simulation system.Firstly, in order to implement behavior selection optimization of the agent and local collaboration of multiple agents, a multi-agent cooperation strategy based on behavior common optimization is proposed. Each agent uses modular fuzzy Q-learning to speculate the behaviors of other agents. Considering their behavior restrictions, individual behavior decision-making is optimized. Then the behavior conflicts among agents are solved by the coordination method sharing joint-intentions to obtain the optimized behavior strategy.Secondly, a layered planning cooperation strategy based on multi-agent behavior graph is presented in the case of communication limited. According to the local environment state information that agents observe, the behavior process of agents is planned using behavior graph in advance. Then combining with the prior knowledge of behavior selection obtained by modular fuzzy Q-learning, initial activity planning is gradually adjusted from lower layer to higher one, so that consistent action sequence of each agent is acquired, which ensures the agent to make action decision fleetly against current environment to cooperate with others neatly.These proposed cooperation strategies above have been applied into CSU_YunLu simulation team. The feasibility is verified in actual antagonism training and competition.更多还原