【摘要】 为了在multi-agent系统中实现agent之间的竞争与协作,该文提出了一种新的在线学习方法,即:改进的模糊Q学习方法,在这种方法中,agent通过增强学习方法来调节模糊推理系统,进而获得最优的模糊规则。为了改善学习的时间,Q学习方法中的奖励值并不是固定的,而是根据状态而变化。将改进的模糊Q学习方法应用到RoboCup仿真环境中,使智能体通过在线学习获得跑位技巧。并通过实验证明了该方法的有效性。更多还原

【Abstract】 <Abstrcat> In order to achieve cooperation and competition among agents in multi-agent systems, this paper proposed a new on-line learning method, named Modified Fuzzy Q-Learning Algorithm(MFQLA).In this method, agent tuned Fuzzy Inference System(FIS) by Reinforcement Learning(RL) method, and found the optimal fuzzy rules. Trying to improve the learning time, the reward values in Q-learning method are not constant. MFQLA tuned the reward values according to current state. In this paper, MFQLA was applied to the environment of Robot World Cup Tournament. It is expected that the learning agent obtained the efficient positioning skills.The results of simulation and real experiments indicate the effectiveness of the proposed method.更多还原