基于模糊神经网络局部强化学习在Robocup中的应用

针对Robocup仿真组比赛中智能体的配合与动作选取,将模糊神经网络（FNN）和局部协调图动态角色分配与传统Q-学习相结合,提出了基于模糊神经网络的局部Q-学习。采用该方法,有效抑制了仿真平台中的噪声干扰,提高了动作选取的精度,解决了传统Q-学习中Q表占用内存空间过大的问题,增强了系统的泛化能力,并进一步缩短了学习时间,更好的满足比赛实时性的要求。将其运用于仿真组比赛的传球和射门模型中,验证了该方法的有效性。

Application of Local Reinforcement Learning Based on Fuzzy Neural Network to Robocup

For the agents' collaboration and actions selection in the Robocup Simulation Team, the method, called Local Q-Learning based on Fuzzy Neural Network, which combined the Fuzzy Neural Network and the dynamic role assignment in a local Co-operation Graphic, was proposed. With this method, the noise interference of the simulation terrace was well restrained, and the action selection precision was improved much. The problem that the Q-table of the traditional Q-learning occupied too much memory was well done. Moreover, this method also improved the generalization capability of the system, decreased the time cost to learn, and met the real-time character of the competition better. It was applied to the experiment of the pass and shoot models in the simulation team, whose result shows the validity of the method.