一种新的CMAC函数逼近器及其再励学习方法

针对复杂再励学习系统状态空间存在维数灾问题，结合多移动机器人协调避障路径规划实际应用，用非均匀模糊分割方法将状态空间分解成模糊子空间，相应地将小脑模型连接控制器网络（Cerebellar Model Articulation Controller,CMAC)函数逼近器改进为模糊CMAC（Fuzzy CMAC,FCMAC)函数逼近器，并将FCMAC函数逼近器置入滞后更新多步Q（Pstphoned-Updating Multi-Stp Q-learning,PUMSQ)学习笮算法，提出FCMAC－PUMSQ学习算法，仿真实验证明，该算法有效且有较好的鲁棒性，FCMAC函数逼近器有较好的收敛速度和泛化能力。

A New Cerebellar Model Articulation Controller Function Approximator and Reinforcement Learning Algorithm

As there is the curse of dimension in a complex reinforcement learning system, the state space was divided into uneven and fuzzy sub spaces with the actual problem of conflict free cooperative mobile robots system. Accordingly, the cerebellar model articulation controller (CMAC) function approximator was developed into fuzzy CMAC (FCMAC) function approximator. Then a new reinforcement learning algorithm named FCMAC PUMSQ which combines FCMAC function approximator with postponed updating multi step Q learning was presented. The simulation results show that the algorithm is effective and has good robust. The FCMAC function approximator has good generalization ability and convergent speed.