基于广义模糊CMAC的强化学习算法

Reinforcement learning algorithm based on general fuzzifiedcerebellar model articulation controller

针对传统的小脑模型,提出了一种广义模糊小脑模型神经网络(GFCMAC)。它采用模糊隶属度函数作为接收域函数,可以获得较常规CMAC连续性强且有解析微分的复杂函数近似,具有计算量少,学习效率高等优点。研究了GFCMAC接收域函数的映射方法、隶属度函数及其参数的选取规律和学习算法。结合强化学习,提出了一种基于GFCAMC的强化学习算法,讨论了其实现过程。应用于船舶航向控制的仿真结果表明,在有各种风浪干扰下,船舶航向跟踪快且操舵动作合理,适合船舶转向控制要求。

Based on the conventional cerebellar model articulation controller (CMAC), a general fuzzified CMAC(GFCMAC) is proposed, in which the fuzzy membership functions are utilized as the receptive field functions. By using GFCMAC, the approximation of complex functions can be obtained which is more continuous than using conventional CMAC. The mapping of receptive field functions, the selection law of membership functions with their parameters and the learning algorithm are presented. A GFCMAC based reinforcement learning algorithm is discussed, its implementation is with reinforcement learning, also given. Applying the reinforcement learning algorithm in a ship steering (control,)?the?simulation?results?show?that?the?ship?course?can?be?properly?controlled?when?changeable?wind?and?wave?exist.