论文部分内容阅读
为解决标准Q学习算法收敛速度慢的问题,提出一种基于多个并行小脑模型(Cerebellar Model Articulation Controller:CMAC)神经网络的强化学习方法。该方法通过对输入状态变量进行分割,在不改变状态分辨率的前提下,降低每个状态变量的量化级数,有效减少CMAC的存储空间,将之与Q学习方法相结合,其输出用于逼近状态变量的Q值,从而提高了Q学习方法的学习速度和控制精度,并实现了连续状态的泛化。将该方法用于直线倒立摆的平衡控制中,仿真结果表明了其正确性和有效性。
In order to solve the problem of slow convergence rate of standard Q learning algorithm, a reinforcement learning method based on multiple Cerebellar Model Articulation Controller (CMAC) neural networks is proposed. By dividing the input state variables, the method reduces the quantization level of each state variable without changing the state resolution, effectively reduces the storage space of the CMAC and combines it with the Q learning method. The output of the method is used for Approximates the Q value of the state variables, thereby improving the learning speed and the control precision of the Q learning method, and realizes the generalization of continuous states. The method is applied to the balance control of linear inverted pendulum. Simulation results show its correctness and validity.