论文部分内容阅读
研究了一种基于变尺度编码CMAC神经网络的增强学习控制器设计方法,并应用于以自行车平衡为模型的非线性随机系统的学习控制中.该方法通过对Markov决策过程状态空间的变尺度重叠量化编码,实现基于CMAC的多分辨率值函数逼近,从而有效地提高了增强学习控制器对连续状态空间马氏决策问题的泛化性能.针对自行车学习控制的仿真研究表明,采用变尺度编码CMAC进行值函数逼近的增强学习控制器能够获得优于已有表格型方法和均匀编码CMAC方法的学习效率和泛化性能.
A design method of CMAC neural network based on enhanced learning controller is studied and applied to the learning control of a nonlinear stochastic system with a bicycle balance as a model.Through the scaling of the state space of Markov decision process Quantization coding to realize the multi-resolution value function approximation based on CMAC, which effectively enhances the generalization performance of the enhanced learning controller for Markov decision problems in continuous state space.A simulation study on bicycle learning control shows that using CMAC The enhanced learning controller which performs the approximation of the value function can obtain learning efficiency and generalization performance over the existing tabular methods and the uniform coding CMAC method.