CMAC階層式多解析度量化及誤差修正之研究

[[abstract]]The Cerebellar Model Articulation Controller (CMAC) is a lattice-based Associative Memory Network (AMN), which applies table look-up technique to address memory and quickly generate correct outputs for given inputs. Owing to its fast learning property, good local generalization capability, and easy of implementation by hardware. CMAC has been applied in many application such as robotic control, data recognition, pattern classification, and signal process. The input space quantization method determines the degree of approximation accuracy and the efficiency of memory utilization directly. However, the conventional CMAC performs an equal-size method for the input space quantization without considering for the various distributions of training data sets. In the paper, a hierarchical multi-resolution quantization of CMAC is presented, the concept consists in applying a Shannon's Entropy Measure for analyzing the distributions of training data sets and the optimum quantization division of input space. The approach will be applied on a second-order sliding mode control system, from experimental results show that our proposed method is able to achieve better learning accuracy, convergence speed, and memory utilization efficiency compared to the conventional CMAC.