基于信度分配的并行集成CMAC及其在建模中的应用

Albus CMAC(cerebella model articulation controller) 神经网络是一种模拟人类小脑学习结构的小脑模型关节控制器, 它具有很强的记忆与输出泛化能力, 但对于在线学习来说, Albus CMAC仍难满足快速性的要求. 本文在常规CMAC神经网络的基础上, 针对其在学习精度与存储容量之间的矛盾, 引入信度分配概念, 提出了一种基于信度分配的并行集成CMAC. 它将大规模网络切割为多个子网络分别训练后再组合, 大大地提高了计算效率. 通过对复杂非线性函数建模的仿真研究表明, 该方案提高了系统建模的泛化能力和算法的收敛速度. 文章最后讨论了学习常数和泛化参数对该神经网络在线学习效果的影响.

Credit-assignment-based parallel ensemble CMAC and its applications in modeling

Albus CMAC (cerebella-model-articulation-controller) is a neural network that simulates the structure of the human cerebella and performs the articulation controller. Although it has a large memory capability and is capable of output generalization, Albus CMAC is still hard to meet the requirements of rapidity for online learning. To solve the conflict between the accuracy and memory capability of Albus CMAC, we introduce the concept of credit assignment and propose the parallel ensemble CMAC based on credit assignment. A large-scale network is separated into several subnetworks; these sub-networks are trained synchronously, and then are combined. It greatly improves the computational efficiency. In simulating the model of the complex nonlinear function, results show that the proposed scheme improves the generalization capability of the system model and raises the convergence rate of the improved arithmetic. Finally, how the learning parameter and the generalized parameter influence the effect of online learning of this neural network is discussed.