ニューラルネットワークを用いた適応学習制御(<小特集>制御工学への知能科学からの接近)  [in Japanese] Adaptive and Learning Control Using Neural Networks(<Special Feature>Intelligent Systems Science Approach to Control Systems Science)  [in Japanese]

Lots of neural-net based adaptive & learning controllers have been considered for nonlinear systems. The reason why the neural network is employed for such systems is that the neural network has the capability of highly approximating nonlinear properties. However, it is pointed out that much time is required until the good control performance is obtained. On the other hand, a CMAC has been proposed by Albus. According to the CMAC, there is little time for training it, although it has an disadvantage that the accuracy of nonlinear approximation is not good. In this paper, a new design scheme of adaptive and learning controller is discussed, which is a fusion of the multilayered neural network (NN) and the CMAC. The NN effectively works in the initial stage of training, and it automatically changes from the NN to the CMAC if the learning progresses. According to this scheme, the faults in the NN and the CMAC are supplemented each other, and a good control performance can be obtained with a few training.