A study on a tuning method of neural network

Summary form only given. In the case of an expert system based on neural networks (NNs) the number of training cycles is as important as the recognition rate because the expert system always has to learn new data. So far high-speed processing of the NN is realized by hardware, supercomputers, or high-speed emulation boards. However, it is difficult to realize a large-scale NN. Therefore, the high speed obtained by tuning parameters is actual and expected. The parameters are not independent, so that even if the optimum values of each parameter are collected independently they probably do not give the optimum status of the NN. A tuning method is proposed which makes it possible to narrow the enormous number of combinations of parameters down to the practical ones