GMDH-type neural networks with a feedback loop and their application to the identification of large-spatial air pollution patterns

Abstract

The GMDH (group method of data handling)-type neural networks with a feedback loop have been proposed in our early work. The architecture of these networks is generated by using the heuristic self-organization method that is the basic theory of the GMDH method. The number of hidden layers and the number of neurons in the hidden layers are determined so as to minimize the error criterion defined by Akaike's information criterion (AIC). Furthermore, the optimum neurons that can handle the complexity of the nonlinear system are selected from a variety of prototype functions, such as the sigmoid function, the radial basis function, the high order polynomial and the linear function. In this study, the GMDH-type neural networks with a feedback loop is applied to the identification of large-spatial air pollution patterns. The source-receptor matrix that represents a relationship between the multiple air pollution sources and the air pollution concentration at the multiple monitoring stations is accurately identified by using the GMDH-type neural networks with a feedback loop. The identification results of the GMDH-type neural networks are compared with those identified by other identification methods.