A Study On Modular Neural Network For Identification And Regression

Modular neural network, a new connectionism model proposed in the past few years, consists of a group of neural networks whose decisions are combined in a fashion of competition and cooperation to improve the performance of the whole system. A great deal of successful applications demonstrate that modular neural network outperforms single neural network in terms of generalization and reliability and undoubtedly provides for us with a new tool for problems-solving. Moreover, various theoretical explanations proposed recently justify the effectiveness of some conventional methods for modular neural network. Having recognized that modular neural network has an enormous potential and bright prospect in application, a large number of researchers plunge themselves into the field, yielding a lot of relevant theories and application achievements. At present modular neural network is a very hot topic in many diverse areas such as pattern recognition, control, decision-making, machine learning, information processing.The research for modular neural network concentrates on two aspects: how to combine the decisions of the component networks, and how to generate the component networks in the whole system. In this paper, the first aspect is mainly studied in the context of regression problems. Two methods for integrating the component networks are proposed and then tested by using several artificial problems, which make some meaningful conclusions.This paper composes five chapters in all. Chapter 1 briefly introduces the importance of modular neural network and consequently gives a simple review from seven viewpoints, i.e., concept, motivation, classification, task decomposition, design, learning. In Chapter 2 a new combination method for modular neural network is proposed. It is a dynamic combination method, for the networks composed the overall system and their corresponding combination weights vary with the changes of the input pattern. Here it is introduced a selective mechanism-only a part of or all the trained component networks are selected according to a certain rule to make up of an entire system, given an input pattern. A modified Bayesian learning method for modular neural network is proposed in Chapter 3. In Chapter 4 we present another combination method for modular neural network-sequential Bayesian method. Finally in Chapter 5, contributions of this paper are summarized and several issues for future work are indicated.