A data driven approach for process modelling and decision making : RBF neural networks vs. linear models

Abstract

Interest in using data mining techniques and in particular neural network models has led to a tremendous surge in research activities in the past decade. A number of researchers have been involved in developing neural network models for various purposes. They have been viewed as black boxes, where the inputs were known and the outputs were calculated, but the reliability of the networks have not been fully understood. There are several factors that affect the performance of neural network models and for their successful application these factors have to be carefully considered. The most critical ones are the dimension of input layer, the number of hidden layer neurons and the size of the training data set. The lack of systematic evaluation of these techniques caused that the results to date are mixed and contradictory.Furthermore, almost all research to date involves multilayer perceptron neural network (MLPNN) models but none of the studies provide a comprehensive performance comparison of other neural network techniques (such as radial basis function neural networks - RBF NNs)with traditional linear methods for modelling different real-life problems. Therefore, there is a need for a thorough comparative study over the range of different data sets (i.e. processes) to address the key issues in the application of these relatively new techniques.In order to foster this process and to help clarify the current state of these modelling techniques, this research systematically investigates the application of RBF NN models for different prediction tasks and compare their performance with traditionally used linear methods. The effects of various factors, such as the dimension of the input data set, the number of hidden layer neurons, the level of noise in the data and the training and test sample size, have been systematically investigated within the experimental work. Also, the effectiveness of the direct and iterative forecast strategies was examined in relation to the performance of neural networks and linear models. The data sets were chosen to reflect the real-life processes and to span as many of a desired group of attributes as possible (static,dynamic, nonlinear, noisy, low dimensional, high dimensional, stochastic, etc.) given the size limitation. A novel hybrid approach for modelling flows within the pipeline networks is also proposed. In this approach, RBF NN is used as a routine for error-correction of deterministic models. The results indicate that the developed hybrid modelling approach can be very powerful tool for hydraulic modelling and process control within the wastewater pipe networks. This research has shown clearly that the advantage of using RBF NN models over traditionally used linear models is their superiority to model nonlinear characteristics of various processes. Therefore, such models are capable to assist users in constructing more accurate and more robust tools than what traditional linear models are able to offer