Comparison of neural networks and fuzzy relational systems in dynamic modelling

Both neural networks and fuzzy relational models show great potential for modelling poorly understood and highly non-linear systems. Recent papers have suggested that both techniques could be used to form the model in model-based controller designs. Although the two techniques can be targeted at the same role, they are fundamentally different. Fuzzy relational models attempt to capture relationships between qualitative states and therefore represent the type of qualitative models used in everyday commonsense reasoning. Neural networks instead try to imitate the hardware involved in thinking and can generate their results from the complex interactions between the separate network elements. Although the two methods are quite different in conception, the authors are not aware of any work that has been done to compare their performance in dynamic modelling, and this paper is an attempt to remedy this. In order to compare the two techniques, the well known Box-Jenkins furnace data were used. The software used to generate the neural networks was 'Neural-Works Explorer', and in-house software was used to generate the fuzzy relational models. The predictive performance of both methods was compared on the dataset for a variety of model configurations. The paper describes the results of these tests and discusses the effects of changing model parameters on predictive and practical performance. For the neural models the factors investigated were: network configuration, transfer function type, and various combinations of learning schemes. The factors considered for the fuzzy relational models were the model structure and reference set definitions. As well as predictive performance criteria, practical modelling criteria such as modelling time, sensitivity, etc, were also used to compare the two methods.