Intelligent data modelling using neurofuzzy algorithms

Often the quality of the available numerical and linguistic knowledge conventionally used to identify neurofuzzy systems is poor. This problem is overcome by the use of advanced model identification algorithms presented in this paper. Parsimonious models are identified via data-driven construction algorithms which match the structure to the data and allow the application of neurofuzzy modelling to high-dimensional real world problems. However, the inherent structure of neurofuzzy models can produce redundant degrees of freedom which are poorly identified by the data. As a solution to this problem Bayesian regularisation is applied to these models, smoothing out any irregularities in the structure, hence controlling unidentified rules. This is important in control and system identification scenarios where data may only be gathered around a collection of operating points