Neurofuzzy construction algorithms s.n.
Full text not available from this repository.
Neurofuzzy systems have been developed as grey box modelling technique ideal for the task of system identification. Neurofuzzy models combine the mathematical structure of Associative Memory Networks (AMNs) with the transparency of fuzzy systems. This produces a modelling technique to which mathematical analysis can be applied, while being more transparent than traditional black box models. Unfortunately, conventional neurofuzzy models suffer from the curse of dimensionality, where the size and complexity of the model grows exponentially with the input dimension. This curse restricts the use of neurofuzzy systems to low dimensional problems. Parsimony is an important issue in modelling where the best models are obtained using the simplest possible structure. This principle can often be employed to alleviate the curse of dimensionality, where an alternative more representative neurofuzzy system can be employed. Several alternative neurofuzzy systems that try to exploit structural dependencies are presented in this report. Another problem, common to most modelling techniques, is model identification, i.e. finding the appropriate model structure (and representation) that best performs the modelling task. This report presents some algorithms that construct parsimonious neurofuzzy systems off-line from a set of system observations. The aim is to construct a parsimonious neurofuzzy model, hence alleviating the curse of dimensionality and providing a qualative insight into the physical behaviour of the system in the form of fuzzy rules. These algorithms are all based on the general iterative identification algorithm, COSMOS, also presented in this report.
Actions (login required)