Inference error minimisation: Fuzzy modelling of ambiguous functions (Q5947596)

The problem of modelling ambiguous (multi-valued) mappings from noisy experimental data and the use of these models in control strategies are considered. The use of consistent fuzzy rulebases as function approximators in modelling and control is discussed via an equation representation of the rulebases and a general subsethood interpretation, in which rules are understood as examples of a function to be modelled, obtained by applying that function to some fuzzy sets. The concepts of proper rules and antecedents formalize some intuitive notions. These concepts can be generalized to multivalued mappings. Different learning algorithms based on the extension principle and error functions have been presented to fit noisy experimental data to a fuzzy system with prescribed antecedents. Algorithms to select the best antecedent-consequent pairings for given sets are also discussed. A feedback linearisation control application based on consistent fuzzy model inversion of a multivalued mapping has been reported. The methods presented here pave the way to automatic supervision of knowledge acquisition: detection of contradictions, rulebase simplification (elimination of redundant rules) and preservation of readability, providing tools for fusion of rulebases from different experts and learning approaches.