A TSK Neuro-Fuzzy Approach for Modeling Highly Dynamic Systems

This paper introduces a new type of TSK-based neuro-fuzzy approach and its application to modeling highly dynamic systems. In details, our proposal performs an adaptive supervised learning on a collection of time series in order to create a so-called Timed Automata Based Fuzzy Controller, i.e. an evolvable fuzzy controller whose dynamic features yield high performances in variable structure systems representation. The adaptive learning is accomplished by merging together theories from the area of times series analysis such as the Adaptive Piecewise Constant Approximation method, with a well-known neuro-fuzzy framework, the Adaptive Neuro Fuzzy Inference System. As will be shown in our experiments, where our proposal has been tested on a Fuzz-IEEE 2011 Fuzzy Competition dataset, this approach reduces the output error measure and achieves a better performance than a standard application of the ANFIS algorithm when applied to highly dynamic systems.