Adrian BREZULIANU
One-Step And Multiple Steps Prediction Using Adaptive Sugeno Fuzzy Systems Architecture

Abstract.
Last years, fuzzy systems have developed a widespread use in signal processing and control [1], [2], [3], [4], [32], [33], [34] including fuzzy logic filters, fuzzy systems predictions, and numerous hybrid approaches [5], [6], [7], [8], [9], [10].
The paper presents the predictive capabilities of the Sugeno Fuzzy Systems Architecture, the composed fuzzy systems being adapted by a genetic algorithm. The performances of this architecture are tested by the prediction quality for a chaotic benchmark signal. Chaotic signal prediction tests were performed for two optimization scenarios: a classical fitness function approach and a modified fitness function with specific penalty terms. An improvement of the generalization performance is achieved by modifying the fitness function of the genetic algorithm in accordance with the auto-correlation model validity criterion. This modification of the fitness function unifies both the model optimization and model validation in the same process, eliminating the cases of adapting results that not verify the model criteria. The prediction performances were tested for one-step and for multiple steps predictions. The second case focuses on improving the generalization results by using the special penalty term on genetic algorithm.

Keywords: Chaotic signal prediction, Sugeno fuzzy systems and genetic algorithm optimization.