Deterministic Annealing Integrated with epsilon-Insensitive Learning in Neuro-fuzzy Systems

In this paper a new method of parameters estimation for neuro-fuzzy system with parameterized consequents is presented. The novelty of the learning algorithm consists of an application of the deterministic annealing method integrated with e-insensitive learning. This method allows to improve neuro-fuzzy modeling quality in the sense of an increase in generalization ability and outliers robustness. To demonstrate performance of the proposed procedure two numerical experiments concerning benchmark problems of prediction and identification are given.