A hybrid neural architecture for fuzzy rules extraction

Abstract

This article proposes a hybrid neural architecture (i.e. neuro-fuzzy) capable to generate Takagi-Sugeno fuzzy rules and to adjust the membership functions and the output functions. The main idea is to apply this algorithm to function approach tasks, whose input-output relationships are the only available information. The proposed algorithm spares previous data analysis. Initially, the user only defines the number of fuzzy rules that the system should produce. This model has the convenience of not requesting the previous and empiric knowledge of the fuzzy rules structure. The proposed model is initiated with an architecture fully connected; in other words, implicit rules do not exist in the connections of the net. Instead of this, the connections between the entrance layer and the T-norms are linked by synaptic weights. These weights must be adjusted and, in a posterior stage, are partially eliminated. The training process occurs in four stages: in the first stage is adjusted the synapses and the function parameters. The second stage occurs when the quadratic error reaches acceptable values. Then, a part of the synapses introduced between the first and the second layer are eliminated. In this stage the proposed net is equaled to the ANFIS model, because the disabled synapses reveal the base rules generated automatically by the algorithm. In the third stage, the adjustment process continues from an identical way to the ANFIS. Finally, in the fourth stage, is eliminated the lower relevant rules. This process makes the base of rules more intelligible.