A hebbian based rule reduction approach to neuro-fuzzy modeling
Date of Issue2008
School of Computer Engineering
The research of neuro-fuzzy modeling is divided into two branches, the precise modeling, implemented by TSK-type fuzzy rules, and the linguistic modeling, realized through Mamdani-type fuzzy rules. Linguistic neuro-fuzzy modeling is investigated in the thesis. Neuro-fuzzy system is a hybrid of neural network and fuzzy system. Many neuro-fuzzy systems have been proposed primarily for improving modeling accuracy, but lesser attention has been devoted to address the interpretability of the derived fuzzy rules. The fuzzy rules that contain redundancy and inconsistency in the rule structure post ambiguity for people to comprehend the data and make decisions. The objective of the thesis is to construct a generic linguistic neuro-fuzzy system that is capable of generating interpretable fuzzy rules while maintaining an acceptable modeling accuracy. The major contributions of the thesis are summarized as follows. A Hebbian based Rule Reduction (HeRR) neuro-fuzzy system is proposed to generate interpretable fuzzy rules. The interpretability of the rule set is significantly improved through the merger of redundant fuzzy sets and the removal of inconsistent rules. Consequently, the fuzzy membership functions can be automatically demarcated with clear semantics. Furthermore, an iterative learning scheme is proposed to strike the balance between the interpretability of the rules and the accuracy of the modeling process; this includes a tuning and a reduction phase. During the tuning phase, the fuzzy membership functions are adjusted using the Least Mean Squared (LMS) algorithm, while the fuzzy sets and rules are respectively merged and reduced through the HeRR in the reduction phase. To generalize the neuro-fuzzy system to handle the pattern classification problems, the HeRR is reformulated as a min-max neuro-fuzzy classifier. A novel attribute reduction algorithm is proposed to eliminate redundancy in the rule set using rough set theory for the knowledge reduction. The algorithm is integrated into the HeRR, termed RS-HeRR for pattern classification. The removal of inefficient input attributes not only improves the interpretability of the rules, but may also enhance the classification accuracy, due to the clarity of the reduced rule set. An extensive set of experiments has been undertaken to demonstrate the efficacy of HeRR and RS-HeRR. Most importantly, they have been applied to the problems of ICU artificial ventilation modeling and bank failure prediction. In contrast to other well-established neuro-fuzzy systems and classifiers, the proposed system is able to deliver superior modeling accuracy through interpretable fuzzy rules.
DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence