Elastic Fuzzy System Modeling Research

Because of its lack of important feature extraction mechanism, the classical data-driven fuzzy system extracts all features indistinguishably when trained by high dimensional data. Because lots of noise features are contained in high-dimensional training data and the classical data-driven fuzzy system has some defects in the stricture of fuzzy rules, it is facing some challenges when the classical data-driven fuzzy system is trained by high dimensional data. Firstly, the generalization performance of the classical data-driven fuzzy system is degraded when it deals with high-dimensional training data collected in the noisy environment because of the deficiency of the effective feature extraction mechanism. Secondly, all features are used by the classical data-driven fuzzy system to construct the fuzzy rules. The interpretation of classical data-driven fuzzy system is degenerated and the obtained fuzzy rule is too complex when trained by high dimensional data. Finally, the constructed fuzzy rules conclude all of the input feature space, which do not conform to the human reasoning mechanism due to the lack of individuality. The relative studies are addressed to solve the challenges faced by the classical data-driven fuzzy system when trained by high dimensional data.Firstly, an elastic fuzzy system framework is proposed to overcome the above challenges occurred by introducing the feature extraction mechanism in fuzzy systems. The elastic fuzzy system can filter the noise features in the high-dimensional data and it has strong robustness because of feature extraction mechanism. Because only important features of input data are adopted to construct the model in the elastic fuzzy system, the obtained fuzzy rules have different feature sub-spaces and more strong interpretation.Secondly, by using soft subspace clustering algorithm as feature extraction mechanism, two kinds of elastic fuzzy system modeling models are proposed based on elastic fuzzy system framework. These two elastic fuzzy system modeling models have shown the distinctive characteristics of elastic fuzzy system. The advantages of the proposed methods have been validated effectively by the experimental studies on the synthetic and real-world datasets.Finally, due to existing the equivalence between fuzzy systems and RBF neural networks under some conditions, a more robust RBF neural network construction method is proposed to deal with the high dimensional data with by introducing the feature extraction mechanism used in the elastic fuzzy system. Compared with traditional RBF neural network modeling methods, the proposed method can extract different subspace features for different nodes in the hidden layer by the adopted feature extraction mechanism. When the training data contains lots of noisy features, the proposed method still can keep good generalization performance by using the feature extraction mechanism to remove noise features. The experimental studies on the synthetic and real-world datasets show that the proposed method has strong robustness in the noisy environment.