Simultaneous structure identification and fuzzy rule generation for Takagi-Sugeno models

Area of Science:

• Artificial Intelligence
• Machine Learning
• Fuzzy Logic Systems

Background:

• Interpretability of fuzzy rule-based systems diminishes with high-dimensional data.
• Identifying fuzzy rules and selecting relevant features are challenging in high-dimensional datasets.
• Existing feature selection methods often overlook nonlinear feature interactions crucial for learning systems.

Purpose of the Study:

• To propose an integrated method for Takagi-Sugeno fuzzy systems that simultaneously identifies fuzzy rules and selects features.
• To address the challenge of structure identification in high-dimensional data by considering nonlinear feature interactions.
• To develop a computationally attractive approach for feature selection and rule generation.

Main Methods:

• An integrated learning mechanism is proposed to find irrelevant features and generate fuzzy rules concurrently.
• The method accounts for nonlinear interactions between features and between features and the fuzzy rule-based system.
• The approach is non-iterative, contrasting with traditional forward or backward selection methods.

Main Results:

• The proposed method successfully identifies a small set of useful features and generates effective fuzzy rules.
• It effectively handles subtle nonlinear interactions present in the data.
• Demonstrated effectiveness on four well-studied function-approximation problems.

Conclusions:

• The integrated method offers an efficient and effective solution for structure identification in Takagi-Sugeno fuzzy systems with high-dimensional data.
• It enhances interpretability by selecting salient features and generating meaningful rules.
• The non-iterative nature makes it computationally attractive for practical applications.