Data-driven interval type-2 neural fuzzy system with high learning accuracy and improved model interpretability.

Area of Science:

• Computational Intelligence
• Machine Learning
• Fuzzy Systems

Background:

• Existing type-2 fuzzy systems (FSs) often prioritize accuracy over the interpretability of fuzzy rules.
• A gap exists in developing fuzzy models that are both accurate and easily understandable.

Purpose of the Study:

• To propose a data-driven interval type-2 fuzzy system (IT2 NFS) with enhanced model interpretability (DIT2NFS-IP).
• To develop a system that balances high accuracy with transparent fuzzy rule representation.

Main Methods:

• Utilized interval type-2 fuzzy sets in the antecedent and interval consequents for rule simplicity.
• Employed a self-splitting clustering algorithm for initial rule base generation.
• Implemented a two-phase parameter-learning algorithm involving gradient descent and recursive least squares, incorporating a novel cost function for accuracy and interpretability.

Main Results:

• The proposed DIT2NFS-IP demonstrated improved model interpretability compared to existing type-1 and type-2 FSs.
• The system achieved competitive accuracy in data-based modeling and prediction tasks.
• The two-phase learning approach effectively optimized both accuracy and rule transparency.

Conclusions:

• The DIT2NFS-IP offers a viable solution for developing accurate and interpretable fuzzy models.
• This approach advances the field of type-2 fuzzy systems by integrating interpretability as a key design objective.
• The findings are validated across multiple benchmark datasets, confirming the system's robustness.