Efficient Approach for RLS Type Learning in TSK Neural Fuzzy Systems.

Area of Science:

• Computational intelligence
• Adaptive systems
• Fuzzy logic systems

Background:

• Recursive Least Square (RLS) algorithms are used for training neural fuzzy systems.
• Reduced covariance matrices in RLS decrease computational load but can degrade performance.
• Existing methods face challenges in balancing computational efficiency and accuracy.

Purpose of the Study:

• To propose an enhanced local learning concept for Takagi-Sugeno-Kang neural fuzzy systems.
• To improve the performance and learning speed of RLS-based neural fuzzy systems.
• To address the limitations of reduced covariance matrices in RLS.

Main Methods:

• Implementing an enhanced local learning concept with a firing threshold for rule updates.
• Utilizing a large backpropagation learning constant to accelerate structure learning.
• Comparing the proposed method against RLS with full and reduced covariance matrices.

Main Results:

• The proposed enhanced local learning method achieves superior performance compared to using a full covariance matrix.
• The method effectively stops learning for less fired rules, reducing system disturbances.
• Significant improvements in the learning speed during the structure learning phase were observed.

Conclusions:

• The enhanced local learning approach offers a more effective way to train Takagi-Sugeno-Kang neural fuzzy systems.
• Selective rule updating enhances both performance and computational efficiency.
• This method provides a robust solution for self-constructing neural fuzzy inference networks.