Development of quantum-based adaptive neuro-fuzzy networks

Area of Science:

• Computational Intelligence
• Machine Learning
• Fuzzy Systems

Background:

• Adaptive neuro-fuzzy networks (ANFNs) are powerful tools for modeling complex systems.
• Existing methods for ANFN construction can be complex and data-intensive.
• There is a need for systematic approaches to automatically generate fuzzy rules for ANFNs.

Purpose of the Study:

• To develop a novel method for constructing quantum-based adaptive neuro-fuzzy networks (QANFNs) with a Takagi-Sugeno-Kang (TSK) fuzzy type.
• To introduce fuzzy subtractive quantum clustering for automatic fuzzy rule generation.
• To propose a linear-regression QANFN (LR-QANFN) for incremental modeling of localized nonlinearities.

Main Methods:

• Fuzzy granulation of input-output data sets.
• Fuzzy subtractive quantum clustering for prototype identification and rule generation.
• Construction of QANFNs and LR-QANFNs incorporating linear regression and local fuzzy if-then rules.

Main Results:

• The proposed QANFN and LR-QANFN models achieved better performance compared to radial basis function networks and previous linguistic models.
• Effective modeling of localized nonlinearities was achieved through the incremental approach of LR-QANFN.
• Successful application demonstrated on automobile MPG prediction, Boston Housing data, and a water purification process.

Conclusions:

• The developed fuzzy subtractive quantum clustering provides an effective method for automatic fuzzy rule generation in QANFNs.
• QANFNs and LR-QANFNs offer a robust and efficient approach for complex system modeling.
• The proposed models show significant potential for real-world applications requiring accurate prediction and control.