Fuzzy Knowledge-Based Prediction Through Weighted Rule Interpolation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Fuzzy Systems

Background:

• Fuzzy rule interpolation (FRI) enables approximate reasoning with sparse data, overcoming limitations of traditional methods requiring dense rule bases.
• Existing FRI methods often assume equal attribute significance or use external data for weights, limiting their applicability.
• Weighted interpolative reasoning techniques exist but are often specific to classification or lack integration with the FRI process.

Purpose of the Study:

• To present a novel weighted rule interpolation scheme for prediction tasks using only fuzzy sparse knowledge.
• To learn attribute weights from the rule base to discern attribute significance within the FRI process.
• To integrate learned attribute weights into the internal mechanism of FRI for improved prediction.

Main Methods:

• Developed a weighted rule interpolation scheme specifically for prediction tasks.
• Integrated attribute weights, learned from the rule base, into the FRI mechanism.
• Demonstrated the scheme using scale and move transformation-based FRI for prediction problems.

Main Results:

• The proposed weighted FRI scheme effectively performs prediction tasks using sparse fuzzy knowledge.
• Attribute weights learned from the rule base enhance the significance discrimination of individual attributes.
• Systematic evaluation on 12 benchmark prediction tasks showed the efficacy of the proposed approach compared to state-of-the-art FRI techniques.

Conclusions:

• The presented weighted rule interpolation scheme offers an effective method for prediction tasks with sparse fuzzy knowledge.
• Learning attribute weights directly from the rule base and integrating them into FRI improves approximate reasoning.
• The approach demonstrates significant potential for enhancing fuzzy rule-based systems in prediction applications.