Knowledge-based fuzzy MLP for classification and rule generation

Area of Science:

• Artificial Intelligence
• Machine Learning
• Pattern Recognition

Background:

• Traditional multilayer perceptrons (MLPs) often lack mechanisms for incorporating prior knowledge, potentially limiting their classification performance and interpretability.
• Fuzzy logic approaches enhance MLPs but can be further optimized by integrating domain-specific knowledge.
• Rule generation from neural networks aids in understanding decision-making processes.

Purpose of the Study:

• To propose a novel knowledge-based classification and rule generation scheme using a fuzzy multilayer perceptron (MLP).
• To enhance the learning speed and classification performance of fuzzy MLPs by encoding prior knowledge.
• To enable the generation of interpretable rules, including negative rules for ambiguous cases.

Main Methods:

• Encoding domain knowledge into connection weights as class a priori probabilities.
• Incorporating hidden nodes representing pattern classes and their complementary regions.
• Refining network architecture (nodes and links) during training via node growing and link pruning.
• Generating classification rules from the trained network using input, output, and connection weights.

Main Results:

• The proposed knowledge-based fuzzy MLP demonstrated superior learning speed and classification performance compared to conventional and standard fuzzy MLPs.
• The scheme effectively handles both convex and concave decision regions.
• Generated rules, including negative rules, provide justification for classification decisions and aid in inferencing ambiguous cases.

Conclusions:

• The integration of prior knowledge into fuzzy MLPs significantly improves classification efficiency and accuracy.
• The developed method offers a robust approach for knowledge-based classification and interpretable rule generation.
• This technique holds promise for applications requiring explainable AI and high-performance pattern recognition.