Gradient-based Optimization for Compact and Explainable Fuzzy Rule-based Classification

Javier Fumanal-Idocin, Raquel Fernandez-Peralta, Javier Andreu-Perez

Keywords: Explainable AI, Rule based Reasoning, Fuzzy logic, Gradient-based optimization, Non-differentiability

TL;DR: We propose a fuzzy neural architecture that uses gradient-based learning without rule-explosion problems.

Abstract: Rule-based models are valued in high-stakes decision-making for their transparency, but their discrete nature limits optimization and scalability. We propose the Fuzzy Rule-based Reasoner (FRR), a gradient-based rule learner that enforces user-defined complexity constraints while maintaining strong performance. FRR combines interpretable fuzzy logic partitions with sufficient (single-rule) decision-making, avoiding the combinatorial growth of additive ensembles. Across 40 datasets, FRR outperforms traditional rule-based methods (by about 5\% over RIPPER), matches the accuracy of tree-based models like CART with rule bases 90\% smaller, and achieves 96\% of the accuracy of additive rule-based models while using only 3\% of their rule base size.

Serve As Reviewer: ~Javier_Fumanal-Idocin1, ~Raquel_Fernandez-Peralta1

Submission Number: 7