Human-Readable Neuro-Fuzzy Networks from Frequent Yet Discernible Patterns in Reward-Based Environments

John Wesley Hostetter, Adittya Soukarjya Saha, Md. Mirajul Islam, Tiffany Barnes, Min ChiIJCAI 2025 · DOI

Introduction

Reinforcement Learning (RL) has achieved remarkable success in training agents for complex decision-making tasks. However, the reliance on Deep Neural Networks (DNNs) often results in "black-box" policies that lack transparency. While Neuro-Fuzzy Networks (NFNs) offer a potential solution by encoding logic into human-readable "IF-THEN" rules, they have historically struggled with rigid semantics and the "curse of dimensionality," where rule bases become too large for human interpretation.

This paper presents a novel framework for developing transparent NFNs that maintain high performance while ensuring rules remain concise and discernible.

Methodology: Self-Organizing and Refining NFNs

The proposed approach utilizes a systematic design process to build and simplify NFNs. Initially, the network undergoes self-organization using algorithms like CLIP and ECM to automatically discover fuzzy sets and representative "exemplars" from raw environmental data.

To address the issue of rigid semantics—where the definitions of linguistic terms like "Fast" or "Near" are fixed—the authors treat the NFN as an auto-encoder. By tasking the network with reconstructing original inputs from internal fuzzy representations, the model learns to dynamically adjust the shapes and positions of its fuzzy sets to better capture environmental nuances.

Dynamic adjustment of fuzzy sets during the self-organization process.

The core innovation for enhancing readability is the Frequent-Yet-Discernible (FYD) method. Inspired by social network analysis, the FYD method models the relationships between rules and conditions as a graph. It filters conditions based on two criteria:

  1. Frequency: How often a condition is activated.
  2. Discernibility: How unique a condition is relative to others (measured via graph centrality).

Redundant or rare conditions are then pruned using the "Kneedle" algorithm to identify the optimal "knee point" in the rule-base complexity.

Key Insights: Semantic Recovery and Discernibility

A critical finding is that frequency alone does not equate to importance. By combining frequency with discernibility, the FYD method identifies the "Goldilocks" zone of information that is both relevant and distinct. Furthermore, the auto-encoder refinement ensures "semantic recovery," allowing the model to optimize its internal logic while maintaining a clear correspondence to physical reality.

Learned fuzzy sets for the four state variables of the Cart Pole environment, demonstrating the human-readable output of the refined NFN.

Results and Impact

The framework was evaluated on classic control benchmarks (e.g., Cart Pole, Mountain Car) and a high-dimensional real-world dataset from an Intelligent Tutoring System (ITS).

  • Performance: In control tasks, simplified FYD-NFNs matched or exceeded the performance of standard DNNs.
  • Simplification: In the ITS domain, which features 142 input variables, the FYD method reduced the average number of conditions per rule from 142 to just 9.7—a significant leap in interpretability.
  • Educational Outcomes: Policies generated by FYD-NFNs helped students complete training 0.6 hours faster on average compared to state-of-the-art rule extraction methods. Unlike mathematical inequalities produced by other methods, FYD rules use linguistic terms (e.g., "IF student has FEW hints..."), making them immediately actionable for educators.

Conclusion

The FYD-NFN framework successfully bridges the gap between the high performance of deep learning and the interpretability of fuzzy logic. By prioritizing frequent and discernible patterns, the method produces transparent policies that are robust enough for real-world deployment in sensitive areas like education. Future work will explore federated learning applications and real-time user explanations.

Summary updated

BibTeX
@inproceedings{hostetter2025human,
  author = {John Wesley Hostetter and Adittya Soukarjya Saha and Md. Mirajul Islam and Tiffany Barnes and Min Chi},
  title = {{Human-Readable Neuro-Fuzzy Networks from Frequent Yet Discernible Patterns in Reward-Based Environments}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/505},
}
Figure