Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Gradient-Based Neuroplastic Adaptation for Concurrent Optimization of Neuro-Fuzzy Networks

John Wesley Hostetter, Min Chi

TL;DR

The paper proposes gradient-based neuroplastic adaptation (GBNA) to concurrently optimize neuro-fuzzy networks (NFNs) in both their parameters and structure, addressing limitations of prior sequential design. By relaxing binary rule connections and employing differentiable sampling (STGE or STE), GBNA enables online, gradient-driven reconfiguration of fuzzy logic rules, while incorporating delayed neurogenesis and online adaptation to handle novel stimuli. To tackle high-dimensional vision tasks, it introduces layer normalization and alpha-entmax to promote sparsity and numerical stability in rule activations, and provides a rewrite of Takagi-Sugeno-Kang NFNs for stable computation. The approach is validated by online reinforcement learning experiments in the DOOM FPS environment, where NFNs are trained alongside CNN front-ends and show competitive performance with DNN baselines across multiple scenarios, highlighting the potential for interpretable, adaptable NFNs in vision-based RL. Overall, GBNA enables a general, application-independent path to simultaneously learn NFN structure and parameters, with practical implications for interpretable AI that can integrate with complex neural architectures and dynamic environments.

Abstract

Neuro-fuzzy networks (NFNs) are transparent, symbolic, and universal function approximations that perform as well as conventional neural architectures, but their knowledge is expressed as linguistic IF-THEN rules. Despite these advantages, their systematic design process remains a challenge. Existing work will often sequentially build NFNs by inefficiently isolating parametric and structural identification, leading to a premature commitment to brittle and subpar architecture. We propose a novel application-independent approach called gradient-based neuroplastic adaptation for the concurrent optimization of NFNs' parameters and structure. By recognizing that NFNs' parameters and structure should be optimized simultaneously as they are deeply conjoined, settings previously unapproachable for NFNs are now accessible, such as the online reinforcement learning of NFNs for vision-based tasks. The effectiveness of concurrently optimizing NFNs is empirically shown as it is trained by online reinforcement learning to proficiently play challenging scenarios from a vision-based video game called DOOM.

Gradient-Based Neuroplastic Adaptation for Concurrent Optimization of Neuro-Fuzzy Networks

TL;DR

The paper proposes gradient-based neuroplastic adaptation (GBNA) to concurrently optimize neuro-fuzzy networks (NFNs) in both their parameters and structure, addressing limitations of prior sequential design. By relaxing binary rule connections and employing differentiable sampling (STGE or STE), GBNA enables online, gradient-driven reconfiguration of fuzzy logic rules, while incorporating delayed neurogenesis and online adaptation to handle novel stimuli. To tackle high-dimensional vision tasks, it introduces layer normalization and alpha-entmax to promote sparsity and numerical stability in rule activations, and provides a rewrite of Takagi-Sugeno-Kang NFNs for stable computation. The approach is validated by online reinforcement learning experiments in the DOOM FPS environment, where NFNs are trained alongside CNN front-ends and show competitive performance with DNN baselines across multiple scenarios, highlighting the potential for interpretable, adaptable NFNs in vision-based RL. Overall, GBNA enables a general, application-independent path to simultaneously learn NFN structure and parameters, with practical implications for interpretable AI that can integrate with complex neural architectures and dynamic environments.

Abstract

Neuro-fuzzy networks (NFNs) are transparent, symbolic, and universal function approximations that perform as well as conventional neural architectures, but their knowledge is expressed as linguistic IF-THEN rules. Despite these advantages, their systematic design process remains a challenge. Existing work will often sequentially build NFNs by inefficiently isolating parametric and structural identification, leading to a premature commitment to brittle and subpar architecture. We propose a novel application-independent approach called gradient-based neuroplastic adaptation for the concurrent optimization of NFNs' parameters and structure. By recognizing that NFNs' parameters and structure should be optimized simultaneously as they are deeply conjoined, settings previously unapproachable for NFNs are now accessible, such as the online reinforcement learning of NFNs for vision-based tasks. The effectiveness of concurrently optimizing NFNs is empirically shown as it is trained by online reinforcement learning to proficiently play challenging scenarios from a vision-based video game called DOOM.

Paper Structure

This paper contains 55 sections, 15 equations, 11 figures, 5 tables.

Table of Contents

  1. Introduction
  2. First Contribution
  3. Second Contribution
  4. Third Contribution
  5. Fourth Contribution
  6. Background & Related Work
  7. Fuzzy Inference Systems
  8. Neuro-Fuzzy Networks
  9. Fuzzy Reinforcement Learning
  10. Neuro-Fuzzy Networks
  11. Describing the Input with Linguistic Terms
  12. Condition Layer
  13. Activating the Fuzzy Logic Rules
  14. Rule Layer
  15. Determining the Decisions
  16. ...and 40 more sections

Figures (11)

  • Figure 1: Attacking stationary targets with a pistol (left) and a rocket launcher (right).
  • Figure 2: Tracking an enemy and predicting their trajectory to launch an attack.
  • Figure 3: Holding the line against a horde of enemy monsters.
  • Figure 4: Handling imminent threats while balancing situational awareness.
  • Figure 5: Surviving a room with an acidic floor by gathering health.
  • ...and 6 more figures