The Morphetic Epsilon-Delayed Neuro-Fuzzy Network: A General Architecture for Transparent Rule-Based Decision-Making

Abstract

Neuro-fuzzy networks are a transparent function approximation technique that embodies human-readable knowledge while retaining the performance of conventional neural architectures. Their transparency may facilitate human-in-the-loop eXplainable AI and knowledge transfer. Still, despite these advantages, there remains no universally agreed upon systematic design of said systems as their construction is often application-dependent. Here, we introduce a new class of neuro-fuzzy architectures called Morphetic Epsilon-Delayed Neuro-Fuzzy Networks inspired by network morphism and the Gumbel-Max mechanism. It is the first neuro-fuzzy network to construct itself purely from gradient signals. We show that our technique is generalizable across online and offline reinforcement learning scenarios and can handle computer vision domains.

John Wesley Hostetter

Computer Science Ph.D. Student

My research interests include fuzzy logic, artificial intelligence, and machine learning.