Formal Descriptive Modeling for Self-verification of Fuzzy Network Systems

Abstract

The ongoing development of deep learning systems has generated novel use cases with abundant potential. However, identifying the criteria under which these use cases are developed remains an intractable problem from the point of view of verification and validation of these systems. This problem is notorious for traditional neural networks. Fuzzy-neural networks, genetic-fuzzy trees, and neuro-fuzzy inference systems are types of fuzzy networks that represent a bridge in the gap between fuzzy inference and traditional inference models popularly used in deep learning systems. In recent years, research into the development of fuzzy network systems has revealed that these possess an innate capability for verification and validation (V &V), even for non-deterministically derived systems e.g. genetic-fuzzy trees, in a way that traditional neural networks do not. This paper presents a framework referred to as formal descriptive modeling (FDM) for extracting these criteria from a fuzzy network of any shape or size and trained under any conditions. A model use case is presented in the form of V &V of a sample fuzzy network designed to administer controls to a flight sub-system for a material transfer problem with certain model requirements. The extraction and identification of internal system criteria, application of those to external human-derived design criteria, and the methodology for deriving the logical principles defining those criteria are demonstrated in reference to the sample problem.