Marjorie Darrah and Brian J. Taylor, Institute for Scientific Research, Inc.; Michael Webb, Lockheed Martin Technology Services; and Rhett Livingston, ProLogic, Inc.
This paper describes a geometric algorithm to extract deterministic rules from a dynamic cell structure (DCS) neural network and the rationale for extracting these rules. The DCS is a type of self-organizing map neural network that has been used in a real-time adaptive flight control application. The purpose for extracting rules in this instance is to determine whether such rules, along with other techniques, could be used in the verification and validation (V&V) of a neural network serving in a safety-critical role. This paper introduces a geometric approach to creating rules that mimic an instance of the trained DCS with 100% agreement. The paper will explain the intelligent flight control application of the DCS, describe the geometric method used for rule extraction, provide experimental results of the rule extraction techniques, and examine the relevance of the rules to the V&V process.