Run-time assessment of neural network control systems

Abstract

A novel approach for validation of soft computing systems (neural network) embedded in safety-critical online self-adaptive systems is presented in the form of run-time risk assessment methodology. The approach is based on run-time operational monitoring of the neural network and data fusion techniques for combining outputs from various monitors. The run-time monitoring is based on the stability analysis of dynamic systems similar, in principle, to the Lyapunov analysis. The outputs from various monitors are fused together to form a single measure of confidence from the data fusion techniques of Dempster-Shafer (Murphy's rule of combination) and Fuzzy Inference System. The developed concept of run-time monitoring and data fusion can serve as a powerful tool for assessing risk in real-time and a complementary means of validating on-line self-adaptive systems in cases where traditional validation techniques fail or cannot be applied. Using the data collected from an F-15 flight simulator, heuristic evidence is provided that supports the prospects of using run-time monitoring and data fusion techniques to form a run-time risk assessment methodology. This work can be viewed as a step towards a solution to the V&V of online self-adaptive systems. This is a complex yet very important problem facing the dependability research community. It is believed that the application boundaries for adaptive and intelligent systems will widen as the underlying software/system verification and validation theory becomes better understood and derived techniques achieve a higher level of maturity. Current work on this approach consists of fine-tuning the fuzzy inference system (modifying rules, membership functions) and providing a signaling system, similar to traffic control system that can warn the pilot/aircraft validation engineers of an imminent threat due to neural network misbehavior. © 2006 Springer Science+Business Media, Inc.