Toward Verification and Validation of Adaptive Aircraft Controllers

Traditional fixed-gain control has proven to be unsuccessful to deal with complex, strongly nonlinear, uncertain, and changing systems such as a damaged aircraft. Control systems with components that can adapt toward changes in the plant, e.g., using a neural network, have been actively investigated as they offer many advantages (e.g., better performance, controllability of aircraft despite of a damaged wing). However, neuro-adaptive controllers have not been used in safety-critical applications, because performance and safety guarantees cannot be provided at development time - a major prerequisite for safety certification (e.g., by the FAA or NASA). In this paper, we describe our approach toward V&V of neuro-adaptive controllers. We have developed tools which dynamically estimate the neural network performance and safety envelope, using a Bayesian approach. We discuss our V&V approach, the tool architecture and simulation experiments within NASA's IFCS (intelligent flight control system) project