A validation approach for neural network-based online adaptive systems

摘要

Traditional software validation methods cannot guarantee the safe behavior of online self-adaptive systems. These systems are characterized by continual adaptation to changing environmental conditions. We present a novel methodology for validating adaptive software systems based on online operational monitoring. The methodology inherits its theoretical underpinnings from the generic stability and convergence analysis of Lyapunov's theory. Lyapunov theory is well established in mathematics and control theory, but has not been used before for software validation. The presented validation technique is applied to a neural network-based online self-adaptive system, the intelligent flight control system. In this application, environmental changes include system failure modes, such as a stuck stabilator, broken aileron and/or rudder, sensor failure, etc. Our case study for validation is a specific online self-adaptive system, the intelligent flight control that utilizes dynamic cell structures' neural networks to perform online adaptation. The theoretical foundation and practicability of the presented validation technique make the presented validation approach generally applicable to other types of online adaptive systems.