A framework for the analysis of adaptive systems using bayesian statistics

Abstract

Safety-critical adaptive software systems, as, for example, used in aircraft must ensure that system must remain in safe regions during adaptation in order to avoid catastrophic failures. We present a framework, which uses hierarchical statistical models and is based upon techniques from computer experiment design and active learning to characterize the boundaries between safe and unsafe regions with a minimal number of test cases. The boundaries are then represented as parametric geometric shapes that can provide easy to understand feedback to the system designer. We illustrate our framework using the NASA adaptive flight control system IFCS.