Anomaly-Aware Adaptation Approach for Self-Adaptive Cyber-Physical System of Systems Using Reinforcement Learning

Abstract

A cyber-physical system of systems (CPSoS) is a system composed of multiple constituent systems that interact with both physical and cyber environments. Self-adaptivity is essential for CPSoS because it works on both cyber and physical uncertainties in various environments. Main obstacles to achieving self-adaptive CPSoS are time constraints and system anomalies. An adaptation should be processed within a certain period and it should consider anomalies caused by system changes due to mechanical faults, cyber-attacks, or emergent behaviors. However, since existing adaptation approaches cannot fully handle both aspects, this paper proposes an advanced approach, A4, for a self-adaptive system that can handle known anomalies in runtime. This approach learns the known anomalies before runtime and mitigates their impact when they are detected. We evaluated the A4 approach for virtual and physical CPSoS and showed that A4 was more efficient than other approaches.