Comprehensible Decisions in Complex Self-Adaptive Systems

Abstract

To cope with uncertain and statically unforeseen environment behaviour of complex systems, self-adaptivity has gained wide acceptance. While adaptation decisions are required to be close to optimal decisions, they at the same time should be efficient, comprehensible, and reusable. To achieve this, we have developed an engineering and analysis approach for self-learning self-adaptive systems based on our notion of timed adaptation rules. Through continuous evaluation and learning, inaccurate rules can be improved and new rules can be learned at run-time to cope with changing environments and system goals. A separate verification phase enables us to provide offline and online guarantees of evolving adaptation logics based on human-comprehensible formal models. Our approach, which incorporates the precise retracing of previous adaptation decisions, enables the understanding of the contexts in which certain adaptation decisions have been made, and assessing whether they have gained their expected effect in time within the system. This comprehensibility of complex decisions in self-adaptive systems enables the precise understanding and reuse of adaptation logics and provides trust in autonomous decision making.