Safe, intelligent and explainable self-adaptive systems ; Sichere, intelligente und erklärbare selbst-adaptive Systeme

Intelligent cyber-physical systems, such as self-driving cars, smart homes or e-health solutions, will increasingly influence our daily lives. They will deal with increasingly uncertain and changing environments and simultaneously must adhere to strict safety requirements. In addition, we need to trust those systems, as we will hand over control on our daily life to them. This increasing list of non-functional requirements makes the design of intelligent cyber-physical systems (CPS) a challenging task. In this thesis, we advance research on intelligent CPS by introducing intelligent self-adaptive systems that autonomously evolve their adaptation logic in response to changes in the system structure, their environment and their goals. We make their widespread integration possible, by introducing a safe design framework, based on a novel methodology. Our framework enables the integrated design and formal verification of intelligent and explainable self-adaptive systems. Our key idea is to combine a resource-efficient process for self-adaptation with dynamic evolution of the adaptation logics and continuous verification activities. To obtain trust in those systems, we additionally collect structured runtime knowledge to build an explanation basis for autonomous decisions. Our main contributions are an efficient and comprehensible rule- and distance-based adaptation process, a quantitative and context-dependent goal model that provides the basis for our adaptation process, a resource-efficient run-time evolution of adaptation logics that combines a continuous evaluation and observation-based optimization, of adaptation rules and a stochastic search-based learning of new comprehensible adaptation rules, and a continuous verification methodology that is based on a formalization of our rule- and distance-based adaptation process in timed automata. We have implemented our framework and evaluated its applicability and performance on three case studies from different domains, namely a smart temperature control system, an ...