Quality Management for AI-Generated Self-Adaptive Resource Controllers

Abstract

Many complex systems requires the use of controllers to allow an automated, self-adaptive management of components and resources. Controllers are software components that observe a system, analyse its quality, and recommend and enact decisions to maintain or improve quality. While controllers have been for many years, recently Artificial Intelligence (AI) techniques such as Machine Learning (ML) and specifically reinforcement learning (RL) are used to construct these controllers, causing uncertainties about the quality of them due to their construction. We investigate quality metrics for RL-constructed software-based controllers that allow for their continuous quality control, which is particularly motivated by increasing automation and also the usage of artificial intelligence and control theoretic solutions for controller construction and operation. We introduce self-adaptation and control principles and define a quality-oriented controller reference architecture for controllers for self-adaptive systems. This forms the basis for the central contribution, a quality analysis metrics framework for controllers themselves.