Lifelong Self-Adaptation: Dealing with Drift in Machine-Learning-Supported Self-Adaptation

As software systems become more complex and must operate continuously in the face of uncertainty, it becomes essential to integrate techniques for managing changes and uncertainty during their operation. One of the most common approaches is to use \emph{external mechanisms} based on the feedback loop concept, which is the basis of understanding self-adaptive software systems. Machine learning techniques have been recently employed to support the feedback loop functions, resulting in a new type of adaptation known as machine-learning-supported self-adaptation. However, these classical machine learning methods are essentially designed to deal with a predefined set of learning tasks, making it difficult to deal with new emerging learning tasks, changes, and new data distribution shifts (concept drift) without human intervention.

The primary research problem of this dissertation is how to enable machine-learning-supported self-adaptive systems to deal with concept drift during operation. Two sub-problems are derived from this primary problem, namely, to deal with covariate drift and novel class appearance drift in machine-learning-supported self-adaptive systems. An additional sub-problem is to study the impact of applying machine learning methods on the guarantees of the system.

This dissertation makes four contributions. The first contribution is a systematic literature review that identifies the primary issue of concept drift and the need for unsupervised learning methods to detect novelties at runtime. The second contribution is a general architecture called \emph{lifelong self-adaptation} that leverages the knowledge of lifelong machine learning to tackle the problem of concept drift. This architecture was instantiated to deal with covariate drift and evaluated in two different application domains. The third contribution is the instantiation of the general architecture to address novel class appearance drift and evaluated in an internet-of-things domain. The fourth contribution investigates the impact of applying machine learning methods using computational learning theory on the analysis results, particularly those obtained using statistical model checking to support decision-making for self-adaptation.