Towards Achieving Homeostasis in Data-Driven Production Systems through Automated Continual Learning

Maintaining reliable and adaptive systems is essential in the evolving field of Machine Learning for operational scenarios. Traditional learning models often struggle to adapt to non-stationary real environments where data distributions evolve over time. This leads to potential degradation in model performance, which can result in not meeting system requirements and, in the worst case, software failures. The appropriate monitoring and maintenance of such complex systems often require tools, practices, and strategies that are predominantly time-consuming and human-demanding. The core of this thesis concerns the introduction of a novel computational framework termed Automated Continual Learning (AutoCL), aims to address such challenges by integrating principles from continual learning and automated machine learning to enable homeostasis in artificial systems. Drawing inspiration from biological organisms that maintain internal stability despite external changes, this novel framework seeks to bridge the gap in current complex system infrastructures. Furthermore, considering the holistic integration of automated machine learning with continual learning solutions remains insufficiently explored in the scientific literature, we aim to contribute new knowledge for further scientific investigation.