Systematic Data-Driven Continual Self-Learning
Time: Tue 2023-05-09 15.00
Location: Ka-Sal C (Sven-Olof Öhrvik), Kistagången 16, Kista
Language: English
Doctoral student: Diarmuid Corcoran , Programvaruteknik och datorsystem, SCS
Opponent: Professor Steven Latré, University of Antwerp, Antwerp, Belgium
Supervisor: Professor Magnus Boman, Programvaruteknik och datorsystem, SCS
QC 20230414
Abstract
There is a lot of unexploited potential in using data-driven and self-learning methods to dramatically improve automatic decision-making and control in complex industrial systems. So far, and on a relatively small scale, these methods have demonstrated some potential to achieve performance gains for the automated tuning of complex distributed systems. However, many difficult questions and challenges remain in relation to how to design methods and organise their deployment and operation into large-scale real-world systems. For systematic and scalable integration of state-of-the-art machine learning into such systems, we propose a structured architectural approach.
To understand the essential elements of this architecture, we identify a set of foundational challenges and then derive a set of five research questions. These questions drill into the essential and complex interdependency between data streams, self-learning algorithms that never stop learning and the supporting reference and run-time architectural structures. While there is a need for traditional one-shot supervised models, pushing the technical boundaries of automating all classes of machine learning model training will require a continual approach.
To support continual learning, real-time data streams are complemented with accurate synthetic data generated for use in model training. By developing and integrating advanced simulations, models can be trained before deployment into a live system, for which system accuracy is then measured quantitatively in realistic scenarios. Reinforcement learning, exploring an action space and qualifying effective dynamic action combinations, is here employed for effective network policy learning. While single-agent and centralised model training may be appropriate in some cases, distributed multi-agent self-learning is essential in industrial scale systems, and thus such a scalable and energy-efficient approach is developed, implemented and analysed in detail.
Energy usage minimisation in software and hardware intense communication systems, such as the 5G radio access system, is an important and difficult problem in its own right. Our work has focused on energy-aware approaches to applying self-learning methods both to energy reduction applications and algorithms. Using this approach, we can demonstrate clear energy savings while at the same time improving system performance.
Perhaps most importantly, our work attempts to form an understanding of the broader industrial system issues of applying self-learning approaches at scale. Our results take some clear, formative, steps towards large-scale industrialisation of self-learning approaches in communication systems such as 5G.