Abstract

At the heart of many computer network management and control tasks often lie complex optimization problems. Research over the last decades has been in pursuit for ever more accurate and faster algorithms solving such problems. More recently, we have also witnessed a significant interest in applying machine learning and artificial intelligence to tackle problems, e.g. related to routing, functions placement, or network design. In this book chapter, we review an approach that is motivated by the observation that most existing human-designed network management and control algorithms overlook a simple yet powerful optimization opportunity: many algorithms are executed repeatedly and hence with each execution, generate data: the (problem, solution)-pairs. For example, finding a path between two specific nodes describes the problem, whereas a path-finding algorithm provides the solution, namely the path. In this chapter, we provide an overview of approaches that use machine learning (ML) and artificial intelligence (AI) to learn from (problem, solution)-pairs to improve actually network algorithms. We discuss the applicability for different use cases and identify challenges within those use cases as well as limitations. We review recent trends from the ML and AI research community, and how they might be applied to specific network problems. Based on this, we present future research directions geared toward data-driven algorithms design for network management and control tasks.