Machine Learning for Optimization

Optimization problems are ubiquitous in many fields such as robotics, finance, logistics, transportation and planning. In real-world applications, optimization problems are generally hard, non-linear, discrete and/or large scale. State-of-the-art methods rely on human expertise to design specialized heuristics for different generic classes of problems and settings. They cannot leverage the specifics of the distribution of problem instances as they are repeatedly encountered in practice in a given context, although that information is usually easy to collect and made available for processing.

We are interested in data-driven optimization, and in particular learning underlying patterns in the problem data to be exploited in the optimization process. Our research topics are at the intersection of machine learning and (stochastic) optimization, spanning end-to-end learning, heuristic learning and hybrid methods that combine learning components and classical approaches. While solutions produced by ad-hoc solvers may be used as supervision signals, this often comes at a cost. We are interested in reinforcement learning techniques as a powerful framework for learning without relying on the availability of solved instances.