Abstract

Empirical Model Learning (EML) is a technique to enable Combinatorial Optimization and decision making over complex real-world systems. The approach is based on a two-fold mechanism: 1) using a Machine Learning (ML) model to approximate the input/output behavior of a system, and 2) embedding such Empirical Model into a Combinatorial Optimization model.

The EML approach has been employed with a measure of success to the application of Combinatorial Optimization to systems that are too complicated for an expert-designed, hand-crafted model, and to the generation of adversarial examples and certification of ML models. Specic use cases include: thermal-aware workload dispatching, transprecision computing, hardware dimensioning and algorithm configuration, epidemiological model, and NN verification.

However, the method has potentially much broader applicability, such as providing an alternative approach to deal with uncertainty in optimization, enabling the definition of hierarchies of optimization systems (each one approximated via ML), black-box optimization, and parameter tuning. Research in these direction has been so far constrained by limited resource and by some notable, open, scientific problems.

The goal of the workshop will be to present the expertise accumulated at UniBo on EML topics, highlight outstanding issue, promising research directions, and defining concrete steps for cooperation and advancement

Program

10:00 - 10:10: Welcome + Introduction (Luc de Raedt)

10:10 - 10:55: Talk (Michela Milano, Michele Lombardi, Andrea Borghesi)

• Group presentation

• Empirical Model Learning (the problem, application/success stories)

• Open Issues

10:55-11:00 - Break

11:00-11:45 - Proposals for concrete ideas around the task

11:45-12:30 - Follow-up discussion & Collaboration definition