Machine Learning for Combinatorial Optimisation of Partially-Specified Problems

Date

Nov 25, 2022, 3:00pm to 4:00pm

Objective (one research question):

Data-driven optimization has been the subject of intense research in recent years. Several of the proposed approaches are designed to tackle what we call “partially specified problems”, meaning problems whose structure is only partially known in symbolic form, with the remaining information being implicitly conveyed through data.

We observe how several such approaches (namely Structured Output Prediction, Decision Focused Learning, Surrogate-based Bayesian Optimization, and Empirical Model Learning can be viewed from a unifying perspective by focusing on regret.

The workshop will investigate this link, highlighting how deep it can be brought, its limitations, and opportunities arising from focusing on regret minimization rather than (e.g.) on accuracy.

Organizing partner:

UniBo

Participating partners:

UniBo, UniTN, TU Eindoven, TU Delft, KU Lueven

WP 4 Task:

T4.2

Datasets:

Some benchmarks are available, but the focus is here on formalization

Expected output:

• Progress towards a unified, regret-centric view on SOP, DFL, SBO, EML

• Improved version of an existing draft for a survey paper on the topic

• Ideas for improvements of SOP, DFL, SBO, EML, based on the unified perspective

• Promising directions for the construction of benchmarks

Bibliography:

Work https://arxiv.org/abs/2205.10157 and references therein