GMLaS

Current research areas in the convergence of Combinatorial Optimization (CO) and ML are emerging as a game-changing force in the realm of AI. Such fusion of two distinct facets of the solving paradigm has recently gained traction in the CO field, becoming a ubiquitous technology across a wide range of applications. One area that can significantly benefit from the use of ML is constraint reasoning and optimization. Constraint Programming (CP) is considered as one of the foremost paradigms for solving combinatorial problems in AI. CP technology offers a generic way of modeling and solving combinatorial problems from various domains. The hybridization of Data Mining, Machine, and Deep Learning with state-of-the-art discrete reasoning and optimization is one of the major challenges in Artificial Intelligence. Our proposal explores the challenge of data-driven decision-making through a combined machine learning (ML), discrete reasoning/optimization framework, and associated solver: Graphical Models (GMs) and the toulbar2 solver (winner of the UAI 2022 solver competition on two discrete optimization/counting tasks). Discrete GMs, especially additive GMs such as Cost Function Networks (CFNs) and Markov Random Fields are attractive because they can represent logical and probabilistic information seamlessly. Then, toulbar2 is a SOTA solver for reasoning on such discrete models. It includes both exact solvers and sophisticated meta-heuristics. GMLaS is a significant step toward the ambitious ultimate goal of integrating ML/DM tools to Cost Function Network paradigm, leading to a hybrid data-driven solving framework, dealing natively with optimization problems and with the ambition to solve larger and more complex real problems. One example of problem the project will focus on is computational protein design (CPD), an area in which one of the project partners has deep expertise.