What this is
This is the methodological core that holds the other two directions together. ORAILIX treats machine learning and operations research not as competing paradigms but as complementary layers of a decision stack: the former for pattern extraction, the latter for guaranteed-feasibility optimisation. The work investigates how to compose them without losing the strengths of either.
Why it matters
The industrial systems ORAILIX works on, rail networks, aircraft fleets, financial decision pipelines, are too high-dimensional for classical OR to model exhaustively, and too safety-critical for black-box ML to be trusted alone. Hybrid AI/OR provides a disciplined way to split responsibility: ML handles the parts where the structure is not known a priori; OR handles the parts where feasibility, optimality or regulatory compliance must be provable.
Done well, this gives both performance and auditability, exactly what the trustworthy AI and AI for mobility directions require to translate into practice.
Representative work
- ORAILIX presents hybrid AI/OR approaches at Google. Paris, 2024.
- PACT, Pruning and Clustering-Based Token Reduction for Visual Language Models (CVPR 2025).
- Strategic Deployment of Cloud-RAN Architectures. A problem that neither ML nor OR solves well alone.
- Open RL turbofan degradation challenge. Reproducible hybrid-AI/OR environment, lowering the entry barrier to the field.
Partners
Cross-cutting across both industrial chairs, with additional partnerships through IRT SystemX (JNI3), Safran Tech, SNCF, and informal collaboration with Google research. Connects ORAILIX to the Operations Research community at LIX and to international venues (ISCO, CPAIOR, MIMAR, CVPR on the ML side).
Related news
Follow related dispatches in the news stream. This covers every category.
← Back to ORAILIX