projects | Olivier Juan

Machine Learning

PlanB&B: Model-Based Reinforcement Learning for Branch and Bound

First model-based RL agent for exact combinatorial optimization. PlanB&B learns an internal model of B&B dynamics and uses Gumbel Search (MCTS) to discover branching strategies that surpass both prior RL agents and imitation learning. AAAI 2026.

BBMDP: A Markov Decision Process for Variable Selection in Branch & Bound

A principled vanilla MDP formulation for learning optimal branching strategies in MILP solvers, unlocking k-step RL algorithms previously incompatible with the TreeMDP framework. New state-of-the-art among RL agents on the Ecole benchmark. NeurIPS 2025.

Gyozas

An open-source RL framework for MILP. Ecole-style API with SCIP 8+ support and a Gymnasium-compatible interface.

FMSTS: Reinforcement Learning for Variable Selection in Branch and Bound

First RL approach to fully optimize branching strategy in B&B from scratch. Introduces subtree size as a naturally observable Q-function, with a novel Multiplicative Dueling Architecture (MDA) for MILP variable selection.

Influence Branching for Learning to Solve MIPs Online

A graph-oriented variable selection strategy combined with Thompson sampling to learn branching heuristics online across sequences of similar MIP instances. Submitted to the 20th Mixed Integer Program Workshop computational competition.

Optimization

Apogène

EDF's short-term unit commitment software — from ground-up redesign to production deployment at scale.

DREEV: V2G Fleet Dispatch for Frequency Containment Reserve

MILP formulation and algorithm development for optimal EV fleet dispatch providing Frequency Containment Reserve (FCR) services to the French grid. Part of EDF R&D's Smart Charging P11L1 program, in collaboration with DREEV (EDF–Nuvve joint venture).

Low NO_x Configurations in an Industrial Boiler via Genetic Algorithm & CFD

Coupling a genetic algorithm with CFD simulations (Code_Saturne) to automatically discover optimal operating configurations for a 600 MW tangentially-fired pulverized-coal boiler, minimizing NOx emissions while controlling corrosion risk.