Reinforcement learning (RL) has become a powerful paradigm for sequential decision-making, with major successes in domains such as robotics, game playing, and language modelling. In optimization, however, its impact is still uneven. While RL has shown promise in improving algorithmic components and in learning solution strategies end-to-end, consistent and transferable gains across problem classes remain difficult to achieve. Key challenges include sparse and delayed rewards, highly structured or combinatorial search spaces, expensive evaluations, and the need to generalize across diverse instance distributions.

Over the past years, research at the intersection of RL and optimization has emerged in three largely independent communities:

• Evolutionary algorithms and metaheuristics, where RL is used to control and adapt search components or even construct new algorithms;
• General-purpose solving paradigms (e.g., constraint programming, mixed-integer programming, SAT/SMT), where RL is integrated into highly engineered solving pipelines to learn branching, variable selection, restart, or cutting strategies; and
• Neural combinatorial optimization, where RL is used to learn solution construction or improvement policies directly.

Despite their different development histories and methodologies, it is very clear that these communities face a set of common challenges when applying RL to optimization problems. Among them are the design of effective representations, the stability and scalability of learning procedures, and the ability to generalize across problem families.

The goal of this Dagstuhl Seminar is to bring together leading researchers from these three domains to foster exchange, identify common principles, and accelerate progress toward a unified understanding of RL for optimization. The seminar will focus on a set of concrete topics intended to stimulate cross-domain discussion and to define a shared research agenda for RL in optimization, including:

• Shared challenges across all domains when using RL, particularly deep RL, including representation learning, stability, scalability, and generalization.
• Domain-specific difficulties, and whether solutions from one area (e.g., evolutionary computation) can be transferred to another (e.g., CP or MIP).
• Common methodologies, benchmarks, and theoretical frameworks that could unify the study of RL for optimization.
• Best practices and lessons learned, including the integration of negative results and empirical insights that rarely appear in publications but are crucial for scientific progress.

Creative Commons BY 4.0

Quentin Cappart, Nguyen Dang, Carola Doerr, and Kevin Tierney