RL for Combinatorial Optimization

This field uses reinforcement learning (RL) to build or improve solutions to problems such as vehicle routing, job-shop scheduling, packing, and graph design, where choices interact and the search space grows rapidly. Common approaches train policies with PPO or policy gradients using graph neural networks or Transformers, and increasingly combine RL with local search, evolutionary methods, exact optimization, or tree search. Research also focuses on dynamic settings, changing constraints, and generalizing from small training instances to larger or unfamiliar problems. A central tension is whether learned methods can match strong specialized heuristics and exact solvers in solution quality while offering faster decisions and reliable feasibility at scale.