Reinforcement Learning and Policy Optimization for Combinatorial Problems

I study how modern policy-optimization algorithms can train neural solvers end-to-end without labelled optimal solutions. This includes policy-gradient and actor–critic methods, REINFORCE with greedy and rollout baselines, and recent algorithms such as GRPO, PKPO, RSPO, and PPO. My goal is to make training more stable, sample-efficient, and better aligned with the true optimization objective of the underlying combinatorial problem.