Benchmarking of Multi-agent Reinforcement Learning Strategies for Optimizing Cutting Plane Selection

Cutting planes refine the feasible region of relaxed Integer Programming (IP) problems, but traditional methods relying on static heuristics often fail to generalize effectively. This study investigates Multi-Agent Reinforcement Learning (MARL) frameworks – TD3-TD3, PPO-PPO, and TD3-PPO – for dynamic, adaptive cut selection. MARL improves scalability and exploration by distributing decision-making across specialized agents, outperforming conventional techniques. The hybrid TD3-PPO configuration balances TD3’s sample-efficient learning with PPO’s robust exploration. PPO-PPO demonstrates superior exploration and success rates in a sensor network design problem relevant to process control, while TD3-TD3 offers greater stability. The results highlight MARL’s potential for enhancing IP solvers and solving complex optimization problems.