Exploring multi-agent reinforcement learning for unrelated parallel machine scheduling

Exploring multi-agent reinforcement learning for unrelated parallel machine scheduling

연구 분야: Artificial Intelligence

학회: The Journal of Supercomputing

Scheduling problems present significant challenges in optimization, particularly in resource allocation and production management. This study addresses the Unrelated Parallel Machine Scheduling Problem with setup times and resource constraints (UPMSR) using a Multi-Agent Reinforcement Learning (MARL) framework. We develop a reinforcement learning (RL) environment for dynamic scheduling and compare MARL with Single-Agent RL approaches through various neural network policies. Results show that Single-Agent algorithms, particularly the Maskable Proximal Policy Optimization (PPO) variant, excel in smaller-scale scenarios, balancing decision quality and computational efficiency. Multi-agent PPO exhibits scalable potential but faces challenges in cooperative learning, underscoring the complexities of coordination in distributed decision-making tasks. This work provides insights into the strengths and limitations of MARL techniques, emphasizing their adaptability to dynamic environments and the need to balance sophistication with scalability in scheduling optimization.