Adaptive Multi-Agent Reinforcement Learning for Decentralized Industrial Process Automation

Modern industrial automation increasingly depends on distributed networks of autonomous agents, such as autonomous guided vehicles (AGVs) and machine controllers, to manage dynamic production tasks. Traditional centralized control systems struggle with adaptability and scalability, particularly under real-time uncertainties and high agent densities. Multi-Agent Reinforcement Learning (MARL) has shown promise for decentralized decision-making, but existing methods face challenges including non-stationarity, inefficient exploration, and limited scalability. To address these issues, we propose an Adaptive MARL framework that integrates centralized training with decentralized execution (CTDE), adaptive exploration scheduling, and reward shaping aligned with key industrial performance metrics. Our method enables agents to dynamically adjust their policies based on performance trends, optimizing for throughput, order latency, and collision avoidance. We validate the approach in a simulated smart factory environment involving a fleet of AGVs, comparing it against Multi-Agent Deep Deterministic Policy Gradient (MADDPG), independent MARL, and heuristic baselines. Experimental results show that Adaptive MARL achieves up to 80% improvement in throughput and a 50% reduction in order cycle time, while minimizing collisions. The method scales effectively up to 8 agents and demonstrates resilience to disturbances such as job surges and agent failures. Overall, our adaptive MARL framework outperforms existing baselines in both efficiency and robustness, offering a scalable and intelligent control solution for Industry 4.0 systems.