Multi-Task Multi-Agent Reinforcement Learning for Real-Time Scheduling of a Dual-Resource Flexible Job Shop with Robots

In this paper, a real-time scheduling problem of dual-resource flexible job shop with robots is studied. Multiple independent and their supervised machine sets form own work cells. First, mixed integer programming model established, which considers the problems jobs machines in cells, between based on process plan flexibility. Second, order to make decisions, framework multi-task multi-agent reinforcement learning centralized training decentralized execution proposed. Each agent interacts environment completes three decision-making tasks: sequencing, selection, planning. training, value network used evaluate optimize policy achieve cooperation, attention mechanism introduced into realize information sharing among multiple tasks. execution, each performs task decisions through local observations according trained network. Then, observation, action, reward are designed. Rewards include global rewards, decomposed sub-rewards corresponding The algorithm designed double-deep Q-network. Finally, simulation derived from benchmarks, experimental results show effectiveness proposed method.