Multi-Agent Reinforcement Learning for Distributed Flexible Job Shop Scheduling With Random Job Arrival

Distributed flexible job shops are increasingly becoming the predominant production method in manufacturing due to their advantages in low-cost production and high customization. In practical production environments, jobs arrive randomly but follow a regular pattern. This paper addresses the scheduling problem of the Distributed Flexible Job shop Scheduling Problem (DFJSP) with random job arrivals. The DFJSP consists of three sub-problems: factory selection, job assignment, and operation sequencing. To tackle this issue, the DFJSP is modeled as a Markov Decision Process (MDP), and a multi-agent approach based on deep reinforcement learning (DRL) is proposed. This approach includes a Distribute Agent (DA) and a Sequence Agent (SA). For the MDP of the DA, we designed 12 state features, 5 candidate actions, and a reward based on the current state of production tardiness. The SA is configured with 7 state features, 6 candidate actions, and rewards that reflect delay conditions. A deep Q-network (DQN) framework that incorporates a linearly decreasing threshold probability was designed to effectively balance exploration and exploitation during the training phase. Comparative experiments conducted on randomly generated instances demonstrate the effectiveness of the DA when used both independently and in conjunction with the SA.