Multi-Agent Flexible Job Shop Scheduling by Deep Learning on Graphs With Evolutionary Neural Topology

Increased competition in the market has formed a novel production model of multivariety and small batch to satisfy the demand of different users for personalized products. Since each user pursues different objectives, a multi-agent system is integrated into the novel production model to balance the satisfaction between diverse users and improve the overall benefit of the model. Multi-agent flexible job shop scheduling problem (MAFJSP) is researched in this article, with objectives including total tardiness (TTD) of user agents and delivery time balancing (DTB) of the job shop agent. Deep learning on graphs with evolutionary neural topology (GDL-ENT) is devised to address the MAFJSP. Aheterogeneous graph is designed to carefully represent the environment information containing the operation, machine, and user agents, while a heterogeneous attention network (HAN) is employed to hierarchically extract state features according to user agents. Meanwhile, the automatic generation of the structure and hyperparameters for the neural network in deep learning on graphs is achieved by the evolutionary neural topology search method. The performance and generalization of the GDL-ENT, as well as the influence of main components in the GDL-ENT, are analyzed on the randomly generated test suite, public benchmark instances, and the real-world dataset, respectively. The GDL-ENT has superior competitiveness compared to the state-of-the-art algorithms for solving the MAFJSP from experiment results.