.png){kind=icon}
.png){kind=icon}
Return
DGFFRL: Solving multi-machine collaborative scheduling based on Dual Graph Feature Fusion Reinforcement Learning
.png){kind=icon}
PRE
AI
W
P
Z
J
DOI：10.1016/j.cor.2025.107380.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Graph Neural Networks and Deep Reinforcement Learning have found widespread applications in numerous industrial manufacturing scenarios. However, solving highly intricate problems such as collaborative scheduling for multiple machines remains to be fully explored. This paper tackles this challenge by proposing an end-to-end reinforcement learning framework called Dual Graph Feature Fusion Reinforcement Learning (DGFFRL). Specifically, we construct a dual graph model based on the Machine Resource Constraint Graph and the Disjunctive Graph to accurately capture the intricate constraint relationships present in the Multi-machine Collaborative Job Shop Scheduling Problem (MCJSSP). Then, we utilize GNNs to extract rich feature representations of implicit and explicit constraints for each operation from MRCG and DG. These two types of features are dynamically fused. The resultant fused feature vector serves as the crucial state representation for the DGFFRL agent, driving its decision-making process. Furthermore, we introduce PPO-WD, an algorithm derived from Proximal Policy Optimization. By imposing Wasserstein distance regularization, PPO-WD effectively regulates the magnitude of policy updates while ensuring policy diversity. This synergistic effect significantly enhances the algorithm’s overall performance and the accuracy of scheduling decisions. To validate DGFFRL’s effectiveness and generalizability, we conducted comprehensive large-scale experiments. The results demonstrate that DGFFRL exhibits superior performance advantages and extensive applicability compared to current mainstream scheduling methods.
Journal
C
IF:
4.3
Papers:
6.5K
Citations:
1.8W
Organization
Cited Papers