A deep reinforcement learning method based on a multiexpert graph neural network for flexible job shop scheduling

When addressing flexible job shop scheduling problems (JSPs) via deep reinforcement learning (DRL), disjunctive graphs are commonly selected as the state observations of the agents. The previously developed methods primarily utilize graph neural networks (GNNs) to extract information from disjunctive graphs. However, as the instance scale increases, agents struggle to handle states with varying distributions, leading to reward confusion. To overcome this issue, inspired by the large-scale ’mixture-of-experts (MoE)’ model, we propose a novel module, i.e., a multiexpert GNN (ME-GNN), which integrates several approaches through a gating mechanism. Furthermore, the expert systems within the module facilitate lossless information propagation, providing robust support for solving complex cases. The experimental results demonstrate the effectiveness of our method. On synthetic datasets, our approach reduces the required makespan by 1.19%, and on classic datasets, it achieves a reduction of 1.34%. The multiple experts contained in the ME-GNN module enhance the overall flexibility of the system, effectively shortening the makespan.