Graph neural network and multi-agent reinforcement learning for machine-process-system integrated control to optimize production yield

In this paper, an integrated control framework is proposed for the optimization of production yield by integrating different levels a manufacturing system, including process, and machine levels. The system modeled as graph treating machines nodes material flows links. model enjoys high flexibility able to incorporate all relevant real-time information across in dynamic node feature. Since tool state essential decision making, Recursive Bayesian Estimation (RBE) adopted reduce observations through sensors learning models provide more accurate estimation be included into With model, Graph Neural Network (GNN) applied process features generate embedding that reflects both local global information. For purpose, each then treated distributed agent Multi-Agent Reinforcement Learning (MARL) conditions its policy on from GNN. State-of-the-art GNN MARL algorithms, namely Attention (GAT) Value Decomposition Actor Critic (VDAC), are implemented train learnable parameters GNN-MARL networks learn optimal multi-agent policy. Extensive numerical experiments analysis proves effectiveness framework.