Data-Driven Real-Time Framework for Hybrid Flow Shop Scheduling Based on Dueling Double Deep Q Network

An innovative real-time scheduling framework is proposed to address a large-scale and multi resource constrained hybrid flow shop (HFS) scheduling problem. The framework utilizes the Dueling Double Deep Q Network (D3QN) algorithm to minimize the makespan. Firstly, HFS is transformed into a Manufacturing Petri Net (MPN) model [1], whose Markov state mechanisms are established. Secondly, the MPN simulation production process is repeatedly extrapolated and rewarded by a new mechanism, which generates a substantial amount of sample data. A multi-dimensional production information matrix is used as the MPN state input for the Q Value network, which is trained using the D3QN algorithm. Upon convergence to the optimal value function, the network model can be invoked in the online matching execution mechanism for optimal scheduling. In a
$30\times 16$
case, the average maximum completion time of this algorithm can be reduced to 293.09. The experimental results demonstrate that the network model trained by the D3QN algorithm under optimal hyperparameter settings has higher performance and response speed.