A dynamic flexible job shop scheduling method based on collaborative agent reinforcement learning

This paper presents an innovative approach to solve the Dynamic Flexible Job Shop Scheduling Problem (DFJSP). Our method aims to enhance production efficiency by minimizing the average total tardiness. To achieve this goal, we first construct a model that accurately describes the environmental state using eight feature values ranging from 0 to 1. These feature values comprehensively consider factors such as job progress and machine utilization, thus providing a full reflection of the actual operational status of the workshop. Additionally, we design four composite scheduling rules that address the dual selection issues of machines and jobs in the DFJSPs, enabling optimal decisions to be made within each scheduling cycle. We design a reward function based on changes in the state values, ensuring that the number of reward values increases as the training progresses, alleviating the problem of sparse rewards to some extent. This design facilitates faster learning and improves the convergence speed of the algorithm. In terms of agent design, we employ a Deep Q-Network and a Proximal Policy Optimization algorithms. Both methods effectively handle complex decision spaces and exhibit good stability during training. Through a data-sharing approach, we further develop a collaborative agent model that enables efficient cooperation among multiple agents. Finally, we validate the effectiveness of our proposed model through a series of experiments. The results demonstrate that our model shows advantages, regardless of whether the dataset is small or large. When compared with other methods, our approach maintains high performance across different problem instances. These results fully validate the correctness and effectiveness of our design approach and provide a viable solution for scheduling problems in practical industrial settings.