Application of a Reinforcement Learning-based Automated Order Release in Production

The importance of job shop production is increasing in order to meet the customer-driven greater demand for products with a larger number of variants in small quantities. However, it also leads to higher requirements for the production planning and control. In order to meet logistical target values and customer needs, one approach is the focus on dynamic planning systems, which can reduce ad-hoc control interventions in the running production. In particular, the release of orders at the beginning of the production process has a high influence on the planning quality. Previous approaches used advanced methods such as combinations of reinforcement learning (RL) and simulation to improve specific production environments, which are sometimes highly simplified and not practical enough. This paper presents a practice-based application of an automated order release procedure based on RL using the example of real-world production scenarios. Both, the training environment, and the data processing method are introduced. Primarily, three aspects to achieve a higher practical orientation are addressed: A more realistic problem size compared to previous approaches, a higher customer orientation by means of an objective regarding adherence to delivery date and a control application for development and performance evaluation of the considered algorithms against known order release strategies. Follow-up research will refine the objective function, continue to scale-up the problem size and evaluate the algorithm’s scheduling results in case of changes in the system.