Evaluation Concept for Order Release Based on Simulation and Reinforcement Learning Against Changes to the Production Configuration

Manufacturing an expanding number of diverse product variants while also enabling rapid responses to changes in the production schedule requires flexible production structures such as job shop production. Managing the resulting multitude of heterogeneous material flows leads to high complexity in production control. Hence, advanced dynamic scheduling methods based on simulation and reinforcement learning (RL) agents are required. These systems are confronted with changes in the production environment and unplanned disruptions, forcing the agent to handle deviated applications that require an adjustment. This paper presents an evaluation concept for a necessary retraining of production scheduling using a powerful control interface. The concept is based on the determination of an evaluation logic using logistic target values. By systematically analyzing changes in the production configuration in detail, respective production scenarios are compared with each other, deriving decision rules for the requisite retraining of the agent.