Batch reinforcement learning based dynamic optimization for polyethylene grade transitions

It is necessary for polyethylene grade transitions to establish dynamic optimization in order to preserve competitiveness in the global polymer market. Although the typical technology of iterative learning control has been performed to track the reference trajectories, it is usually difficult to obtain the optimal reference trajectories. A transition from one specific grade to another specific grade can be considered as one batch, therefore, we propose a feasible scheme for dynamic optimization of polyethylene grade transitions based on the batch reinforcement learning, which can derive a best possible policy after a few batches. The scheme aims to integrate the offline reference optimization and online implementations, and explore a better control policy instead of tracking the predetermined trajectories. This designed scheme has three distinct phases, collecting observations from reference and the historical trajectories, learning a policy, and executions. The proposed method is verified by the simulated polyethylene grade transitions, and good performance has been obtained.