Benchmark study of reinforcement learning in controlling and optimizing batch processes

Abstract

AbstractIn this article, multiple reinforcement learning (RL) methods such as value‐based, policy‐based, and actor‐critic algorithms are investigated for typical control tasks found in the chemical industries. Through a critical assessment of these novel techniques, their main advantages are highlighted, but also the challenges that still need to be resolved are discussed. Two batch control tasks are used as benchmarks, namely, production maximization, and setpoint control. Using these testing environments, a direct comparison of different RL approaches is presented, which could guide the algorithm selection in future RL applications for batch process control. Furthermore, the results obtained with a traditional control method, model predictive control (MPC), are shown to provide a baseline for comparison with RL algorithms. The results show that RL has significant applicability in various control tasks and has comparable control performance to traditional methods but with a lower online computational cost. A batch bioreactor simulation and a simulation of an industrial polyol process are used for illustration purposes.