Control of A Polyol Process Using Reinforcement Learning

Abstract

Abstract Reinforcement learning is a branch of machine learning, where an agent gradually learns a control policy via a combination of exploration and interactions with a system. Recent successes of model-free reinforcement learning (RL) has attracted tremendous attention from the process control community. For instance, RL has been successfully applied in very complex control tasks (e.g., games such as chess or Go that contain large state spaces) and is shown to be robust to uncertainties. These findings indicate that there is a significant potential to leverage RL methods to improve the control of chemical processes. In this work, RL was applied to a detailed and accurate simulation of an industrial polyol process. To manufacture the desired product, the RL controller is required to achieve the target ending conditions determined by four key parameters; meanwhile, economic factors are also considered in this process, including batch reaction time and total feed amounts. The obtained results show a high consistency between RL and the current optimal operating conditions. Additionally, an improvement opportunity was identified by extending current control bounds of the manipulated variables. This work illustrates that RL is capable of handling complicated industrial systems, even under realistic operating constraints.