A survey and comparative evaluation of actor-critic methods in process control

Actor-critic (AC) methods have emerged as an important class of reinforcement learning (RL) paradigm that enables model-free control by acting on a process and learning from the consequence. To that end, these methods utilize artificial neural networks, which are synergized for action evaluation and optimal action prediction. This feature is highly desirable for process control, especially when the knowledge about a process is limited or when it is susceptible to uncertainties. In this work, we summarize important concepts of AC methods and survey their process control applications. This treatment is followed by a comparative evaluation of the set-point tracking and robustness of controllers based on five prominent AC methods, namely, DDPG, TD3, SAC, PPO, and TRPO, in five case studies of varying process nonlinearity. The training demands and control performances indicate the superiority of DDPG and TD3 methods, which rely on off-policy, deterministic search for optimal action policies. Overall, the knowledge base and results of this work are expected to serve practitioners in their efforts toward further development of autonomous process control strategies.