Supervised Imitation Training for Deep Reinforcement Learning Control of a Fluid Catalytic Cracking System

Reinforcement Learning (RL) based intelligent automation techniques have been widely applied in various industrial sectors, ranging from unmanned aerial vehicles to autonomous driving vehicles. However, training a sophisticated RL controller for the industry-level application is usually time-consuming and expensive. To tackle this challenge, a supervised imitation training approach is proposed in this work. This approach utilizes state-of-art industrial controllers, particularly the PID controllers, to facilitate training RL controllers. The actor network within the RL controller is pre-trained by fitting the data which are collected from traditional controllers. As a result, the agent at first behaves similarly to the conventional controller, but its performance could be improved during the online stage for training the RL agent. Experiments were conducted for a simulated fluid catalytic cracking, where the effectiveness of the supervised imitation approach is verified.