A Reinforcement Learning Model for Industrial Filling Process Control

In this paper, we propose an industrial filling process based on reinforcement learning. Despite the desired control can be achieved by using a simple PID controller, in many practical cases action space can be discrete as coarse and fine feed, there can be some residual material which may lead to overflow after the predefined setpoint is reached, or type and temperature of the material may affect the flow of the filling process. For such instances where the PID controller may struggle, in this work, we define a data-driven filling process based on reinforcement learning. We record different filling processes as episodes, define state-action pairs, and rewards to minimize the total filling duration, and set penalty functions for overflow and underflow situations. We use Monte Carlo method to find the best action in each state. The performance of the algorithm is illustrated with numerical examples.