Reinforcement Learning Techniques To Support Industrial Process Control

This paper will discuss some recent industry-university research efforts on reinforcement learning (RL) in the context of industrial process control. Reinforcement learning is a type of machine learning where a software agent is allowed to decide what actions to take on an environment. Rewards are given for actions that result in desirable outcomes so that, over time, the reinforcement learning system improves the actions that it takes in the sense of maximizing its rewards. There are many different approaches to the implementation of reinforcement learning and many possible applications to industrial process control and related domains. For example, an agent could be constructed as a deep neural network and could be trained to perform control on an industrial process, possibly replacing traditional controllers such as proportional-integral-derivative (PID) or model predictive control (MPC). However, this approach turns the controller into a true 'black box' lacking the intuitiveness of the established technologies. Another approach is, instead of replacing the controller, to have the agent adjust the controller tuning parameters. This allows for checking of 'reasonableness' of the learned tuning parameters, while benefiting from the automated learning aspect of the RL algorithm.

One drawback to reinforcement learning is that it can be 'data hungry', meaning that a lot of trial-and-error may be required for the agent to learn to do its task well. In an industrial context, extensive trial-and-error on a working mill would be generally unacceptable. For this reason, research is being performing into offline learning, and other techniques to 'improve sample efficiency', i.e. to lessen the amount of online experimentation needed. In this research, these aspects of reinforcement learning are explored through application of reinforcement learning to PID tuning, and with meta-RL which greatly reduces online experimentation by pretraining the agent on different environments.

TAPPI conference proceedings and presentations, technical papers, and publication articles provide technical and management data and solutions on topics covering the Pulp, Paper, Tissue, Corrugated Packaging, Flexible Packaging, Nanotechnology and Converting Industries.

Simply select the quantity, add to your cart and your conference paper, presentation or article will be available for immediate download.