Reinforcement Learning: Plant Optimization

Control of complex, dynamic industrial machinery has always been a difficult and labor intensive task. Not only do these tasks involve many variables and carry a high cost to failure, but they also commonly involve large “dead times”, or time which it takes to see effects in the system after some change. PSIORI is tasked with developing self-learning reinforcement learning controllers that help alleviate all the problems of classical control design for various industrial processes, including pulp bleaching, ore thickening, and ore sorting.

PSIORI is currently developing algorithms that employ the latest advances in Deep Reinforcement Learning to solve these dynamic industrial control problems. By using reinforcement learning, the operation of these plants can be learned from both live data and simulation data. The end result is an agent that can control the plants with an efficient control policy without operator interaction.

Having an efficient controller leads to increased profit and decreased or eliminated downtimes. Controllers designed in this fashion can also be easily transferred to other plants and can be quickly updated when details of a plant change. Reinforcement learning can be deployed for small and big problems with minimal alterations, and is thus an option for any industrial control problem.