Multi-Agent Reinforcement Learning System for Multiloop Control of Chemical Processes
Yifei Yue, S. Lakshminarayanan
Abstract
Reinforcement Learning (RL) provide a model-free method of controlling processes. Recent advancements in Deep Reinforcement Learning have developed RL agents that can potentially control multivariate processes. Multi-Agent Reinforcement Learning (MARL) is a subfield of RL where multiple agents are trained in a shared environment. In this study, a MARL control system comprising multiple Twin-Delayed Deep Deterministic Policy Gradient (TD3) agents is trained to control a multiloop CSTR process. The MARL system achieves stable closed-loop response and good disturbance rejection.
Topics & Concepts
Reinforcement learningComputer scienceReinforcementMarlControl (management)Process (computing)Control systemControl theory (sociology)Artificial intelligenceControl engineeringEngineeringStructural engineeringPaleontologyStructural basinElectrical engineeringOperating systemBiologyAdvanced Control Systems OptimizationReinforcement Learning in RoboticsFault Detection and Control Systems