Enhancing Reinforcement Learning Robustness via Integrated Multiple-Model Adaptive Control

Reinforcement learning (RL) has attracted considerable attention from both industry and academia for its success in solving complex problems. However, the performance of RL algorithms often decreases in environments characterized by uncertainties, unmodeled dynamics, and nonlinearities. This paper presents a novel robust RL algorithm designed to ensure closed-loop stability for industrial processes. The algorithm considers a wide range of potential scenarios across various operating conditions and different ranges of parameter uncertainties. Using the multiple-model adaptive control methodology, the algorithm evaluates all scenarios and ranks them based on their likelihood of accurately characterizing the actual process. The validity of the results is demonstrated using a benchmark continuous stirred tank reactor (CSTR) problem. Copyright (C) 2024 The Authors.