Offline Reinforcement Learning Architecture for Industrial Process Optimization: Application to the Tennessee Eastman Process

The application of reinforcement learning to process control holds significant promise in enhancing operational efficiency and robustness in industrial systems. However, the challenges associated with live online learning, including safety risks and data scarcity, have driven interest toward offline reinforcement learning methodologies. This thesis presents the development and validation of a data-driven control policy for industrial process systems using a combination of behavior cloning algorithm with a value-based learning algorithm to ensure safe policy initialization as well as conservative improvement. The proposed architecture is demonstrated on the Tennessee Eastman Process, a well established benchmark for complex chemical plant control. The data set, consisting of sequential state-action-reward transitions, was preprocessed and structured to configure and then train an agent to derive an optimal policy that maximizes long-term operational rewards without the need for online exploration. The results of this work demonstrate a practical pathway for deploying offline RL techniques in industrial environments, showcasing outputs that were safe, constraint-aware, yet still optimizing beyond historical operating behavior. The proposed framework highlights the potential for reinforcement learning to drive advances in automated process control, ultimately contributing to improved operational reliability, energy efficiency, and economic performance.