Simulation and Optimization of Drying Operations with Reinforcement Learning (PDF)

• Overview: This research project, conducted in partnership with a North American wood drying company, explored the use of Deep Reinforcement Learning (Deep-RL) combined with a detailed simulator (Digital Twin) for optimizing industrial processes. The project aimed to evaluate whether advanced learning algorithms could improve the efficiency of resource allocation in industrial operations, providing actionable insights for real-world applications.

• Goal: The primary objective was to train a Deep-RL agent capable of efficiently allocating resources within a simulated industrial environment. This approach leverages the potential of simulation-based learning to optimize complex processes while minimizing real-world trial-and-error costs. The project also aimed to assess how simulation-based DRL training could be applied in broader contexts, such as other industrial automation tasks.

• Details: The simulation environment was developed using OpenAI Gym and Simpy to replicate the company’s operational processes accurately. A Deep-RL agent was trained using the Proximal Policy Optimization (PPO) algorithm, combined with reward engineering techniques like “Heuristic-Guided Reinforcement Learning” to guide the agent’s learning process. I advocated for the use of Deep-RL and simulation-based training, believing it represents a forward-looking approach for optimizing industrial processes and robotics. This project showcased the potential of Deep-RL and Digital Twin technology to drive efficiency and innovation in industrial settings.