Optimization of Code for Scalability of Deep Reinforcement Learning Agents in Slow and Stochastic Industrial Process Patterns

Resumo

This work investigates the impact of code optimization techniques on the scalability and training time of deep reinforcement learning (DRL) agents applied to stochastic and slow-dynamic industrial process control environments. DRL combines deep learning and reinforcement learning to address high-dimensional decision-making problems, but it often demands significant computational resources and time, especially in complex industrial scenarios. To address these challenges, this research explores the use of vectorization, parallelism, and hyperparameter optimization via Optuna to improve performance and training efficiency. Experiments were conducted using the “TempControl-v0” environment from the PIDGym library, simulating thermal inertia processes akin to industrial furnaces. Key metrics analyzed include total training time, average episode reward, and CPU/GPU utilization. The results demonstrate that strategic code optimization can significantly enhance the performance and scalability of DRL agents, making their application more feasible in real-world industrial contexts.