Reinforcement Learning for PID Fine-Tuning in Nonlinear Temperature Control Systems

This paper compares the traditional tuning techniques, i.e., Ziegler-Nichols and Cohen-Coon, with artificial intelligence-based techniques, i.e., Reinforcement Learning (RL) and Genetic Algorithms (GA), for tuning Proportional-Integral-Derivative (PID) controllers in nonlinear temperature control systems, which are modeled by an augmented Newton's Law of Cooling with a heat dissipation coefficient that varies with temperature. The analysis is done in a simulated setting that includes sensor noise, control delays, and different levels of nonlinearity. System performance is evaluated based on stability, response time, energy efficiency, and disturbance robustness. Results from the simulations demonstrate that the GA-based approach consistently achieves the highest accuracy and stability across all tested conditions, outperforming both classical methods and RL-based strategies, which show limitations in handling complex dynamics and noise. These findings offer practical insights for selecting effective PID tuning methods for nonlinear thermal systems.