Model-free tracking control of an optical fiber drawing process using deep reinforcement learning

Abstract

A deep reinforcement learning (DRL) approach for tracking control of an optical fiber drawing process is developed and evaluated. The DRL-based control is capable of regulating the fiber diameter to track either steady or varying reference trajectories in the presence of stochasticity and non-linear delayed dynamics of the system. With about 3.5 hours of real-time training, it outperformed other control models such as open-loop control, proportional-integral (PI) control, and quadratic dynamic matrix control (QDMC) in terms of diameter error. It does not require analytical or numerical model of the system dynamics unlike model-based approaches such as linear-quadratic regulator (LQR) or model predictive control (MPC). It can also track reference trajectories that it has never experienced in the training process.¹