A Deep Reinforcement Learning approach for Vertical Stabilization of tokamak plasmas

Reinforcement Learning has emerged as a promising approach to implement efficient data-driven controllers for a variety of applications. In this paper, a Deep Deterministic Policy Gradient (DDPG) algorithm is used to train a Vertical Stabilization agent, to be considered as a possible alternative to the model-based solutions usually adopted in existing machines. The agent is trained and validated considering the ITER tokamak magnetic control as case study environment. The tuning of the DDPG algorithm's hyper-parameters is motivated through a sensitivity analysis.