A Modular Approach based on a Deep Reinforcement Learning Technique for the Plasma Magnetic Control in DEMO

In this paper we propose a modular approach, based on a deep reinforcement learning technique, for the control of a plasma with a limited configuration in the DEMO tokamak. Three different reinforcement learning agents are used to perform the magnetic confinement of the plasma, i.e. to stabilize the vertical plasma instability, to control the radial centroid position, and to ramp-up the plasma current. This modular approach allows us to simplify the training procedure of the control policy, since it requires a lower overall computational load. Performance of the proposed approach are characterized by numerical simulations.