Keywords: Tokamak control, Deep reinforcement learning, Curriculum learning, Distributed computing
TL;DR: This paper describes the use of curriculum learning to reduce training time and improve performance of RL-based controllers for routine use on tokamaks.
Abstract: Tokamaks are the leading candidates to achieve nuclear fusion as a sustainable source of energy, and plasma control plays a crucial role in their operations. However, the complex behavior of plasma dynamics makes control of these devices challenging through traditional methods. Recent works proved the usefulness of reinforcement learning as an efficient alternative, in order to fulfill these high-dimensional and non-linear situations. Despite their performance, controlling relevant plasma configurations requires expensive and long training sessions on simulations. In this work, we leverage the use of a curriculum strategy to achieve significant speed-up in learning a controller for the control coils, which tracks plasma quantities such as shape, position and current. To this end, we developed a fast, asynchronous and reliable framework to enable interactions between a distributed actor-critic and a C++ code simulating the WEST tokamak. By sequentially increasing task complexity, results show a clear reduction in convergence time and training cost. This work is one of the first attempts to enable fast production of robust magnetic controllers, for routine use in the operations of a magnetically confined fusion device.
Submission Number: 13
Loading