The Joint Accelerator Conferences Website (JACoW) is an international collaboration that publishes the proceedings of accelerator conferences held around the world.
TY - CONF AU - Ivanov, A.N. AU - Agapov, I.V. AU - Eichler, A. AU - Tomin, S. ED - Liu, Lin ED - Byrd, John M. ED - Neuenschwander, Regis T. ED - Picoreti, Renan ED - Schaa, Volker R. W. TI - Physics-Enhanced Reinforcement Learning for Optimal Control J2 - Proc. of IPAC2021, Campinas, SP, Brazil, 24-28 May 2021 CY - Campinas, SP, Brazil T2 - International Particle Accelerator Conference T3 - 12 LA - english AB - We propose an approach for incorporating accelerator physics models into reinforcement learning agents. The proposed approach is based on the Taylor mapping technique for simulation of the particle dynamics. The resulting computational graph is represented as a polynomial neural network and embedded into the traditional reinforcement learning agents. The application of the model is demonstrated in a nonlinear simulation model of beam transmission. The comparison of the approach with the traditional numerical optimization as well as neural networks based agents demonstrates better convergence of the proposed technique. PB - JACoW Publishing CP - Geneva, Switzerland SP - 4150 EP - 4152 KW - network KW - lattice KW - controls KW - simulation KW - alignment DA - 2021/08 PY - 2021 SN - 2673-5490 SN - 978-3-95450-214-1 DO - doi:10.18429/JACoW-IPAC2021-THPAB191 UR - https://jacow.org/ipac2021/papers/thpab191.pdf ER -