Robust regression with CUDA and its application to plasma reflectometry

Tutkimustuotos: Lehtiartikkelivertaisarvioitu


  • JET contributors


  • Univ Lisbon, Universidade de Lisboa, Inst Super Tecn, Inst Plasmas & Fusao Nucl
  • Univ Lisbon, Universidade de Lisboa, Inst Super Tecn, IPFN
  • Inst Plasma Res, Institute for Plasma Research (IPR)
  • Czech Academy of Sciences
  • Culham Sci Ctr, Culham Science Centre, CCFE
  • Queens Univ, Queens University Belfast, Dept Pure & Appl Phys
  • Univ Tartu, University of Tartu
  • Univ Napoli Federico II, University of Naples Federico II, Consorzio CREATE
  • Asociac EURATOM CIEMAT, Euratom, Lab Nacl Fus
  • CNR, Istituto Fisica del Plasma "Piero Caldirola" (IFP-CNR), Consiglio Nazionale delle Ricerche (CNR), Ist Fis Plasma
  • ITER Org
  • Princeton Plasma Phys Lab, Princeton Physics Laboratory, Princeton University, United States Department of Energy (DOE)
  • Uppsala University
  • European Commiss
  • Chinese Academy of Sciences
  • Natl Fus Res Inst, National Fusion Research Institute (NFRI)
  • Univ Politecn Madrid, Universidad Politecnica de Madrid, Grp I2A2
  • CNRS, Centre National de la Recherche Scientifique (CNRS), CNRS - Institute for Engineering & Systems Sciences (INSIS), Ecole Polytechnique, Sorbonne Universite, UMR 7648, Ecole Polytech
  • Forsch Zentrum Julich GmbH, Helmholtz Association, Research Center Julich, Inst Energie & Klimaforsch Plasmaphys
  • Univ Elect Sci & Technol China, University of Electronic Science & Technology of China
  • VTT Technical Research Centre of Finland


In many applications, especially those involving scientific instrumentation data with a large experimental error, it is often necessary to carry out linear regression in the presence of severe outliers which may adversely affect the results. Robust regression methods do exist, but they are much more computationally intensive, making it difficult to apply them in real-time scenarios. In this work, we resort to graphics processing unit (GPU)-based computing to carry out robust regression in a time-sensitive application. We illustrate the results and the performance gains obtained by parallelizing one of the most common robust regression methods, namely, least median of squares. Although the method has a complexity of O(n(3) log n), with GPU computing, it is possible to accelerate it to the point that it becomes usable within the required time frame. In our experiments, the input data come from a plasma diagnostic system installed at Joint European Torus, the largest fusion experiment in Europe, but the approach can be easily transferred to other applications.


JulkaisuReview of Scientific Instruments
TilaJulkaistu - marraskuuta 2015
OKM-julkaisutyyppiA1 Julkaistu artikkeli, soviteltu

ID: 38581424