The European Physical Journal Applied Physics

Table of Contents - November 2000 - Volume 12, Issue 02  

Research Article

Using neural networks to speed up optimization algorithms*

M. Bazana1 and S. Russenschucka2

The Institute of Computer Science, University of Wrocław, Przesmyckiego 21, 51-151 Wrocław, Poland

CERN, 1211 Geneva 23, Switzerland

Abstract

The paper presents the application of Radial-basis-function (RBF) neural networks to speed up deterministic search algorithms used for the design and optimization of superconducting LHC magnets. The optimization of the iron yoke of the main dipoles requires a number of numerical field computations per trial solution as the field quality depends on the excitation of the magnets. This results in computation times of about 30 minutes for each objective function evaluation (on a DEC-Alpha 600/333) and only the most robust (deterministic) optimization algorithms can be applied. Using a RBF function approximator, the achieved speed-up of the search algorithm is in the order of 25% for problems with two parameters and about 18% for problems with three and five design variables.

(Received May 11 2000)

(Accepted August 29 2000)

(Online publication November 15 2000)

PACS:

  • 02.60.Ed – Interpolation; curve fitting;
  • 07.05.Mh – Neural networks, fuzzy logic, artificial intelligence;
  • 02.60.Pn – Numerical optimization

stephan.russenschuck@cern.ch

Footnotes

*  This paper has been presented at NUMELEC 2000.