TAKING THE NEXT STEP WITH INTELLIGENT MONTE CARLO

Booth, T.E. ; Carlson, J.A.

Los Alamos National Laboratory

关键词: Statistics;    99 General And Miscellaneous//Mathematics, Computing, And Information Science;    Computers;    Transport;    Efficiency;   

DOI  :  10.2172/765262
RP-ID  :  LA-UR-00-4892
RP-ID  :  W-7405-ENG-36
RP-ID  :  765262

美国|英语

来源: UNT Digital Library

【 摘 要 】

For many scientific calculations, Monte Carlo is the only practical method available. Unfortunately, standard Monte Carlo methods converge slowly as the square root of the computer time. We have shown, both numerically and theoretically, that the convergence rate can be increased dramatically if the Monte Carlo algorithm is allowed to adapt based on what it has learned from previous samples. As the learning continues, computational efficiency increases, often geometrically fast. The particle transport work achieved geometric convergence for a two-region problem as well as for problems with rapidly changing nuclear data. The statistics work provided theoretical proof of geometic convergence for continuous transport problems and promising initial results for airborne migration of particles. The statistical physics work applied adaptive methods to a variety of physical problems including the three-dimensional Ising glass, quantum scattering, and eigenvalue problems.