[Previous] | [Session 2] | [Next]

[2.02] Genetic Algorithms and Neural Networks for Astrophysical Applications

G. Edirisinghe (U. Tennessee & Johns Hopkins U.), D. Edirisinghe (U. Tennessee & Carnegie Mellon U.), O. Messer (U. Tennessee), M. W. Guidry (U. Tennessee & ORNL Physics Division*)

We report on the development of general-purpose algorithms for global parameter minimization in scientific applications where complex data sets resemble a pattern recognition problem. Our basic approach is to implement model calculations for the physical problem in parallel on a Beowulf cluster, with a genetic algorithm to optimize the parameters of the calculation and a neural network trained on observational data to compute the fitness function for members of the genetic algorithm population in each successive generation. We shall use this approach to investigate galaxy collisions using a gravity tree plus SPH hydrodynamics implemented with the code GADGET[1], our own genetic algorithm code for global parameter minimization, and our own neural network code for comparison of calculations with observational data. We shall also discuss other potential applications such as to the analysis of element production in large networks for r-process, hot-CNO, and rp-process calculations.

1. GADGET: A code for collisionless and gasdynamical cosmological simulations, Springel V., Yoshida N., White S. D. M., 2001, New Astronomy, 6, 51

*Managed by UT-Battelle, LLC, for the U.S. Department of Energy under contract DE-AC05-00OR22725.

[Previous] | [Session 2] | [Next]

Bulletin of the American Astronomical Society, 35 #3
© 2003. The American Astronomical Soceity.