This work reports the theoretical and computational results concerning some recent gradient-based methods of Barzilai-Borwein type for Mathematical Programming and their parallel implementation for the solution of the training problem in large-scale support vector machines applications. Recently developed steplength updating rules are implemented as the core of a parallel decomposition technique in an existing object-oriented code. Good performances are obtained by the code tuning and optimization on well known challenging data sets, up to 64 processors. Moreover, a meaningful new formulation of the standard binary optimal separating hyperplane problem is given, which has interesting theoretical properties.
Parallel Gradient Methods for Some Classes of Large-Scale Nonlinear Programming Problems
ZANGHIRATI, Gaetano;
2008
Abstract
This work reports the theoretical and computational results concerning some recent gradient-based methods of Barzilai-Borwein type for Mathematical Programming and their parallel implementation for the solution of the training problem in large-scale support vector machines applications. Recently developed steplength updating rules are implemented as the core of a parallel decomposition technique in an existing object-oriented code. Good performances are obtained by the code tuning and optimization on well known challenging data sets, up to 64 processors. Moreover, a meaningful new formulation of the standard binary optimal separating hyperplane problem is given, which has interesting theoretical properties.I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
