Towards low-cost, high-accuracy classifiers for linear solver selection

Sanjukta Bhowmick, Brice Toth, Padma Raghavan

Research output: Chapter in Book/Report/Conference proceedingConference contribution

11 Scopus citations

Abstract

The time to solve linear systems depends to a large extent on the choice of the solution method and the properties of the coefficient matrix. Although there are several linear solution methods, in most cases it is impossible to predict apriori which linear solver would be best suited for a given linear system. Recent investigations on selecting linear solvers for a given system have explored the use of classification techniques based on the linear system parameters for solver selection. In this paper, we present a method to develop low-cost high-accuracy classifiers. We show that the cost for constructing a classifier can be significantly reduced by focusing on the computational complexity of each feature. In particular, we filter out low information linear system parameters and then order the remaining parameters to decrease the total computation cost for classification at a prescribed accuracy. Our results indicate that the speedup factor of the time to compute the feature set using our ordering can be as high as 262. The accuracy and computation time of the feature set generated using our method is comparable to a near-optimal one, thus demonstrating the effectiveness of our technique.

Original languageEnglish (US)
Title of host publicationComputational Science - ICCS 2009 - 9th International Conference, Proceedings
Pages463-472
Number of pages10
EditionPART 1
DOIs
StatePublished - 2009
Event9th International Conference on Computational Science, ICCS 2009 - Baton Rouge, LA, United States
Duration: May 25 2009May 27 2009

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
NumberPART 1
Volume5544 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference9th International Conference on Computational Science, ICCS 2009
Country/TerritoryUnited States
CityBaton Rouge, LA
Period5/25/095/27/09

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint

Dive into the research topics of 'Towards low-cost, high-accuracy classifiers for linear solver selection'. Together they form a unique fingerprint.

Cite this