### 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 language | English (US) |
---|---|

Title of host publication | Computational Science - ICCS 2009 - 9th International Conference, Proceedings |

Pages | 463-472 |

Number of pages | 10 |

Edition | PART 1 |

DOIs | |

State | Published - Aug 21 2009 |

Event | 9th International Conference on Computational Science, ICCS 2009 - Baton Rouge, LA, United States Duration: May 25 2009 → May 27 2009 |

### Publication series

Name | Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) |
---|---|

Number | PART 1 |

Volume | 5544 LNCS |

ISSN (Print) | 0302-9743 |

ISSN (Electronic) | 1611-3349 |

### Conference

Conference | 9th International Conference on Computational Science, ICCS 2009 |
---|---|

Country | United States |

City | Baton Rouge, LA |

Period | 5/25/09 → 5/27/09 |

### Fingerprint

### ASJC Scopus subject areas

- Theoretical Computer Science
- Computer Science(all)

### Cite this

*Computational Science - ICCS 2009 - 9th International Conference, Proceedings*(PART 1 ed., pp. 463-472). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 5544 LNCS, No. PART 1). https://doi.org/10.1007/978-3-642-01970-8_45

**Towards low-cost, high-accuracy classifiers for linear solver selection.** / Bhowmick, Sanjukta; Toth, Brice; Raghavan, Padma.

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

*Computational Science - ICCS 2009 - 9th International Conference, Proceedings.*PART 1 edn, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), no. PART 1, vol. 5544 LNCS, pp. 463-472, 9th International Conference on Computational Science, ICCS 2009, Baton Rouge, LA, United States, 5/25/09. https://doi.org/10.1007/978-3-642-01970-8_45

}

TY - GEN

T1 - Towards low-cost, high-accuracy classifiers for linear solver selection

AU - Bhowmick, Sanjukta

AU - Toth, Brice

AU - Raghavan, Padma

PY - 2009/8/21

Y1 - 2009/8/21

N2 - 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.

AB - 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.

UR - http://www.scopus.com/inward/record.url?scp=68849090240&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=68849090240&partnerID=8YFLogxK

U2 - 10.1007/978-3-642-01970-8_45

DO - 10.1007/978-3-642-01970-8_45

M3 - Conference contribution

AN - SCOPUS:68849090240

SN - 3642019692

SN - 9783642019692

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 463

EP - 472

BT - Computational Science - ICCS 2009 - 9th International Conference, Proceedings

ER -