A direct orthogonal sparse static methodology for a finite continuation hybrid LP solver

Authors:
Pablo Guerrero-García;Ángel Santos-Palomo
Affiliations:
Department of Applied Mathematics,Complejo Tecnológico, University of Málaga, Málaga, Spain;Department of Applied Mathematics,Complejo Tecnológico, University of Málaga, Málaga, Spain
Venue:
PARA'04 Proceedings of the 7th international conference on Applied Parallel Computing: state of the Art in Scientific Computing
Year:
2004

Citing 6
Cited 0

Sparse Linear Least Squares Problems in Optimization

Computational Optimization and Applications
An Object-Oriented Approach to the Design of a User Interface for a Sparse Matrix Package

SIAM Journal on Matrix Analysis and Applications
A New Finite Continuation Algorithm for Linear Programming

SIAM Journal on Optimization
A software package for sparse orthogonal factorization and updating

ACM Transactions on Mathematical Software (TOMS)
Computational Techniques of the Simplex Method

Computational Techniques of the Simplex Method
Large-scale linear programming using the Cholesky factorization.

Large-scale linear programming using the Cholesky factorization.

Quantified Score

Hi-index	0.00

Visualization

Abstract

A finite continuation method for solving linear programs (LPs) has been recently put forward by K. Madsen and H. B. Nielsen which, to improve its performance, can be thought of as a Phase-I for a non-simplex active-set method (also known as basis-deficiency-allowing simplex variation); this avoids having to start the simplex method from a highly degenerate square basis. An efficient sparse implementation of this combined hybrid approach to solve LPs requires the use of the same sparse data structure in both phases, and a way to proceed in Phase-II when a non-square working matrix is obtained after Phase-I. In this paper a direct sparse orthogonalization methodology based on Givens rotations and a static sparsity data structure is proposed for both phases, with a Linpack-like downdating without resorting to hyperbolic rotations. Its sparse implementation (recently put forward by us) is of reduced-gradient type, regularization is not used in Phase-II, and occasional refactorizations can take advantage of row orderings and parallelizability issues to decrease the computational effort.