## Assessing the Potential of Interior Methods for Nonlinear Optimization (2002)

### Cached

### Download Links

Citations: | 7 - 1 self |

### BibTeX

@MISC{Morales02assessingthe,

author = {José Luis Morales and Jorge Nocedal and Richard A. Waltz and Guanghui Liu and Jean-pierre Goux},

title = {Assessing the Potential of Interior Methods for Nonlinear Optimization},

year = {2002}

}

### OpenURL

### Abstract

A series of numerical experiments with interior point (LOQO, KNITRO) and active-set sequential quadratic programming (SNOPT, filterSQP) codes are reported and analyzed. The tests were performed with small, medium-size and moderately large problems, and are examined by problem classes. Detailed observations on the performance of the codes, and several suggestions on how to improve them are presented. Overall, interior methods appear to be strong competitors of active-set SQP methods, but all codes show much room for improvement. 1

### Citations

2141 | Numerical optimization - Nocedal, Wright - 1999 |

379 |
AMPL: A Modeling Language for Mathematical Programming
- Fourer, Gay, et al.
- 2002
(Show Context)
Citation Context ...mplemented in lterSQP [13], but this is still an ongoing area of research. 3 Testing Environment We performed tests using 509 problems from the CUTE collection [2] that have been translated into AMPL =-=[16]-=- by Benson and Vanderbei [24]. We selected asubsetof the 740 problems available in the Benson-Vanderbei test set as of December, 2000, according to the following criteria. All linear programming (LP) ... |

374 | SNOPT: An SQP algorithm for large-scale constrained optimization
- Gill, Murray, et al.
- 2002
(Show Context)
Citation Context ...ive with established codes for large-scale nonlinear programming. Several production packages, such as LANCELOT [8], MINOS [20], and CONOPT [11] would have served our purposes, but wehavechosen SNOPT =-=[18]-=- and lterSQP [14] because wehave experience with both codes, and because they constitute two complementary approaches to SQP algorithms: SNOPT is a line search method using convex models and lterSQP i... |

263 | Benchmarking optimization software with performance profiles
- Dolan, Moré
(Show Context)
Citation Context ...e step is a local minimizer of the quadratic model of f subject to the trust region constraint. All the results in this paper will be presented using the performance proles proposed by Dolan and More =-=[10]-=-. Let tp�s denote the time to solve problem p by solver s. Then de ne the ratios rp�s = tp�s � (8) tp where t p is the lowest time required by anycodetosolve problem p. Ifacode does not solve a proble... |

179 | Nonlinear programming without a penalty function
- Fletcher, Leyer
(Show Context)
Citation Context ...hed codes for large-scale nonlinear programming. Several production packages, such as LANCELOT [8], MINOS [20], and CONOPT [11] would have served our purposes, but wehavechosen SNOPT [18] and lterSQP =-=[14]-=- because wehave experience with both codes, and because they constitute two complementary approaches to SQP algorithms: SNOPT is a line search method using convex models and lterSQP is a trust region ... |

169 | D.F.Shanno: An interior point algorithm for nonconvex nonlinear programming
- Vanderbei
- 1999
(Show Context)
Citation Context ...he CUTE collection. By this we mean problems that have upto10,000 variables or constraints. Avariety ofinterior (or barrier) methods for nonlinear programming have been proposed in the last few years =-=[1,5,7,12,15,17,25,26]-=-, but only a few implementations are currently available for public use. We have chosen to experiment with the software packages LOQO [25] and KNITRO [5], which are available through the NEOS system [... |

139 |
The conjugate gradient method and trust regions in large scale optimization
- Steihaug
- 1983
(Show Context)
Citation Context ...x, if necessary. KNITRO isan (inexact) Newton method with a spherical trust region. The step is computed by an unpreconditioned conjugate gradient iteration, which is terminated using Steihaug's rule =-=[23]-=-. The step is accepted if it provides su cient reduction in f. SNOPT is a quasi-Newton method with a backtracking line search requiring su cient reduction in f. But it is not a traditional unconstrain... |

130 |
CUTE: constrained and unconstrained testing environment
- Toint
- 1995
(Show Context)
Citation Context ...d for variations of the lter mechanism implemented in lterSQP [13], but this is still an ongoing area of research. 3 Testing Environment We performed tests using 509 problems from the CUTE collection =-=[2]-=- that have been translated into AMPL [16] by Benson and Vanderbei [24]. We selected asubsetof the 740 problems available in the Benson-Vanderbei test set as of December, 2000, according to the followi... |

111 | A trust region method based on interior point techniques for nonlinear programming
- Gilbert, Nocedal
(Show Context)
Citation Context ...ter is chosen at each iteration so that the step d is a descent direction for the merit function (6). A global convergence theory of an algorithm similar to that implemented in KNITRO is presented in =-=[4]-=-, whereas rate of convergence results are given in [6]. 2.3 SNOPT SNOPT 5.3-4 [18] is a line search SQP method in which the search direction is determined by anactiver-set method for convex quadratic ... |

103 | LANCELOT: a Fortran package for large-scale nonlinear optimization (Release A). Number 17 - Toint - 1992 |

95 |
Y.Zhang: On the formulation and theory of Newton interior-point method for nonlinear programming
- El-Bakry, Tsuchiya
- 1996
(Show Context)
Citation Context ...he CUTE collection. By this we mean problems that have upto10,000 variables or constraints. Avariety ofinterior (or barrier) methods for nonlinear programming have been proposed in the last few years =-=[1,5,7,12,15,17,25,26]-=-, but only a few implementations are currently available for public use. We have chosen to experiment with the software packages LOQO [25] and KNITRO [5], which are available through the NEOS system [... |

95 | Interior-point methods for nonconvex nonlinear programming: filter methods and merit functions
- Benson, Shanno, et al.
(Show Context)
Citation Context ...ection d is a descent direction for . A dynamic rule (see (11)) is used to update the value of the barrier parameter at each iteration. The KKT system (3) is solved using a sparse LDL T factorization =-=[22]-=-. In order to handle non-convex problems, LOQO monitors the Cholesky factors of r 2 L(x� y)+A(x) T S ;1 YA(x)� and if the factorization implies that this matrix is not su ciently positive de nite, a m... |

83 |
MINOS 5.4 User's Guide
- Murtagh, Saunders
- 1983
(Show Context)
Citation Context ...ermine if the interior methods implemented in these two packages are competitive with established codes for large-scale nonlinear programming. Several production packages, such as LANCELOT [8], MINOS =-=[20]-=-, and CONOPT [11] would have served our purposes, but wehavechosen SNOPT [18] and lterSQP [14] because wehave experience with both codes, and because they constitute two complementary approaches to SQ... |

82 |
The NEOS server
- Czyzyk, Moré
- 1998
(Show Context)
Citation Context ...], but only a few implementations are currently available for public use. We have chosen to experiment with the software packages LOQO [25] and KNITRO [5], which are available through the NEOS system =-=[9]-=-, as well as in machine executable form. These packages implement two distinct interior point approaches: LOQO is a line search algorithm that has much in common with interior algorithms for linear an... |

75 | An Interior Point Algorithm for Large Scale Nonlinear
- Byrd, Hribar, et al.
- 1997
(Show Context)
Citation Context ...he CUTE collection. By this we mean problems that have upto10,000 variables or constraints. Avariety ofinterior (or barrier) methods for nonlinear programming have been proposed in the last few years =-=[1,5,7,12,15,17,25,26]-=-, but only a few implementations are currently available for public use. We have chosen to experiment with the software packages LOQO [25] and KNITRO [5], which are available through the NEOS system [... |

65 |
A primal-dual interior method for nonconvex nonlinear programming
- Gay, Overton, et al.
- 1997
(Show Context)
Citation Context |

63 | Primal-dual interior methods for nonconvex nonlinear programming
- Forsgren, Gill
- 1998
(Show Context)
Citation Context |

39 |
Trust region algorithms for optimization with nonlinear equality and inequality constraints
- Omojokun
- 1989
(Show Context)
Citation Context ...decreased by a xed factor of ve. Each step is computed as the sum d = v + t of a normal component v whose objective is to improve feasibility, and a tangential component t that aims toward optimality =-=[3,21]-=-. The main cost of the iteration lies in the computation of the tangential component, which is performed using a projected conjugate gradient (CG) iteration. Each CG iteration involves two main comput... |

26 | On the local behavior of an interior point method for nonlinear programming
- Byrd, Liu, et al.
- 1997
(Show Context)
Citation Context ...a descent direction for the merit function (6). A global convergence theory of an algorithm similar to that implemented in KNITRO is presented in [4], whereas rate of convergence results are given in =-=[6]-=-. 2.3 SNOPT SNOPT 5.3-4 [18] is a line search SQP method in which the search direction is determined by anactiver-set method for convex quadratic programming. The line search enforces decrease of an a... |

23 |
A globally and superlinearly convergent primal-dual interor point trust region method for large scale constrained optimization
- Yamashita, Yabe, et al.
- 2005
(Show Context)
Citation Context |

22 |
Robust trust region methods for constrained optimization
- Byrd
- 1987
(Show Context)
Citation Context ...decreased by a xed factor of ve. Each step is computed as the sum d = v + t of a normal component v whose objective is to improve feasibility, and a tangential component t that aims toward optimality =-=[3,21]-=-. The main cost of the iteration lies in the computation of the tangential component, which is performed using a projected conjugate gradient (CG) iteration. Each CG iteration involves two main comput... |

21 |
A primal-dual trust-region algorithm for nonconvex nonlinear programming
- Toint
(Show Context)
Citation Context |

21 | Benchmarking optimization software with COPS - Dolan, Moré - 2000 |

18 | Global convergence of a trustregion SQP-filter algorithm for general nonlinear programming - Toint, Wächter |

13 | A feasible BFGS interior point algorithm for solving strongly convex minimization problems - Armand, Gilbert, et al. - 2000 |

13 | CONOPT: a large scale GRG code - Drud - 1994 |

12 | A new technique for inconsistent qp problems in the sqp method - Spellucci - 1998 |

6 |
LANCELOT: aFORTRAN package for large-scale nonlinear optimization (Release A)", Number 17
- Conn, Gould, et al.
- 1992
(Show Context)
Citation Context ... try to determine if the interior methods implemented in these two packages are competitive with established codes for large-scale nonlinear programming. Several production packages, such as LANCELOT =-=[8]-=-, MINOS [20], and CONOPT [11] would have served our purposes, but wehavechosen SNOPT [18] and lterSQP [14] because wehave experience with both codes, and because they constitute two complementary appr... |

6 |
AMPL models. http://www.sor.princeton.edu/rvdb/ampl/nlmodels
- Vanderbei
(Show Context)
Citation Context ...ut this is still an ongoing area of research. 3 Testing Environment We performed tests using 509 problems from the CUTE collection [2] that have been translated into AMPL [16] by Benson and Vanderbei =-=[24]-=-. We selected asubsetof the 740 problems available in the Benson-Vanderbei test set as of December, 2000, according to the following criteria. All linear programming (LP) and quadratic programming (QP... |

4 | A numerical study of limited memory BFGS methods
- Morales
(Show Context)
Citation Context ...herefore, for large problems it will be ine - cient tokeep both a sparse and a dense version of essentially the same n n matrix. An evaluation of the limited memory strategy used in SNOPT is given in =-=[19]-=-. FilterSQP is an (exact) Newton method using an `1 trust region. The lter mechanism reduces, in the unconstrained case, to a standard su cient decrease condition on f. The search direction is compute... |

3 | A globally and superlineraly convergent primaldual point trust region method for large scale constrained optimization - Yamashita, Yabe, et al. - 1997 |

2 | donlp2. program and users guide available as donlp2.tar from netlib/opt - Spellucci |

1 | Numerical Analysis Report NA/171 - Forsgren, Gill - 1997 |

1 |
CONOPT { a large scale GRG code.ORSA
- Drud
- 1994
(Show Context)
Citation Context ...erior methods implemented in these two packages are competitive with established codes for large-scale nonlinear programming. Several production packages, such as LANCELOT [8], MINOS [20], and CONOPT =-=[11]-=- would have served our purposes, but wehavechosen SNOPT [18] and lterSQP [14] because wehave experience with both codes, and because they constitute two complementary approaches to SQP algorithms: SNO... |

1 |
Global convergence of trust-region SQP- lter algorithms for nonlinear programming
- Toint
(Show Context)
Citation Context ...ve-set quadratic programming solvers to carry information from one active-set to the next. Global convergence results have been established for variations of the lter mechanism implemented in lterSQP =-=[13]-=-, but this is still an ongoing area of research. 3 Testing Environment We performed tests using 509 problems from the CUTE collection [2] that have been translated into AMPL [16] by Benson and Vanderb... |