TY - JOUR
T1 - Tackling box-constrained optimization via a new projected quasi-newton approach
AU - Kim, Dongmin
AU - Sra, Suvrit
AU - Dhillon, Inderjit S.
PY - 2010
Y1 - 2010
N2 - Numerous scientific applications across a variety of fields depend on box-constrained convex optimization. Box-constrained problems therefore continue to attract research interest. We address box-constrained (strictly convex) problems by deriving two new quasi-Newton algorithms. Our algorithms are positioned between the projected-gradient [J. B. Rosen, J. SIAM, 8 (1960), pp. 181-217] and projected-Newton [D. P. Bertsekas, SIAM J. Control Optim., 20 (1982), pp. 221- 246] methods. We also prove their convergence under a simple Armijo step-size rule. We provide experimental results for two particular box-constrained problems: nonnegative least squares (NNLS), and nonnegative Kullback-Leibler (NNKL) minimization. For both NNLS and NNKL our algorithms perform competitively as compared to well-established methods on medium-sized problems; for larger problems our approach frequently outperforms the competition.
AB - Numerous scientific applications across a variety of fields depend on box-constrained convex optimization. Box-constrained problems therefore continue to attract research interest. We address box-constrained (strictly convex) problems by deriving two new quasi-Newton algorithms. Our algorithms are positioned between the projected-gradient [J. B. Rosen, J. SIAM, 8 (1960), pp. 181-217] and projected-Newton [D. P. Bertsekas, SIAM J. Control Optim., 20 (1982), pp. 221- 246] methods. We also prove their convergence under a simple Armijo step-size rule. We provide experimental results for two particular box-constrained problems: nonnegative least squares (NNLS), and nonnegative Kullback-Leibler (NNKL) minimization. For both NNLS and NNKL our algorithms perform competitively as compared to well-established methods on medium-sized problems; for larger problems our approach frequently outperforms the competition.
KW - Box-constrained convex optimization
KW - Kullback-Leibler divergence minimization
KW - Nonnegative least squares
KW - Projected-Newton methods
KW - Quasi-Newton
UR - http://www.scopus.com/inward/record.url?scp=79251469431&partnerID=8YFLogxK
U2 - 10.1137/08073812X
DO - 10.1137/08073812X
M3 - Article
AN - SCOPUS:79251469431
SN - 1064-8275
VL - 32
SP - 3548
EP - 3563
JO - SIAM Journal on Scientific Computing
JF - SIAM Journal on Scientific Computing
IS - 6
ER -