cgdma

cgdma − CUTEr CG_DESCENT test driver

cgdma

The cgdma main program test drives CG_DESCENT on SIF problems from the CUTEr distribution.

The CG_DESCENT package is a nonlinear congugate-gradient method for large-scale unconstrained minimization designed by William Hager and Hongchao Zhang (U. Florida).

The CG_DESCENT file cg_desecent.f should be compiled, and the resulting file cg_descent.o placed in (or symbolically linked to) the directory $MYCUTER/double/bin.

There is no single-precision version.

If no CG_DESCENT.SPC file is present in the current directory, the default version is copied from $CUTER/common/src/pkg/cg_descent/. Default specifications are as follows:

    1.0D-5    eps          The stopping tolerance
    1.0D-1    delta        Wolfe line search parameter
    9.0D-1    sigma        Wolfe line search parameter
    1.0D-6    epsilon      approximate Wolfe threshold factor
    5.0D-1    theta        update
    6.6D-1    gamma        required decay factor in interval
    5.0D+0    rho          growth factor in bracket
    1.0D-2    eta          lower bound for cg’s beta_k
    1.0D-2    psi0         factor used in very initial starting guess
    1.0D-1    psi1         factor previous step multiplied by in QuadStep
    2.0D+0    psi2         factor previous step is multipled by for startup
   1.0D-12    QuadCutOff   lower bound on rel change in f before QuadStep
    1.0D+0    restart_fac  restart cg in restart_fac*n iterations
    5.0D+2    maxit_fac    terminate in maxit_fac*n iterations
    0.0D+0    feps         stop when value change <= feps*|f|
        50    nexpand      number of grow/shrink allowed in bracket
        50    nsecant      number of secant steps allowed in line search
         T    QuadStep     use initial quad interpolation in line search
         F    PrintLevel   F (no print) T (intermediate results)
         F    PrintFinal   F (no print) T (print error messages, final error)
         F    StopRule     F (|grad|_infty <= tol) T (... <= tol*(1+|f|))
         F    ERule        F (eps_k = eps|f|) T (eps_k = eps)
         T    AWolfe       F (Wolfe) T (+approx Wolfe) 2 (epsilon pert)
         F    Step         F (no initial line search guess) T (guess in gnorm)

The reader is referred to the paper quoted below and the code itself if he or she wishes to modify these parameters.

CUTER

Parent directory for CUTEr

MYCUTER

Home directory of the installed CUTEr distribution.

I. Bongartz, A.R. Conn, N.I.M. Gould, D. Orban and Ph.L. Toint

CUTEr (and SifDec): A Constrained and Unconstrained Testing Environment, revisited, N.I.M. Gould, D. Orban and Ph.L. Toint, ACM TOMS, 29:4, pp.373-394, 2003.

CUTE: Constrained and Unconstrained Testing Environment, I. Bongartz, A.R. Conn, N.I.M. Gould and Ph.L. Toint, ACM TOMS, 21:1, pp.123-160, 1995.

CG_DESCENT, A conjugate gradient method with guaranteed descent, W. W. Hager and H. Zhang, Department of Mathematics, University of Florida, Gainesville, FL 32611, USA, January, 2004.

sdcgd(1), cgd(1).