Mercurial > hg > octave-lyh
view scripts/optimization/__dogleg__.m @ 8986:22c8272af34b
improvements to fsolve & family
author | Jaroslav Hajek <highegg@gmail.com> |
---|---|
date | Tue, 17 Mar 2009 08:49:08 +0100 |
parents | eb63fbe60fab |
children | b7210faa3ed0 |
line wrap: on
line source
## Copyright (C) 2008, 2009 Jaroslav Hajek ## ## This file is part of Octave. ## ## Octave is free software; you can redistribute it and/or modify it ## under the terms of the GNU General Public License as published by ## the Free Software Foundation; either version 3 of the License, or (at ## your option) any later version. ## ## Octave is distributed in the hope that it will be useful, but ## WITHOUT ANY WARRANTY; without even the implied warranty of ## MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU ## General Public License for more details. ## ## You should have received a copy of the GNU General Public License ## along with Octave; see the file COPYING. If not, see ## <http://www.gnu.org/licenses/>. ## -*- texinfo -*- ## @deftypefn{Function File} {@var{x}} = __dogleg__ (@var{r}, @var{b}, @var{x}, @var{d}, @var{delta}) ## Undocumented internal function. ## @end deftypefn ## Solve the double dogleg trust-region problem: ## Minimize norm(r*x-b) subject to the constraint norm(d.*x) <= delta, ## x being a convex combination of the gauss-newton and scaled gradient. ## TODO: error checks ## TODO: handle singularity, or leave it up to mldivide? function x = __dogleg__ (r, b, d, delta) ## Get Gauss-Newton direction. x = r \ b; xn = norm (d .* x); if (xn > delta) ## GN is too big, get scaled gradient. s = (r' * b) ./ d; sn = norm (s); if (sn > 0) ## Normalize and rescale. s = (s / sn) ./ d; ## Get the line minimizer in s direction. tn = norm (r*s); snm = (sn / tn) / tn; if (snm < delta) ## Get the dogleg path minimizer. bn = norm (b); dxn = delta/xn; snmd = snm/delta; t = (bn/sn) * (bn/xn) * snmd; t -= dxn * snmd^2 - sqrt ((t-dxn)^2 + (1-dxn^2)*(1-snmd^2)); alpha = dxn*(1-snmd^2) / t; else alpha = 0; endif else alpha = delta / xn; snm = 0; endif ## Form the appropriate convex combination. x = alpha * x + ((1-alpha) * min (snm, delta)) * s; endif endfunction