src/glpscl.c
changeset 1 c445c931472f
     1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/src/glpscl.c	Mon Dec 06 13:09:21 2010 +0100
     1.3 @@ -0,0 +1,476 @@
     1.4 +/* glpscl.c */
     1.5 +
     1.6 +/***********************************************************************
     1.7 +*  This code is part of GLPK (GNU Linear Programming Kit).
     1.8 +*
     1.9 +*  Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008,
    1.10 +*  2009, 2010 Andrew Makhorin, Department for Applied Informatics,
    1.11 +*  Moscow Aviation Institute, Moscow, Russia. All rights reserved.
    1.12 +*  E-mail: <mao@gnu.org>.
    1.13 +*
    1.14 +*  GLPK is free software: you can redistribute it and/or modify it
    1.15 +*  under the terms of the GNU General Public License as published by
    1.16 +*  the Free Software Foundation, either version 3 of the License, or
    1.17 +*  (at your option) any later version.
    1.18 +*
    1.19 +*  GLPK is distributed in the hope that it will be useful, but WITHOUT
    1.20 +*  ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
    1.21 +*  or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
    1.22 +*  License for more details.
    1.23 +*
    1.24 +*  You should have received a copy of the GNU General Public License
    1.25 +*  along with GLPK. If not, see <http://www.gnu.org/licenses/>.
    1.26 +***********************************************************************/
    1.27 +
    1.28 +#include "glpapi.h"
    1.29 +
    1.30 +/***********************************************************************
    1.31 +*  min_row_aij - determine minimal |a[i,j]| in i-th row
    1.32 +*
    1.33 +*  This routine returns minimal magnitude of (non-zero) constraint
    1.34 +*  coefficients in i-th row of the constraint matrix.
    1.35 +*
    1.36 +*  If the parameter scaled is zero, the original constraint matrix A is
    1.37 +*  assumed. Otherwise, the scaled constraint matrix R*A*S is assumed.
    1.38 +*
    1.39 +*  If i-th row of the matrix is empty, the routine returns 1. */
    1.40 +
    1.41 +static double min_row_aij(glp_prob *lp, int i, int scaled)
    1.42 +{     GLPAIJ *aij;
    1.43 +      double min_aij, temp;
    1.44 +      xassert(1 <= i && i <= lp->m);
    1.45 +      min_aij = 1.0;
    1.46 +      for (aij = lp->row[i]->ptr; aij != NULL; aij = aij->r_next)
    1.47 +      {  temp = fabs(aij->val);
    1.48 +         if (scaled) temp *= (aij->row->rii * aij->col->sjj);
    1.49 +         if (aij->r_prev == NULL || min_aij > temp)
    1.50 +            min_aij = temp;
    1.51 +      }
    1.52 +      return min_aij;
    1.53 +}
    1.54 +
    1.55 +/***********************************************************************
    1.56 +*  max_row_aij - determine maximal |a[i,j]| in i-th row
    1.57 +*
    1.58 +*  This routine returns maximal magnitude of (non-zero) constraint
    1.59 +*  coefficients in i-th row of the constraint matrix.
    1.60 +*
    1.61 +*  If the parameter scaled is zero, the original constraint matrix A is
    1.62 +*  assumed. Otherwise, the scaled constraint matrix R*A*S is assumed.
    1.63 +*
    1.64 +*  If i-th row of the matrix is empty, the routine returns 1. */
    1.65 +
    1.66 +static double max_row_aij(glp_prob *lp, int i, int scaled)
    1.67 +{     GLPAIJ *aij;
    1.68 +      double max_aij, temp;
    1.69 +      xassert(1 <= i && i <= lp->m);
    1.70 +      max_aij = 1.0;
    1.71 +      for (aij = lp->row[i]->ptr; aij != NULL; aij = aij->r_next)
    1.72 +      {  temp = fabs(aij->val);
    1.73 +         if (scaled) temp *= (aij->row->rii * aij->col->sjj);
    1.74 +         if (aij->r_prev == NULL || max_aij < temp)
    1.75 +            max_aij = temp;
    1.76 +      }
    1.77 +      return max_aij;
    1.78 +}
    1.79 +
    1.80 +/***********************************************************************
    1.81 +*  min_col_aij - determine minimal |a[i,j]| in j-th column
    1.82 +*
    1.83 +*  This routine returns minimal magnitude of (non-zero) constraint
    1.84 +*  coefficients in j-th column of the constraint matrix.
    1.85 +*
    1.86 +*  If the parameter scaled is zero, the original constraint matrix A is
    1.87 +*  assumed. Otherwise, the scaled constraint matrix R*A*S is assumed.
    1.88 +*
    1.89 +*  If j-th column of the matrix is empty, the routine returns 1. */
    1.90 +
    1.91 +static double min_col_aij(glp_prob *lp, int j, int scaled)
    1.92 +{     GLPAIJ *aij;
    1.93 +      double min_aij, temp;
    1.94 +      xassert(1 <= j && j <= lp->n);
    1.95 +      min_aij = 1.0;
    1.96 +      for (aij = lp->col[j]->ptr; aij != NULL; aij = aij->c_next)
    1.97 +      {  temp = fabs(aij->val);
    1.98 +         if (scaled) temp *= (aij->row->rii * aij->col->sjj);
    1.99 +         if (aij->c_prev == NULL || min_aij > temp)
   1.100 +            min_aij = temp;
   1.101 +      }
   1.102 +      return min_aij;
   1.103 +}
   1.104 +
   1.105 +/***********************************************************************
   1.106 +*  max_col_aij - determine maximal |a[i,j]| in j-th column
   1.107 +*
   1.108 +*  This routine returns maximal magnitude of (non-zero) constraint
   1.109 +*  coefficients in j-th column of the constraint matrix.
   1.110 +*
   1.111 +*  If the parameter scaled is zero, the original constraint matrix A is
   1.112 +*  assumed. Otherwise, the scaled constraint matrix R*A*S is assumed.
   1.113 +*
   1.114 +*  If j-th column of the matrix is empty, the routine returns 1. */
   1.115 +
   1.116 +static double max_col_aij(glp_prob *lp, int j, int scaled)
   1.117 +{     GLPAIJ *aij;
   1.118 +      double max_aij, temp;
   1.119 +      xassert(1 <= j && j <= lp->n);
   1.120 +      max_aij = 1.0;
   1.121 +      for (aij = lp->col[j]->ptr; aij != NULL; aij = aij->c_next)
   1.122 +      {  temp = fabs(aij->val);
   1.123 +         if (scaled) temp *= (aij->row->rii * aij->col->sjj);
   1.124 +         if (aij->c_prev == NULL || max_aij < temp)
   1.125 +            max_aij = temp;
   1.126 +      }
   1.127 +      return max_aij;
   1.128 +}
   1.129 +
   1.130 +/***********************************************************************
   1.131 +*  min_mat_aij - determine minimal |a[i,j]| in constraint matrix
   1.132 +*
   1.133 +*  This routine returns minimal magnitude of (non-zero) constraint
   1.134 +*  coefficients in the constraint matrix.
   1.135 +*
   1.136 +*  If the parameter scaled is zero, the original constraint matrix A is
   1.137 +*  assumed. Otherwise, the scaled constraint matrix R*A*S is assumed.
   1.138 +*
   1.139 +*  If the matrix is empty, the routine returns 1. */
   1.140 +
   1.141 +static double min_mat_aij(glp_prob *lp, int scaled)
   1.142 +{     int i;
   1.143 +      double min_aij, temp;
   1.144 +      min_aij = 1.0;
   1.145 +      for (i = 1; i <= lp->m; i++)
   1.146 +      {  temp = min_row_aij(lp, i, scaled);
   1.147 +         if (i == 1 || min_aij > temp)
   1.148 +            min_aij = temp;
   1.149 +      }
   1.150 +      return min_aij;
   1.151 +}
   1.152 +
   1.153 +/***********************************************************************
   1.154 +*  max_mat_aij - determine maximal |a[i,j]| in constraint matrix
   1.155 +*
   1.156 +*  This routine returns maximal magnitude of (non-zero) constraint
   1.157 +*  coefficients in the constraint matrix.
   1.158 +*
   1.159 +*  If the parameter scaled is zero, the original constraint matrix A is
   1.160 +*  assumed. Otherwise, the scaled constraint matrix R*A*S is assumed.
   1.161 +*
   1.162 +*  If the matrix is empty, the routine returns 1. */
   1.163 +
   1.164 +static double max_mat_aij(glp_prob *lp, int scaled)
   1.165 +{     int i;
   1.166 +      double max_aij, temp;
   1.167 +      max_aij = 1.0;
   1.168 +      for (i = 1; i <= lp->m; i++)
   1.169 +      {  temp = max_row_aij(lp, i, scaled);
   1.170 +         if (i == 1 || max_aij < temp)
   1.171 +            max_aij = temp;
   1.172 +      }
   1.173 +      return max_aij;
   1.174 +}
   1.175 +
   1.176 +/***********************************************************************
   1.177 +*  eq_scaling - perform equilibration scaling
   1.178 +*
   1.179 +*  This routine performs equilibration scaling of rows and columns of
   1.180 +*  the constraint matrix.
   1.181 +*
   1.182 +*  If the parameter flag is zero, the routine scales rows at first and
   1.183 +*  then columns. Otherwise, the routine scales columns and then rows.
   1.184 +*
   1.185 +*  Rows are scaled as follows:
   1.186 +*
   1.187 +*                         n
   1.188 +*     a'[i,j] = a[i,j] / max |a[i,j]|,  i = 1,...,m.
   1.189 +*                        j=1
   1.190 +*
   1.191 +*  This makes the infinity (maximum) norm of each row of the matrix
   1.192 +*  equal to 1.
   1.193 +*
   1.194 +*  Columns are scaled as follows:
   1.195 +*
   1.196 +*                         n
   1.197 +*     a'[i,j] = a[i,j] / max |a[i,j]|,  j = 1,...,n.
   1.198 +*                        i=1
   1.199 +*
   1.200 +*  This makes the infinity (maximum) norm of each column of the matrix
   1.201 +*  equal to 1. */
   1.202 +
   1.203 +static void eq_scaling(glp_prob *lp, int flag)
   1.204 +{     int i, j, pass;
   1.205 +      double temp;
   1.206 +      xassert(flag == 0 || flag == 1);
   1.207 +      for (pass = 0; pass <= 1; pass++)
   1.208 +      {  if (pass == flag)
   1.209 +         {  /* scale rows */
   1.210 +            for (i = 1; i <= lp->m; i++)
   1.211 +            {  temp = max_row_aij(lp, i, 1);
   1.212 +               glp_set_rii(lp, i, glp_get_rii(lp, i) / temp);
   1.213 +            }
   1.214 +         }
   1.215 +         else
   1.216 +         {  /* scale columns */
   1.217 +            for (j = 1; j <= lp->n; j++)
   1.218 +            {  temp = max_col_aij(lp, j, 1);
   1.219 +               glp_set_sjj(lp, j, glp_get_sjj(lp, j) / temp);
   1.220 +            }
   1.221 +         }
   1.222 +      }
   1.223 +      return;
   1.224 +}
   1.225 +
   1.226 +/***********************************************************************
   1.227 +*  gm_scaling - perform geometric mean scaling
   1.228 +*
   1.229 +*  This routine performs geometric mean scaling of rows and columns of
   1.230 +*  the constraint matrix.
   1.231 +*
   1.232 +*  If the parameter flag is zero, the routine scales rows at first and
   1.233 +*  then columns. Otherwise, the routine scales columns and then rows.
   1.234 +*
   1.235 +*  Rows are scaled as follows:
   1.236 +*
   1.237 +*     a'[i,j] = a[i,j] / sqrt(alfa[i] * beta[i]),  i = 1,...,m,
   1.238 +*
   1.239 +*  where:
   1.240 +*                n                        n
   1.241 +*     alfa[i] = min |a[i,j]|,  beta[i] = max |a[i,j]|.
   1.242 +*               j=1                      j=1
   1.243 +*
   1.244 +*  This allows decreasing the ratio beta[i] / alfa[i] for each row of
   1.245 +*  the matrix.
   1.246 +*
   1.247 +*  Columns are scaled as follows:
   1.248 +*
   1.249 +*     a'[i,j] = a[i,j] / sqrt(alfa[j] * beta[j]),  j = 1,...,n,
   1.250 +*
   1.251 +*  where:
   1.252 +*                m                        m
   1.253 +*     alfa[j] = min |a[i,j]|,  beta[j] = max |a[i,j]|.
   1.254 +*               i=1                      i=1
   1.255 +*
   1.256 +*  This allows decreasing the ratio beta[j] / alfa[j] for each column
   1.257 +*  of the matrix. */
   1.258 +
   1.259 +static void gm_scaling(glp_prob *lp, int flag)
   1.260 +{     int i, j, pass;
   1.261 +      double temp;
   1.262 +      xassert(flag == 0 || flag == 1);
   1.263 +      for (pass = 0; pass <= 1; pass++)
   1.264 +      {  if (pass == flag)
   1.265 +         {  /* scale rows */
   1.266 +            for (i = 1; i <= lp->m; i++)
   1.267 +            {  temp = min_row_aij(lp, i, 1) * max_row_aij(lp, i, 1);
   1.268 +               glp_set_rii(lp, i, glp_get_rii(lp, i) / sqrt(temp));
   1.269 +            }
   1.270 +         }
   1.271 +         else
   1.272 +         {  /* scale columns */
   1.273 +            for (j = 1; j <= lp->n; j++)
   1.274 +            {  temp = min_col_aij(lp, j, 1) * max_col_aij(lp, j, 1);
   1.275 +               glp_set_sjj(lp, j, glp_get_sjj(lp, j) / sqrt(temp));
   1.276 +            }
   1.277 +         }
   1.278 +      }
   1.279 +      return;
   1.280 +}
   1.281 +
   1.282 +/***********************************************************************
   1.283 +*  max_row_ratio - determine worst scaling "quality" for rows
   1.284 +*
   1.285 +*  This routine returns the worst scaling "quality" for rows of the
   1.286 +*  currently scaled constraint matrix:
   1.287 +*
   1.288 +*              m
   1.289 +*     ratio = max ratio[i],
   1.290 +*             i=1
   1.291 +*  where:
   1.292 +*                 n              n
   1.293 +*     ratio[i] = max |a[i,j]| / min |a[i,j]|,  1 <= i <= m,
   1.294 +*                j=1            j=1
   1.295 +*
   1.296 +*  is the scaling "quality" of i-th row. */
   1.297 +
   1.298 +static double max_row_ratio(glp_prob *lp)
   1.299 +{     int i;
   1.300 +      double ratio, temp;
   1.301 +      ratio = 1.0;
   1.302 +      for (i = 1; i <= lp->m; i++)
   1.303 +      {  temp = max_row_aij(lp, i, 1) / min_row_aij(lp, i, 1);
   1.304 +         if (i == 1 || ratio < temp) ratio = temp;
   1.305 +      }
   1.306 +      return ratio;
   1.307 +}
   1.308 +
   1.309 +/***********************************************************************
   1.310 +*  max_col_ratio - determine worst scaling "quality" for columns
   1.311 +*
   1.312 +*  This routine returns the worst scaling "quality" for columns of the
   1.313 +*  currently scaled constraint matrix:
   1.314 +*
   1.315 +*              n
   1.316 +*     ratio = max ratio[j],
   1.317 +*             j=1
   1.318 +*  where:
   1.319 +*                 m              m
   1.320 +*     ratio[j] = max |a[i,j]| / min |a[i,j]|,  1 <= j <= n,
   1.321 +*                i=1            i=1
   1.322 +*
   1.323 +*  is the scaling "quality" of j-th column. */
   1.324 +
   1.325 +static double max_col_ratio(glp_prob *lp)
   1.326 +{     int j;
   1.327 +      double ratio, temp;
   1.328 +      ratio = 1.0;
   1.329 +      for (j = 1; j <= lp->n; j++)
   1.330 +      {  temp = max_col_aij(lp, j, 1) / min_col_aij(lp, j, 1);
   1.331 +         if (j == 1 || ratio < temp) ratio = temp;
   1.332 +      }
   1.333 +      return ratio;
   1.334 +}
   1.335 +
   1.336 +/***********************************************************************
   1.337 +*  gm_iterate - perform iterative geometric mean scaling
   1.338 +*
   1.339 +*  This routine performs iterative geometric mean scaling of rows and
   1.340 +*  columns of the constraint matrix.
   1.341 +*
   1.342 +*  The parameter it_max specifies the maximal number of iterations.
   1.343 +*  Recommended value of it_max is 15.
   1.344 +*
   1.345 +*  The parameter tau specifies a minimal improvement of the scaling
   1.346 +*  "quality" on each iteration, 0 < tau < 1. It means than the scaling
   1.347 +*  process continues while the following condition is satisfied:
   1.348 +*
   1.349 +*     ratio[k] <= tau * ratio[k-1],
   1.350 +*
   1.351 +*  where ratio = max |a[i,j]| / min |a[i,j]| is the scaling "quality"
   1.352 +*  to be minimized, k is the iteration number. Recommended value of tau
   1.353 +*  is 0.90. */
   1.354 +
   1.355 +static void gm_iterate(glp_prob *lp, int it_max, double tau)
   1.356 +{     int k, flag;
   1.357 +      double ratio = 0.0, r_old;
   1.358 +      /* if the scaling "quality" for rows is better than for columns,
   1.359 +         the rows are scaled first; otherwise, the columns are scaled
   1.360 +         first */
   1.361 +      flag = (max_row_ratio(lp) > max_col_ratio(lp));
   1.362 +      for (k = 1; k <= it_max; k++)
   1.363 +      {  /* save the scaling "quality" from previous iteration */
   1.364 +         r_old = ratio;
   1.365 +         /* determine the current scaling "quality" */
   1.366 +         ratio = max_mat_aij(lp, 1) / min_mat_aij(lp, 1);
   1.367 +#if 0
   1.368 +         xprintf("k = %d; ratio = %g\n", k, ratio);
   1.369 +#endif
   1.370 +         /* if improvement is not enough, terminate scaling */
   1.371 +         if (k > 1 && ratio > tau * r_old) break;
   1.372 +         /* otherwise, perform another iteration */
   1.373 +         gm_scaling(lp, flag);
   1.374 +      }
   1.375 +      return;
   1.376 +}
   1.377 +
   1.378 +/***********************************************************************
   1.379 +*  NAME
   1.380 +*
   1.381 +*  scale_prob - scale problem data
   1.382 +*
   1.383 +*  SYNOPSIS
   1.384 +*
   1.385 +*  #include "glpscl.h"
   1.386 +*  void scale_prob(glp_prob *lp, int flags);
   1.387 +*
   1.388 +*  DESCRIPTION
   1.389 +*
   1.390 +*  The routine scale_prob performs automatic scaling of problem data
   1.391 +*  for the specified problem object. */
   1.392 +
   1.393 +static void scale_prob(glp_prob *lp, int flags)
   1.394 +{     static const char *fmt =
   1.395 +         "%s: min|aij| = %10.3e  max|aij| = %10.3e  ratio = %10.3e\n";
   1.396 +      double min_aij, max_aij, ratio;
   1.397 +      xprintf("Scaling...\n");
   1.398 +      /* cancel the current scaling effect */
   1.399 +      glp_unscale_prob(lp);
   1.400 +      /* report original scaling "quality" */
   1.401 +      min_aij = min_mat_aij(lp, 1);
   1.402 +      max_aij = max_mat_aij(lp, 1);
   1.403 +      ratio = max_aij / min_aij;
   1.404 +      xprintf(fmt, " A", min_aij, max_aij, ratio);
   1.405 +      /* check if the problem is well scaled */
   1.406 +      if (min_aij >= 0.10 && max_aij <= 10.0)
   1.407 +      {  xprintf("Problem data seem to be well scaled\n");
   1.408 +         /* skip scaling, if required */
   1.409 +         if (flags & GLP_SF_SKIP) goto done;
   1.410 +      }
   1.411 +      /* perform iterative geometric mean scaling, if required */
   1.412 +      if (flags & GLP_SF_GM)
   1.413 +      {  gm_iterate(lp, 15, 0.90);
   1.414 +         min_aij = min_mat_aij(lp, 1);
   1.415 +         max_aij = max_mat_aij(lp, 1);
   1.416 +         ratio = max_aij / min_aij;
   1.417 +         xprintf(fmt, "GM", min_aij, max_aij, ratio);
   1.418 +      }
   1.419 +      /* perform equilibration scaling, if required */
   1.420 +      if (flags & GLP_SF_EQ)
   1.421 +      {  eq_scaling(lp, max_row_ratio(lp) > max_col_ratio(lp));
   1.422 +         min_aij = min_mat_aij(lp, 1);
   1.423 +         max_aij = max_mat_aij(lp, 1);
   1.424 +         ratio = max_aij / min_aij;
   1.425 +         xprintf(fmt, "EQ", min_aij, max_aij, ratio);
   1.426 +      }
   1.427 +      /* round scale factors to nearest power of two, if required */
   1.428 +      if (flags & GLP_SF_2N)
   1.429 +      {  int i, j;
   1.430 +         for (i = 1; i <= lp->m; i++)
   1.431 +            glp_set_rii(lp, i, round2n(glp_get_rii(lp, i)));
   1.432 +         for (j = 1; j <= lp->n; j++)
   1.433 +            glp_set_sjj(lp, j, round2n(glp_get_sjj(lp, j)));
   1.434 +         min_aij = min_mat_aij(lp, 1);
   1.435 +         max_aij = max_mat_aij(lp, 1);
   1.436 +         ratio = max_aij / min_aij;
   1.437 +         xprintf(fmt, "2N", min_aij, max_aij, ratio);
   1.438 +      }
   1.439 +done: return;
   1.440 +}
   1.441 +
   1.442 +/***********************************************************************
   1.443 +*  NAME
   1.444 +*
   1.445 +*  glp_scale_prob - scale problem data
   1.446 +*
   1.447 +*  SYNOPSIS
   1.448 +*
   1.449 +*  void glp_scale_prob(glp_prob *lp, int flags);
   1.450 +*
   1.451 +*  DESCRIPTION
   1.452 +*
   1.453 +*  The routine glp_scale_prob performs automatic scaling of problem
   1.454 +*  data for the specified problem object.
   1.455 +*
   1.456 +*  The parameter flags specifies scaling options used by the routine.
   1.457 +*  Options can be combined with the bitwise OR operator and may be the
   1.458 +*  following:
   1.459 +*
   1.460 +*  GLP_SF_GM      perform geometric mean scaling;
   1.461 +*  GLP_SF_EQ      perform equilibration scaling;
   1.462 +*  GLP_SF_2N      round scale factors to nearest power of two;
   1.463 +*  GLP_SF_SKIP    skip scaling, if the problem is well scaled.
   1.464 +*
   1.465 +*  The parameter flags may be specified as GLP_SF_AUTO, in which case
   1.466 +*  the routine chooses scaling options automatically. */
   1.467 +
   1.468 +void glp_scale_prob(glp_prob *lp, int flags)
   1.469 +{     if (flags & ~(GLP_SF_GM | GLP_SF_EQ | GLP_SF_2N | GLP_SF_SKIP |
   1.470 +                    GLP_SF_AUTO))
   1.471 +         xerror("glp_scale_prob: flags = 0x%02X; invalid scaling option"
   1.472 +            "s\n", flags);
   1.473 +      if (flags & GLP_SF_AUTO)
   1.474 +         flags = (GLP_SF_GM | GLP_SF_EQ | GLP_SF_SKIP);
   1.475 +      scale_prob(lp, flags);
   1.476 +      return;
   1.477 +}
   1.478 +
   1.479 +/* eof */