/* glpios11.c (process cuts stored in the local cut pool) */ /*********************************************************************** * This code is part of GLPK (GNU Linear Programming Kit). * Copyright (C) 2005-2018 Free Software Foundation, Inc. * Written by Andrew Makhorin . * * GLPK 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. * * GLPK 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 GLPK. If not, see . ***********************************************************************/ #include "draft.h" #include "env.h" #include "ios.h" /*********************************************************************** * NAME * * ios_process_cuts - process cuts stored in the local cut pool * * SYNOPSIS * * #include "glpios.h" * void ios_process_cuts(glp_tree *T); * * DESCRIPTION * * The routine ios_process_cuts analyzes each cut currently stored in * the local cut pool, which must be non-empty, and either adds the cut * to the current subproblem or just discards it. All cuts are assumed * to be locally valid. On exit the local cut pool remains unchanged. * * REFERENCES * * 1. E.Balas, S.Ceria, G.Cornuejols, "Mixed 0-1 Programming by * Lift-and-Project in a Branch-and-Cut Framework", Management Sc., * 42 (1996) 1229-1246. * * 2. G.Andreello, A.Caprara, and M.Fischetti, "Embedding Cuts in * a Branch&Cut Framework: a Computational Study with {0,1/2}-Cuts", * Preliminary Draft, October 28, 2003, pp.6-8. */ struct info { /* estimated cut efficiency */ IOSCUT *cut; /* pointer to cut in the cut pool */ char flag; /* if this flag is set, the cut is included into the current subproblem */ double eff; /* cut efficacy (normalized residual) */ double deg; /* lower bound to objective degradation */ }; static int CDECL fcmp(const void *arg1, const void *arg2) { const struct info *info1 = arg1, *info2 = arg2; if (info1->deg == 0.0 && info2->deg == 0.0) { if (info1->eff > info2->eff) return -1; if (info1->eff < info2->eff) return +1; } else { if (info1->deg > info2->deg) return -1; if (info1->deg < info2->deg) return +1; } return 0; } static double parallel(IOSCUT *a, IOSCUT *b, double work[]); #ifdef NEW_LOCAL /* 02/II-2018 */ void ios_process_cuts(glp_tree *T) { IOSPOOL *pool; IOSCUT *cut; GLPAIJ *aij; struct info *info; int k, kk, max_cuts, len, ret, *ind; double *val, *work, rhs; /* the current subproblem must exist */ xassert(T->curr != NULL); /* the pool must exist and be non-empty */ pool = T->local; xassert(pool != NULL); xassert(pool->m > 0); /* allocate working arrays */ info = xcalloc(1+pool->m, sizeof(struct info)); ind = xcalloc(1+T->n, sizeof(int)); val = xcalloc(1+T->n, sizeof(double)); work = xcalloc(1+T->n, sizeof(double)); for (k = 1; k <= T->n; k++) work[k] = 0.0; /* build the list of cuts stored in the cut pool */ for (k = 1; k <= pool->m; k++) info[k].cut = pool->row[k], info[k].flag = 0; /* estimate efficiency of all cuts in the cut pool */ for (k = 1; k <= pool->m; k++) { double temp, dy, dz; cut = info[k].cut; /* build the vector of cut coefficients and compute its Euclidean norm */ len = 0; temp = 0.0; for (aij = cut->ptr; aij != NULL; aij = aij->r_next) { xassert(1 <= aij->col->j && aij->col->j <= T->n); len++, ind[len] = aij->col->j, val[len] = aij->val; temp += aij->val * aij->val; } if (temp < DBL_EPSILON * DBL_EPSILON) temp = DBL_EPSILON; /* transform the cut to express it only through non-basic (auxiliary and structural) variables */ len = glp_transform_row(T->mip, len, ind, val); /* determine change in the cut value and in the objective value for the adjacent basis by simulating one step of the dual simplex */ switch (cut->type) { case GLP_LO: rhs = cut->lb; break; case GLP_UP: rhs = cut->ub; break; default: xassert(cut != cut); } ret = _glp_analyze_row(T->mip, len, ind, val, cut->type, rhs, 1e-9, NULL, NULL, NULL, NULL, &dy, &dz); /* determine normalized residual and lower bound to objective degradation */ if (ret == 0) { info[k].eff = fabs(dy) / sqrt(temp); /* if some reduced costs violates (slightly) their zero bounds (i.e. have wrong signs) due to round-off errors, dz also may have wrong sign being close to zero */ if (T->mip->dir == GLP_MIN) { if (dz < 0.0) dz = 0.0; info[k].deg = + dz; } else /* GLP_MAX */ { if (dz > 0.0) dz = 0.0; info[k].deg = - dz; } } else if (ret == 1) { /* the constraint is not violated at the current point */ info[k].eff = info[k].deg = 0.0; } else if (ret == 2) { /* no dual feasible adjacent basis exists */ info[k].eff = 1.0; info[k].deg = DBL_MAX; } else xassert(ret != ret); /* if the degradation is too small, just ignore it */ if (info[k].deg < 0.01) info[k].deg = 0.0; } /* sort the list of cuts by decreasing objective degradation and then by decreasing efficacy */ qsort(&info[1], pool->m, sizeof(struct info), fcmp); /* only first (most efficient) max_cuts in the list are qualified as candidates to be added to the current subproblem */ max_cuts = (T->curr->level == 0 ? 90 : 10); if (max_cuts > pool->m) max_cuts = pool->m; /* add cuts to the current subproblem */ #if 0 xprintf("*** adding cuts ***\n"); #endif for (k = 1; k <= max_cuts; k++) { int i, len; /* if this cut seems to be inefficient, skip it */ if (info[k].deg < 0.01 && info[k].eff < 0.01) continue; /* if the angle between this cut and every other cut included in the current subproblem is small, skip this cut */ for (kk = 1; kk < k; kk++) { if (info[kk].flag) { if (parallel(info[k].cut, info[kk].cut, work) > 0.90) break; } } if (kk < k) continue; /* add this cut to the current subproblem */ #if 0 xprintf("eff = %g; deg = %g\n", info[k].eff, info[k].deg); #endif cut = info[k].cut, info[k].flag = 1; i = glp_add_rows(T->mip, 1); if (cut->name != NULL) glp_set_row_name(T->mip, i, cut->name); xassert(T->mip->row[i]->origin == GLP_RF_CUT); T->mip->row[i]->klass = cut->klass; len = 0; for (aij = cut->ptr; aij != NULL; aij = aij->r_next) len++, ind[len] = aij->col->j, val[len] = aij->val; glp_set_mat_row(T->mip, i, len, ind, val); switch (cut->type) { case GLP_LO: rhs = cut->lb; break; case GLP_UP: rhs = cut->ub; break; default: xassert(cut != cut); } glp_set_row_bnds(T->mip, i, cut->type, rhs, rhs); } /* free working arrays */ xfree(info); xfree(ind); xfree(val); xfree(work); return; } #else void ios_process_cuts(glp_tree *T) { IOSPOOL *pool; IOSCUT *cut; IOSAIJ *aij; struct info *info; int k, kk, max_cuts, len, ret, *ind; double *val, *work; /* the current subproblem must exist */ xassert(T->curr != NULL); /* the pool must exist and be non-empty */ pool = T->local; xassert(pool != NULL); xassert(pool->size > 0); /* allocate working arrays */ info = xcalloc(1+pool->size, sizeof(struct info)); ind = xcalloc(1+T->n, sizeof(int)); val = xcalloc(1+T->n, sizeof(double)); work = xcalloc(1+T->n, sizeof(double)); for (k = 1; k <= T->n; k++) work[k] = 0.0; /* build the list of cuts stored in the cut pool */ for (k = 0, cut = pool->head; cut != NULL; cut = cut->next) k++, info[k].cut = cut, info[k].flag = 0; xassert(k == pool->size); /* estimate efficiency of all cuts in the cut pool */ for (k = 1; k <= pool->size; k++) { double temp, dy, dz; cut = info[k].cut; /* build the vector of cut coefficients and compute its Euclidean norm */ len = 0; temp = 0.0; for (aij = cut->ptr; aij != NULL; aij = aij->next) { xassert(1 <= aij->j && aij->j <= T->n); len++, ind[len] = aij->j, val[len] = aij->val; temp += aij->val * aij->val; } if (temp < DBL_EPSILON * DBL_EPSILON) temp = DBL_EPSILON; /* transform the cut to express it only through non-basic (auxiliary and structural) variables */ len = glp_transform_row(T->mip, len, ind, val); /* determine change in the cut value and in the objective value for the adjacent basis by simulating one step of the dual simplex */ ret = _glp_analyze_row(T->mip, len, ind, val, cut->type, cut->rhs, 1e-9, NULL, NULL, NULL, NULL, &dy, &dz); /* determine normalized residual and lower bound to objective degradation */ if (ret == 0) { info[k].eff = fabs(dy) / sqrt(temp); /* if some reduced costs violates (slightly) their zero bounds (i.e. have wrong signs) due to round-off errors, dz also may have wrong sign being close to zero */ if (T->mip->dir == GLP_MIN) { if (dz < 0.0) dz = 0.0; info[k].deg = + dz; } else /* GLP_MAX */ { if (dz > 0.0) dz = 0.0; info[k].deg = - dz; } } else if (ret == 1) { /* the constraint is not violated at the current point */ info[k].eff = info[k].deg = 0.0; } else if (ret == 2) { /* no dual feasible adjacent basis exists */ info[k].eff = 1.0; info[k].deg = DBL_MAX; } else xassert(ret != ret); /* if the degradation is too small, just ignore it */ if (info[k].deg < 0.01) info[k].deg = 0.0; } /* sort the list of cuts by decreasing objective degradation and then by decreasing efficacy */ qsort(&info[1], pool->size, sizeof(struct info), fcmp); /* only first (most efficient) max_cuts in the list are qualified as candidates to be added to the current subproblem */ max_cuts = (T->curr->level == 0 ? 90 : 10); if (max_cuts > pool->size) max_cuts = pool->size; /* add cuts to the current subproblem */ #if 0 xprintf("*** adding cuts ***\n"); #endif for (k = 1; k <= max_cuts; k++) { int i, len; /* if this cut seems to be inefficient, skip it */ if (info[k].deg < 0.01 && info[k].eff < 0.01) continue; /* if the angle between this cut and every other cut included in the current subproblem is small, skip this cut */ for (kk = 1; kk < k; kk++) { if (info[kk].flag) { if (parallel(info[k].cut, info[kk].cut, work) > 0.90) break; } } if (kk < k) continue; /* add this cut to the current subproblem */ #if 0 xprintf("eff = %g; deg = %g\n", info[k].eff, info[k].deg); #endif cut = info[k].cut, info[k].flag = 1; i = glp_add_rows(T->mip, 1); if (cut->name != NULL) glp_set_row_name(T->mip, i, cut->name); xassert(T->mip->row[i]->origin == GLP_RF_CUT); T->mip->row[i]->klass = cut->klass; len = 0; for (aij = cut->ptr; aij != NULL; aij = aij->next) len++, ind[len] = aij->j, val[len] = aij->val; glp_set_mat_row(T->mip, i, len, ind, val); xassert(cut->type == GLP_LO || cut->type == GLP_UP); glp_set_row_bnds(T->mip, i, cut->type, cut->rhs, cut->rhs); } /* free working arrays */ xfree(info); xfree(ind); xfree(val); xfree(work); return; } #endif #if 0 /*********************************************************************** * Given a cut a * x >= b (<= b) the routine efficacy computes the cut * efficacy as follows: * * eff = d * (a * x~ - b) / ||a||, * * where d is -1 (in case of '>= b') or +1 (in case of '<= b'), x~ is * the vector of values of structural variables in optimal solution to * LP relaxation of the current subproblem, ||a|| is the Euclidean norm * of the vector of cut coefficients. * * If the cut is violated at point x~, the efficacy eff is positive, * and its value is the Euclidean distance between x~ and the cut plane * a * x = b in the space of structural variables. * * Following geometrical intuition, it is quite natural to consider * this distance as a first-order measure of the expected efficacy of * the cut: the larger the distance the better the cut [1]. */ static double efficacy(glp_tree *T, IOSCUT *cut) { glp_prob *mip = T->mip; IOSAIJ *aij; double s = 0.0, t = 0.0, temp; for (aij = cut->ptr; aij != NULL; aij = aij->next) { xassert(1 <= aij->j && aij->j <= mip->n); s += aij->val * mip->col[aij->j]->prim; t += aij->val * aij->val; } temp = sqrt(t); if (temp < DBL_EPSILON) temp = DBL_EPSILON; if (cut->type == GLP_LO) temp = (s >= cut->rhs ? 0.0 : (cut->rhs - s) / temp); else if (cut->type == GLP_UP) temp = (s <= cut->rhs ? 0.0 : (s - cut->rhs) / temp); else xassert(cut != cut); return temp; } #endif /*********************************************************************** * Given two cuts a1 * x >= b1 (<= b1) and a2 * x >= b2 (<= b2) the * routine parallel computes the cosine of angle between the cut planes * a1 * x = b1 and a2 * x = b2 (which is the acute angle between two * normals to these planes) in the space of structural variables as * follows: * * cos phi = (a1' * a2) / (||a1|| * ||a2||), * * where (a1' * a2) is a dot product of vectors of cut coefficients, * ||a1|| and ||a2|| are Euclidean norms of vectors a1 and a2. * * Note that requirement cos phi = 0 forces the cuts to be orthogonal, * i.e. with disjoint support, while requirement cos phi <= 0.999 means * only avoiding duplicate (parallel) cuts [1]. */ #ifdef NEW_LOCAL /* 02/II-2018 */ static double parallel(IOSCUT *a, IOSCUT *b, double work[]) { GLPAIJ *aij; double s = 0.0, sa = 0.0, sb = 0.0, temp; for (aij = a->ptr; aij != NULL; aij = aij->r_next) { work[aij->col->j] = aij->val; sa += aij->val * aij->val; } for (aij = b->ptr; aij != NULL; aij = aij->r_next) { s += work[aij->col->j] * aij->val; sb += aij->val * aij->val; } for (aij = a->ptr; aij != NULL; aij = aij->r_next) work[aij->col->j] = 0.0; temp = sqrt(sa) * sqrt(sb); if (temp < DBL_EPSILON * DBL_EPSILON) temp = DBL_EPSILON; return s / temp; } #else static double parallel(IOSCUT *a, IOSCUT *b, double work[]) { IOSAIJ *aij; double s = 0.0, sa = 0.0, sb = 0.0, temp; for (aij = a->ptr; aij != NULL; aij = aij->next) { work[aij->j] = aij->val; sa += aij->val * aij->val; } for (aij = b->ptr; aij != NULL; aij = aij->next) { s += work[aij->j] * aij->val; sb += aij->val * aij->val; } for (aij = a->ptr; aij != NULL; aij = aij->next) work[aij->j] = 0.0; temp = sqrt(sa) * sqrt(sb); if (temp < DBL_EPSILON * DBL_EPSILON) temp = DBL_EPSILON; return s / temp; } #endif /* eof */