prop_nlobbt.c

Go to the documentation of this file.

/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
/*                                                                           */
/*                  This file is part of the program and library             */
/*         SCIP --- Solving Constraint Integer Programs                      */
/*                                                                           */
/*    Copyright (C) 2002-2020 Konrad-Zuse-Zentrum                            */
/*                            fuer Informationstechnik Berlin                */
/*                                                                           */
/*  SCIP is distributed under the terms of the ZIB Academic License.         */
/*                                                                           */
/*  You should have received a copy of the ZIB Academic License              */
/*  along with SCIP; see the file COPYING. If not visit scipopt.org.         */
/*                                                                           */
/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */

/**@file   prop_nlobbt.c
 * @ingroup DEFPLUGINS_PROP
 * @brief  nlobbt propagator
 * @author Benjamin Mueller
 */

/*---+----1----+----2----+----3----+----4----+----5----+----6----+----7----+----8----+----9----+----0----+----1----+----2*/

#include "blockmemshell/memory.h"
#include "nlpi/nlpi.h"
#include "scip/prop_genvbounds.h"
#include "scip/prop_nlobbt.h"
#include "scip/pub_message.h"
#include "scip/pub_misc.h"
#include "scip/pub_misc_sort.h"
#include "scip/pub_nlp.h"
#include "scip/pub_prop.h"
#include "scip/pub_tree.h"
#include "scip/pub_var.h"
#include "scip/scip_general.h"
#include "scip/scip_lp.h"
#include "scip/scip_mem.h"
#include "scip/scip_message.h"
#include "scip/scip_nlp.h"
#include "scip/scip_nonlinear.h"
#include "scip/scip_numerics.h"
#include "scip/scip_param.h"
#include "scip/scip_prob.h"
#include "scip/scip_probing.h"
#include "scip/scip_prop.h"
#include "scip/scip_randnumgen.h"
#include "scip/scip_solvingstats.h"
#include "scip/scip_timing.h"
#include "scip/scip_tree.h"
#include "scip/scip_var.h"
#include <string.h>

#define PROP_NAME              "nlobbt"
#define PROP_DESC              "propagator template"
#define PROP_PRIORITY          -1100000
#define PROP_FREQ                    -1
#define PROP_DELAY                 TRUE
#define PROP_TIMING            SCIP_PROPTIMING_AFTERLPLOOP

#define DEFAULT_MINNONCONVEXFRAC   0.20      /**< default minimum (# convex nlrows)/(# nonconvex nlrows) threshold to apply propagator */
#define DEFAULT_MINLINEARFRAC      0.02      /**< default minimum (# convex nlrows)/(# linear nlrows) threshold to apply propagator */
#define DEFAULT_FEASTOLFAC         0.01      /**< default factor for NLP feasibility tolerance */
#define DEFAULT_RELOBJTOLFAC       0.01      /**< default factor for NLP relative objective tolerance */
#define DEFAULT_ADDLPROWS          TRUE      /**< should (non-initial) LP rows be used? */
#define DEFAULT_ITLIMITFACTOR       2.0      /**< multiple of root node LP iterations used as total LP iteration
                                              *   limit for nlobbt (<= 0: no limit ) */
#define DEFAULT_NLPITERLIMIT        500      /**< default iteration limit of NLP solver; 0 for no limit */
#define DEFAULT_NLPTIMELIMIT        0.0      /**< default time limit of NLP solver; 0.0 for no limit */
#define DEFAULT_NLPVERLEVEL           0      /**< verbosity level of NLP solver */
#define DEFAULT_RANDSEED             79      /**< initial random seed */

/*
 * Data structures
 */

/** status of bound candidates */
enum BoundStatus
{
   UNSOLVED = 1,                             /**< did not solve LB or UB problem */
   SOLVEDLB = 2,                             /**< solved LB problem */
   SOLVEDUB = 4                              /**< solved UB problem */
};

/** propagator data */
struct SCIP_PropData
{
   SCIP_NLPI*            nlpi;               /**< nlpi used to create the nlpi problem */
   SCIP_NLPIPROBLEM*     nlpiprob;           /**< nlpi problem representing the convex NLP relaxation */
   SCIP_HASHMAP*         var2nlpiidx;        /**< mapping between variables and nlpi indices */
   SCIP_VAR**            nlpivars;           /**< array containing all variables of the nlpi */
   int                   nlpinvars;          /**< total number of nlpi variables */
   SCIP_Real*            nlscore;            /**< score for each nonlinear variable */
   int*                  status;             /**< array containing a bound status for each candidate (type int* is
                                              *   necessary to use sort functions) */
   SCIP_PROP*            genvboundprop;      /**< genvbound propagator */
   SCIP_RANDNUMGEN*      randnumgen;         /**< random number generator */
   SCIP_Bool             skipprop;           /**< should the propagator be skipped? */
   SCIP_Longint          lastnode;           /**< number of last node where obbt was performed */
   int                   currpos;            /**< current position in the nlpivars array */

   int                   nlpiterlimit;       /**< iteration limit of NLP solver; 0 for no limit */
   SCIP_Real             nlptimelimit;       /**< time limit of NLP solver; 0.0 for no limit */
   int                   nlpverblevel;       /**< verbosity level of NLP solver */
   SCIP_NLPSTATISTICS*   nlpstatistics;      /**< statistics from NLP solver */

   SCIP_Real             feastolfac;         /**< factor for NLP feasibility tolerance */
   SCIP_Real             relobjtolfac;       /**< factor for NLP relative objective tolerance */
   SCIP_Real             minnonconvexfrac;   /**< minimum (#convex nlrows)/(#nonconvex nlrows) threshold to apply propagator */
   SCIP_Real             minlinearfrac;      /**< minimum (#convex nlrows)/(#linear nlrows) threshold to apply propagator */
   SCIP_Bool             addlprows;          /**< should (non-initial) LP rows be used? */
   SCIP_Real             itlimitfactor;      /**< LP iteration limit for nlobbt will be this factor times total LP
                                              *   iterations in root node */
};

/*
 * Local methods
 */

/** clears the propagator data */
static
SCIP_RETCODE propdataClear(
   SCIP*                 scip,               /**< SCIP data structure */
   SCIP_PROPDATA*        propdata            /**< propagator data */
   )
{
   assert(propdata != NULL);

   if( propdata->nlpiprob != NULL )
   {
      assert(propdata->nlpi != NULL);

      SCIPfreeBlockMemoryArray(scip, &propdata->status, propdata->nlpinvars);
      SCIPfreeBlockMemoryArray(scip, &propdata->nlscore, propdata->nlpinvars);
      SCIPfreeBlockMemoryArray(scip, &propdata->nlpivars, propdata->nlpinvars);
      SCIPhashmapFree(&propdata->var2nlpiidx);
      SCIP_CALL( SCIPnlpiFreeProblem(propdata->nlpi, &propdata->nlpiprob) );

      propdata->nlpinvars = 0;
   }
   assert(propdata->nlpinvars == 0);

   propdata->skipprop = FALSE;
   propdata->currpos = 0;
   propdata->lastnode = -1;

   return SCIP_OKAY;
}

/** checks whether it is worth to call nonlinear OBBT procedure */
static
SCIP_Bool isNlobbtApplicable(
   SCIP*                 scip,               /**< SCIP data structure */
   SCIP_PROPDATA*        propdata            /**< propagation data */
   )
{
   SCIP_NLROW** nlrows;
   int nnonconvexnlrows;
   int nconvexnlrows;
   int nlinearnlrows;
   int nnlrows;
   int i;

   nlrows = SCIPgetNLPNlRows(scip);
   nnlrows = SCIPgetNNLPNlRows(scip);
   nnonconvexnlrows = 0;
   nconvexnlrows = 0;
   nlinearnlrows = 0;

   for( i = 0; i < nnlrows; ++i )
   {
      if( SCIPnlrowGetNQuadElems(nlrows[i]) == 0 && SCIPnlrowGetExprtree(nlrows[i]) == NULL )
         ++nlinearnlrows;
      else if( SCIPnlrowGetCurvature(nlrows[i]) == SCIP_EXPRCURV_CONVEX )
      {
         if( !SCIPisInfinity(scip, SCIPnlrowGetRhs(nlrows[i])) )
            ++nconvexnlrows;
         if( !SCIPisInfinity(scip, -SCIPnlrowGetLhs(nlrows[i])) )
            ++nnonconvexnlrows;
      }
      else if( SCIPnlrowGetCurvature(nlrows[i]) == SCIP_EXPRCURV_CONCAVE )
      {
         if( !SCIPisInfinity(scip, SCIPnlrowGetRhs(nlrows[i])) )
            ++nnonconvexnlrows;
         if( !SCIPisInfinity(scip, -SCIPnlrowGetLhs(nlrows[i])) )
            ++nconvexnlrows;
      }
      else
      {
         if( !SCIPisInfinity(scip, SCIPnlrowGetRhs(nlrows[i])) )
            ++nnonconvexnlrows;
         if( !SCIPisInfinity(scip, -SCIPnlrowGetLhs(nlrows[i])) )
            ++nnonconvexnlrows;
      }
   }

   SCIPdebugMsg(scip, "nconvex=%d nnonconvex=%d nlinear=%d\n", nconvexnlrows, nnonconvexnlrows, nlinearnlrows);

   return nconvexnlrows > 0
      && nnonconvexnlrows > 0
      && (SCIPisGE(scip, (SCIP_Real)nconvexnlrows, nnonconvexnlrows * propdata->minnonconvexfrac))
      && (SCIPisGE(scip, (SCIP_Real)nconvexnlrows, nlinearnlrows * propdata->minlinearfrac));
}

/** filters variables which achieve their lower or dual bound in the current NLP solution */
static
SCIP_RETCODE filterCands(
   SCIP*                 scip,               /**< SCIP data structure */
   SCIP_PROPDATA*        propdata            /**< propagator data */
   )
{
   SCIP_Real* primal;
   int i;

   assert(propdata->currpos >= 0 && propdata->currpos < propdata->nlpinvars);
   assert(SCIPnlpiGetSolstat(propdata->nlpi, propdata->nlpiprob) <= SCIP_NLPSOLSTAT_FEASIBLE);

   SCIP_CALL( SCIPnlpiGetSolution(propdata->nlpi, propdata->nlpiprob, &primal, NULL, NULL, NULL, NULL) );
   assert(primal != NULL);

   /* we skip all candidates which have been processed already, i.e., starting at propdata->currpos + 1 */
   for( i = propdata->currpos + 1; i < propdata->nlpinvars; ++i )
   {
      SCIP_VAR* var;
      SCIP_Real val;
      int varidx;

      /* only uninteresting variables left -> stop filtering */
      if( SCIPisLE(scip, propdata->nlscore[i], 0.0) )
         break;

      var = propdata->nlpivars[i];
      assert(var != NULL && SCIPhashmapExists(propdata->var2nlpiidx, (void*)var));

      varidx = SCIPhashmapGetImageInt(propdata->var2nlpiidx, (void*)var);
      assert(SCIPgetVars(scip)[varidx] == var);
      val = primal[varidx];

      if( (propdata->status[i] & SOLVEDLB) == 0 && !SCIPisInfinity(scip, -val) /*lint !e641*/
         && SCIPisFeasLE(scip, val, SCIPvarGetLbLocal(var)) )
      {
         SCIPdebugMsg(scip, "filter LB of %s in [%g,%g] with %g\n", SCIPvarGetName(var), SCIPvarGetLbLocal(var),
            SCIPvarGetUbLocal(var), val);
         propdata->status[i] |= SOLVEDLB; /*lint !e641*/
         assert((propdata->status[i] & SOLVEDLB) != 0); /*lint !e641*/
      }

      if( (propdata->status[i] & SOLVEDUB) == 0 && !SCIPisInfinity(scip, val) /*lint !e641*/
         && SCIPisFeasGE(scip, val, SCIPvarGetUbLocal(var)) )
      {
         SCIPdebugMsg(scip, "filter UB of %s in [%g,%g] with %g\n", SCIPvarGetName(var), SCIPvarGetLbLocal(var),
            SCIPvarGetUbLocal(var), val);
         propdata->status[i] |= SOLVEDUB; /*lint !e641*/
         assert((propdata->status[i] & SOLVEDUB) != 0); /*lint !e641*/
      }
   }

   return SCIP_OKAY;
}

/** tries to add a generalized variable bound by exploiting the dual solution of the last NLP solve (see @ref
 *  prop_nlobbt.h for more information)
 */
static
SCIP_RETCODE addGenVBound(
   SCIP*                 scip,               /**< SCIP data structure */
   SCIP_PROPDATA*        propdata,           /**< propagator data */
   SCIP_VAR*             var,                /**< variable used in last NLP solve */
   int                   varidx,             /**< variable index in the propdata->nlpivars array */
   SCIP_BOUNDTYPE        boundtype,          /**< type of bound provided by the genvbound */
   SCIP_Real             cutoffbound         /**< cutoff bound */
   )
{
   SCIP_VAR** lvbvars;
   SCIP_Real* lvbcoefs;
   SCIP_Real* primal;
   SCIP_Real* dual;
   SCIP_Real* alpha;
   SCIP_Real* beta;
   SCIP_Real constant;
   SCIP_Real mu;
   int nlvbvars;
   int i;

   assert(propdata->genvboundprop != NULL);
   assert(var != NULL);
   assert(varidx >= 0 && varidx < propdata->nlpinvars);
   assert(SCIPnlpiGetSolstat(propdata->nlpi, propdata->nlpiprob) <= SCIP_NLPSOLSTAT_LOCOPT);

   SCIP_CALL( SCIPnlpiGetSolution(propdata->nlpi, propdata->nlpiprob, &primal, &dual, &alpha, &beta, NULL) );

   /* not possible to generate genvbound if the duals for the propagated variable do not disappear */
   if( !SCIPisFeasZero(scip, alpha[varidx] - beta[varidx]) )
      return SCIP_OKAY;

   SCIP_CALL( SCIPallocBufferArray(scip, &lvbcoefs, propdata->nlpinvars) );
   SCIP_CALL( SCIPallocBufferArray(scip, &lvbvars, propdata->nlpinvars) );
   constant = boundtype == SCIP_BOUNDTYPE_LOWER ? primal[varidx] : -primal[varidx];
   mu = 0.0;
   nlvbvars = 0;

   /* collect coefficients of genvbound */
   for( i = 0; i < propdata->nlpinvars; ++i )
   {
      if( !SCIPisZero(scip, beta[i] - alpha[i]) )
      {
         lvbvars[nlvbvars] = propdata->nlpivars[i];
         lvbcoefs[nlvbvars] = beta[i] - alpha[i];
         ++nlvbvars;

         constant += (alpha[i] - beta[i]) * primal[i];
      }
   }

   /* first dual multiplier corresponds to the cutoff row if cutoffbound < SCIPinfinity() */
   if( !SCIPisInfinity(scip, cutoffbound) && SCIPisGT(scip, dual[0], 0.0) )
   {
      mu = dual[0];
      constant += mu * cutoffbound;
   }

   /* add genvbound to genvbounds propagator */
   if( !SCIPisInfinity(scip, REALABS(constant)) && (nlvbvars > 0 || SCIPisFeasGT(scip, mu, 0.0)) )
   {
      SCIP_CALL( SCIPgenVBoundAdd(scip, propdata->genvboundprop, lvbvars, var, lvbcoefs, nlvbvars, -mu, constant,
            boundtype) );
      SCIPdebugMsg(scip, "add genvbound for %s\n", SCIPvarGetName(var));
   }

   SCIPfreeBufferArray(scip, &lvbvars);
   SCIPfreeBufferArray(scip, &lvbcoefs);

   return SCIP_OKAY;
}

/** sets the objective function, solves the NLP, and tightens the given variable; after calling this function, the
 *  objective function is set to zero
 *
 *  @note function assumes that objective function is zero
 */
static
SCIP_RETCODE solveNlp(
   SCIP*                 scip,               /**< SCIP data structure */
   SCIP_PROPDATA*        propdata,           /**< propagator data */
   SCIP_VAR*             var,                /**< variable to propagate */
   int                   varidx,             /**< variable index in the propdata->nlpivars array */
   SCIP_BOUNDTYPE        boundtype,          /**< minimize or maximize var? */
   int*                  nlpiter,            /**< buffer to store the total number of nlp iterations */
   SCIP_RESULT*          result              /**< pointer to store result */
   )
{
   SCIP_Real timelimit;
   SCIP_Real* primal;
   SCIP_Real obj;
   int iterlimit;

#ifdef SCIP_DEBUG
   SCIP_Real oldlb;
   SCIP_Real oldub;

   oldlb = SCIPvarGetLbLocal(var);
   oldub = SCIPvarGetUbLocal(var);
#endif

   assert(var != NULL);
   assert(varidx >= 0 && varidx < propdata->nlpinvars);
   assert(result != NULL && *result != SCIP_CUTOFF);

   *nlpiter = 0;

   /* set time and iteration limit */
   SCIP_CALL( SCIPgetRealParam(scip, "limits/time", &timelimit) );
   if( !SCIPisInfinity(scip, timelimit) )
   {
      timelimit -= SCIPgetSolvingTime(scip);
      if( timelimit <= 0.0 )
      {
         SCIPdebugMsg(scip, "skip NLP solve; no time left\n");
         return SCIP_OKAY;
      }
   }
   if( propdata->nlptimelimit > 0.0 )
      timelimit = MIN(propdata->nlptimelimit, timelimit);
   iterlimit = propdata->nlpiterlimit > 0 ? propdata->nlpiterlimit : INT_MAX;
   SCIP_CALL( SCIPnlpiSetRealPar(propdata->nlpi, propdata->nlpiprob, SCIP_NLPPAR_TILIM, timelimit) );
   SCIP_CALL( SCIPnlpiSetIntPar(propdata->nlpi, propdata->nlpiprob, SCIP_NLPPAR_ITLIM, iterlimit) );

   /* set corresponding objective coefficient and solve NLP */
   obj = boundtype == SCIP_BOUNDTYPE_LOWER ? 1.0 : -1.0;
   SCIP_CALL( SCIPnlpiSetObjective(propdata->nlpi, propdata->nlpiprob, 1, &varidx, &obj, 0, NULL, NULL, NULL, 0.0) );

   SCIPdebugMsg(scip, "solve var=%s boundtype=%d nlscore=%g\n", SCIPvarGetName(var), boundtype,
      propdata->nlscore[propdata->currpos]);
   SCIP_CALL( SCIPnlpiSolve(propdata->nlpi, propdata->nlpiprob) );
   SCIPdebugMsg(scip, "NLP solstat = %d\n", SCIPnlpiGetSolstat(propdata->nlpi, propdata->nlpiprob));

   /* collect NLP statistics */
   assert(propdata->nlpstatistics != NULL);
   SCIP_CALL( SCIPnlpiGetStatistics(propdata->nlpi, propdata->nlpiprob, propdata->nlpstatistics) );
   *nlpiter = SCIPnlpStatisticsGetNIterations(propdata->nlpstatistics);
   SCIPdebugMsg(scip, "iterations %d time %g\n", *nlpiter, SCIPnlpStatisticsGetTotalTime(propdata->nlpstatistics));

   /* filter bound candidates first, otherwise we do not have access to the primal solution values */
   if( SCIPnlpiGetSolstat(propdata->nlpi, propdata->nlpiprob) <= SCIP_NLPSOLSTAT_FEASIBLE )
   {
      SCIP_CALL( filterCands(scip, propdata) );
   }

   /* try to tighten variable bound */
   if( SCIPnlpiGetSolstat(propdata->nlpi, propdata->nlpiprob) <= SCIP_NLPSOLSTAT_LOCOPT )
   {
      SCIP_Bool tightened;
      SCIP_Bool infeasible;

      /* try to add a genvbound in the root node */
      if( propdata->genvboundprop != NULL && SCIPgetDepth(scip) == 0 )
      {
         SCIP_CALL( addGenVBound(scip, propdata, var, varidx, boundtype, SCIPgetCutoffbound(scip)) );
      }

      SCIP_CALL( SCIPnlpiGetSolution(propdata->nlpi, propdata->nlpiprob, &primal, NULL, NULL, NULL, NULL) );

      if( boundtype == SCIP_BOUNDTYPE_LOWER )
      {
         SCIP_CALL( SCIPtightenVarLb(scip, var, primal[varidx], FALSE, &infeasible, &tightened) );
      }
      else
      {
         SCIP_CALL( SCIPtightenVarUb(scip, var, primal[varidx], FALSE, &infeasible, &tightened) );
      }

      if( infeasible )
      {
         SCIPdebugMsg(scip, "detect infeasibility after propagating %s\n", SCIPvarGetName(var));
         *result = SCIP_CUTOFF;
      }
      else if( tightened )
      {
         SCIP_Real lb;
         SCIP_Real ub;

         *result = SCIP_REDUCEDDOM;

         /* update bounds in NLP */
         lb = SCIPvarGetLbLocal(var);
         ub = SCIPvarGetUbLocal(var);
         SCIP_CALL( SCIPnlpiChgVarBounds(propdata->nlpi, propdata->nlpiprob, 1, &varidx, &lb, &ub) );

#ifdef SCIP_DEBUG
         SCIPdebugMsg(scip, "tightened bounds of %s from [%g,%g] to [%g,%g]\n", SCIPvarGetName(var), oldlb, oldub, lb, ub);
#endif
      }
   }

   /* reset objective function */
   obj = 0.0;
   SCIP_CALL( SCIPnlpiSetObjective(propdata->nlpi, propdata->nlpiprob, 1, &varidx, &obj, 0, NULL, NULL, NULL, 0.0) );

   return SCIP_OKAY;
}

/** main method of the propagator
 *
 *  creates a convex NLP relaxation and solves the OBBT-NLPs for each possible candidate;
 *  binary and variables with a small domain will be ignored to reduce the computational cost of the propagator; after
 *  solving each NLP we filter out all variable candidates which are on their lower or upper bound; candidates with a
 *  larger number of occurrences are preferred
 */
static
SCIP_RETCODE applyNlobbt(
   SCIP*                 scip,               /**< SCIP data structure */
   SCIP_PROPDATA*        propdata,           /**< propagation data */
   SCIP_RESULT*          result              /**< pointer to store result */
   )
{
   int nlpiterleft;

   assert(result != NULL);
   assert(!propdata->skipprop);
   assert(SCIPgetNNlpis(scip) > 0);

   *result = SCIP_DIDNOTRUN;

   if( propdata->nlpiprob == NULL && !isNlobbtApplicable(scip, propdata) )
   {
      /* do not call the propagator anymore (except after a restart) */
      SCIPdebugMsg(scip, "nlobbt propagator is not applicable\n");
      propdata->skipprop = TRUE;
      return SCIP_OKAY;
   }

   *result = SCIP_DIDNOTFIND;

   /* compute NLP iteration limit */
   if( propdata->itlimitfactor > 0.0 )
      nlpiterleft = (int)(propdata->itlimitfactor * SCIPgetNRootLPIterations(scip));
   else
      nlpiterleft = INT_MAX;

   /* recompute NLP relaxation if the variable set changed */
   if( propdata->nlpiprob != NULL && SCIPgetNVars(scip) != propdata->nlpinvars )
   {
      SCIP_CALL( propdataClear(scip, propdata) );
      assert(propdata->nlpiprob == NULL);
   }

   /* create or update NLP relaxation */
   if( propdata->nlpiprob == NULL )
   {
      int i;

      propdata->nlpinvars = SCIPgetNVars(scip);
      propdata->nlpi = SCIPgetNlpis(scip)[0];
      assert(propdata->nlpi != NULL);

      SCIP_CALL( SCIPnlpiCreateProblem(propdata->nlpi, &propdata->nlpiprob, "nlobbt-nlp") );
      SCIP_CALL( SCIPhashmapCreate(&propdata->var2nlpiidx, SCIPblkmem(scip), propdata->nlpinvars) );
      SCIP_CALL( SCIPduplicateBlockMemoryArray(scip, &propdata->nlpivars, SCIPgetVars(scip), propdata->nlpinvars) ); /*lint !e666*/
      SCIP_CALL( SCIPallocBlockMemoryArray(scip, &propdata->nlscore, propdata->nlpinvars) );
      SCIP_CALL( SCIPallocBlockMemoryArray(scip, &propdata->status, propdata->nlpinvars) );

      SCIP_CALL( SCIPcreateNlpiProb(scip, propdata->nlpi, SCIPgetNLPNlRows(scip), SCIPgetNNLPNlRows(scip),
            propdata->nlpiprob, propdata->var2nlpiidx, NULL, propdata->nlscore, SCIPgetCutoffbound(scip), FALSE, TRUE) );

      /* initialize bound status; perturb nlscores by a factor which ensures that zero scores remain zero */
      assert(propdata->randnumgen != NULL);
      for( i = 0; i < propdata->nlpinvars; ++i )
      {
         propdata->status[i] = UNSOLVED; /*lint !e641*/
         propdata->nlscore[i] *= 1.0 + SCIPrandomGetReal(propdata->randnumgen, SCIPfeastol(scip), 2.0 * SCIPfeastol(scip));
      }

      /* add rows of the LP */
      if( SCIPgetDepth(scip) == 0 )
      {
         SCIP_CALL( SCIPaddNlpiProbRows(scip, propdata->nlpi, propdata->nlpiprob, propdata->var2nlpiidx,
               SCIPgetLPRows(scip), SCIPgetNLPRows(scip)) );
      }
   }
   else
   {
      SCIP_CALL( SCIPupdateNlpiProb(scip, propdata->nlpi, propdata->nlpiprob, propdata->var2nlpiidx,
            propdata->nlpivars, propdata->nlpinvars, SCIPgetCutoffbound(scip)) );
   }

   assert(propdata->nlpiprob != NULL);
   assert(propdata->var2nlpiidx != NULL);
   assert(propdata->nlpivars != NULL);
   assert(propdata->nlscore != NULL);

   /* sort variables w.r.t. their nlscores if we did not solve any NLP for this node */
   if( propdata->currpos == 0 )
   {
      SCIPsortDownRealIntPtr(propdata->nlscore, propdata->status, (void**)propdata->nlpivars, propdata->nlpinvars);
   }

   /* set parameters of NLP solver */
   SCIP_CALL( SCIPnlpiSetRealPar(propdata->nlpi, propdata->nlpiprob, SCIP_NLPPAR_FEASTOL,
         SCIPfeastol(scip) * propdata->feastolfac) );
   SCIP_CALL( SCIPnlpiSetRealPar(propdata->nlpi, propdata->nlpiprob, SCIP_NLPPAR_RELOBJTOL,
         SCIPfeastol(scip) * propdata->relobjtolfac) );
   SCIP_CALL( SCIPnlpiSetIntPar(propdata->nlpi, propdata->nlpiprob, SCIP_NLPPAR_VERBLEVEL, propdata->nlpverblevel) );

   /* main propagation loop */
   while( propdata->currpos < propdata->nlpinvars
      && nlpiterleft > 0
      && SCIPisGT(scip, propdata->nlscore[propdata->currpos], 0.0)
      && *result != SCIP_CUTOFF
      && !SCIPisStopped(scip) )
   {
      SCIP_VAR* var;
      int varidx;
      int iters;

      var = propdata->nlpivars[propdata->currpos];
      assert(var != NULL);

      /* skip binary or almost fixed variables */
      if( SCIPvarGetType(var) == SCIP_VARTYPE_BINARY
         || SCIPisFeasEQ(scip, SCIPvarGetLbLocal(var), SCIPvarGetUbLocal(var)) )
      {
         ++(propdata->currpos);
         continue;
      }

      SCIPdebugMsg(scip, "iterations left %d\n", nlpiterleft);

      /* get index of var in the nlpi */
      assert(SCIPhashmapExists(propdata->var2nlpiidx, (void*)var) );
      varidx = SCIPhashmapGetImageInt(propdata->var2nlpiidx, (void*)var);
      assert(var == SCIPgetVars(scip)[varidx]);

      /* case: minimize var */
      if( (propdata->status[propdata->currpos] & SOLVEDLB) == 0 ) /*lint !e641*/
      {
         SCIP_CALL( solveNlp(scip, propdata, var, varidx, SCIP_BOUNDTYPE_LOWER, &iters, result) );
         nlpiterleft -= iters;
      }

      /* case: maximize var */
      if( *result != SCIP_CUTOFF && (propdata->status[propdata->currpos] & SOLVEDUB) == 0 ) /*lint !e641*/
      {
         SCIP_CALL( solveNlp(scip, propdata, var, varidx, SCIP_BOUNDTYPE_UPPER, &iters, result) );
         nlpiterleft -= iters;
      }

      /* update the current position */
      ++(propdata->currpos);
   }

   return SCIP_OKAY;
}

/*
 * Callback methods of propagator
 */

/** destructor of propagator to free user data (called when SCIP is exiting) */
static
SCIP_DECL_PROPFREE(propFreeNlobbt)
{  /*lint --e{715}*/
   SCIP_PROPDATA* propdata;

   propdata = SCIPpropGetData(prop);
   assert(propdata != NULL);

   SCIP_CALL( propdataClear(scip, propdata) );
   SCIPfreeBlockMemory(scip, &propdata);
   SCIPpropSetData(prop, NULL);

   return SCIP_OKAY;
}

/** solving process initialization method of propagator (called when branch and bound process is about to begin) */
static
SCIP_DECL_PROPINITSOL(propInitsolNlobbt)
{  /*lint --e{715}*/
   SCIP_PROPDATA* propdata;

   assert(scip != NULL);
   assert(prop != NULL);

   propdata = SCIPpropGetData(prop);
   assert(propdata != NULL);

   /* if genvbounds propagator is not available, we cannot create genvbounds */
   propdata->genvboundprop = SCIPfindProp(scip, "genvbounds");

   SCIP_CALL( SCIPcreateRandom(scip, &propdata->randnumgen,
         DEFAULT_RANDSEED, TRUE) );
   SCIP_CALL( SCIPnlpStatisticsCreate(SCIPblkmem(scip), &propdata->nlpstatistics) );
   propdata->lastnode = -1;

   return SCIP_OKAY;
}

/** solving process deinitialization method of propagator (called before branch and bound process data is freed) */
static
SCIP_DECL_PROPEXITSOL(propExitsolNlobbt)
{  /*lint --e{715}*/
   SCIP_PROPDATA* propdata;

   propdata = SCIPpropGetData(prop);
   assert(propdata != NULL);

   SCIPnlpStatisticsFree(SCIPblkmem(scip), &propdata->nlpstatistics);
   SCIPfreeRandom(scip, &propdata->randnumgen);

   SCIP_CALL( propdataClear(scip, propdata) );

   return SCIP_OKAY;
}

/** execution method of propagator */
static
SCIP_DECL_PROPEXEC(propExecNlobbt)
{  /*lint --e{715}*/
   SCIP_PROPDATA* propdata;

   *result = SCIP_DIDNOTRUN;

   propdata = SCIPpropGetData(prop);
   assert(propdata != NULL);

   if( propdata->skipprop || SCIPgetStage(scip) != SCIP_STAGE_SOLVING || SCIPinRepropagation(scip)
      || SCIPinProbing(scip) || SCIPinDive(scip) || !SCIPallowWeakDualReds(scip) || SCIPgetNNlpis(scip) == 0 )
   {
      SCIPdebugMsg(scip, "skip nlobbt propagator\n");
      return SCIP_OKAY;
   }

   /* only run if LP all columns are in the LP, e.g., do not run if pricers are active */
   if( !SCIPallColsInLP(scip) )
   {
      SCIPdebugMsg(scip, "not all columns in LP, skipping obbt\n");
      return SCIP_OKAY;
   }

   /* do not run if SCIP does not have constructed an NLP */
   if( !SCIPisNLPConstructed(scip) )
   {
      SCIPdebugMsg(scip, "NLP not constructed, skipping nlobbt\n");
      return SCIP_OKAY;
   }

   /* consider all variables again if we process a new node */
   if( SCIPnodeGetNumber(SCIPgetCurrentNode(scip)) != propdata->lastnode )
   {
      propdata->lastnode = SCIPnodeGetNumber(SCIPgetCurrentNode(scip));
      propdata->currpos = 0;
   }

   /* call main procedure of nonlinear OBBT propagator */
   SCIP_CALL( applyNlobbt(scip, propdata, result) );

   return SCIP_OKAY;
}

/*
 * propagator specific interface methods
 */

/** creates the nlobbt propagator and includes it in SCIP */
SCIP_RETCODE SCIPincludePropNlobbt(
   SCIP*                 scip                /**< SCIP data structure */
   )
{
   SCIP_PROPDATA* propdata;
   SCIP_PROP* prop;

   propdata = NULL;
   prop = NULL;

   SCIP_CALL( SCIPallocBlockMemory(scip, &propdata) );
   assert(propdata != NULL);
   BMSclearMemory(propdata);

   SCIP_CALL( SCIPincludePropBasic(scip, &prop, PROP_NAME, PROP_DESC, PROP_PRIORITY, PROP_FREQ, PROP_DELAY, PROP_TIMING,
         propExecNlobbt, propdata) );
   assert(prop != NULL);

   SCIP_CALL( SCIPsetPropFree(scip, prop, propFreeNlobbt) );
   SCIP_CALL( SCIPsetPropInitsol(scip, prop, propInitsolNlobbt) );
   SCIP_CALL( SCIPsetPropExitsol(scip, prop, propExitsolNlobbt) );

   SCIP_CALL( SCIPaddRealParam(scip, "propagating/" PROP_NAME "/feastolfac",
         "factor for NLP feasibility tolerance",
         &propdata->feastolfac, TRUE, DEFAULT_FEASTOLFAC, 0.0, 1.0, NULL, NULL) );

   SCIP_CALL( SCIPaddRealParam(scip, "propagating/" PROP_NAME "/relobjtolfac",
         "factor for NLP relative objective tolerance",
         &propdata->relobjtolfac, TRUE, DEFAULT_RELOBJTOLFAC, 0.0, 1.0, NULL, NULL) );

   SCIP_CALL( SCIPaddRealParam(scip, "propagating/" PROP_NAME "/minnonconvexfrac",
         "(#convex nlrows)/(#nonconvex nlrows) threshold to apply propagator",
         &propdata->minnonconvexfrac, TRUE, DEFAULT_MINNONCONVEXFRAC, 0.0, SCIPinfinity(scip), NULL, NULL) );

   SCIP_CALL( SCIPaddRealParam(scip, "propagating/" PROP_NAME "/minlinearfrac",
         "minimum (#convex nlrows)/(#linear nlrows) threshold to apply propagator",
         &propdata->minlinearfrac, TRUE, DEFAULT_MINLINEARFRAC, 0.0, SCIPinfinity(scip), NULL, NULL) );

   SCIP_CALL( SCIPaddBoolParam(scip, "propagating/" PROP_NAME "/addlprows",
         "should non-initial LP rows be used?",
         &propdata->addlprows, FALSE, DEFAULT_ADDLPROWS, NULL, NULL) );

   SCIP_CALL( SCIPaddIntParam(scip, "propagating/" PROP_NAME "/nlpiterlimit",
         "iteration limit of NLP solver; 0 for no limit",
         &propdata->nlpiterlimit, TRUE, DEFAULT_NLPITERLIMIT, 0, INT_MAX, NULL, NULL) );

   SCIP_CALL( SCIPaddRealParam(scip, "propagating/" PROP_NAME "/nlptimelimit",
         "time limit of NLP solver; 0.0 for no limit",
         &propdata->nlptimelimit, TRUE, DEFAULT_NLPTIMELIMIT, 0.0, SCIP_REAL_MAX, NULL, NULL) );

   SCIP_CALL( SCIPaddIntParam(scip, "propagating/" PROP_NAME "/nlpverblevel",
         "verbosity level of NLP solver",
         &propdata->nlpverblevel, TRUE, DEFAULT_NLPVERLEVEL, 0, 5, NULL, NULL) );

   SCIP_CALL( SCIPaddRealParam(scip, "propagating/" PROP_NAME "/itlimitfactor",
         "LP iteration limit for nlobbt will be this factor times total LP iterations in root node",
         &propdata->itlimitfactor, TRUE, DEFAULT_ITLIMITFACTOR, 0.0, SCIP_REAL_MAX, NULL, NULL) );

   return SCIP_OKAY;
}