/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ /* */ /* This file is part of the program and library */ /* SCIP --- Solving Constraint Integer Programs */ /* */ /* Copyright 2002-2022 Zuse Institute Berlin */ /* */ /* Licensed under the Apache License, Version 2.0 (the "License"); */ /* you may not use this file except in compliance with the License. */ /* You may obtain a copy of the License at */ /* */ /* http://www.apache.org/licenses/LICENSE-2.0 */ /* */ /* Unless required by applicable law or agreed to in writing, software */ /* distributed under the License is distributed on an "AS IS" BASIS, */ /* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. */ /* See the License for the specific language governing permissions and */ /* limitations under the License. */ /* */ /* You should have received a copy of the Apache-2.0 license */ /* along with SCIP; see the file LICENSE. If not visit scipopt.org. */ /* */ /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ /**@file branch_distribution.h * @ingroup BRANCHINGRULES * @brief probability based branching rule based on an article by J. Pryor and J.W. Chinneck * @author Gregor Hendel * * The distribution branching rule selects a variable based on its impact on row activity distributions. More formally, * let \f$ a(x) = a_1 x_1 + \dots + a_n x_n \leq b \f$ be a row of the LP. Let further \f$ l_i, u_i \in R\f$ denote the * (finite) lower and upper bound, respectively, of the \f$ i \f$-th variable \f$x_i\f$. * Viewing every variable \f$x_i \f$ as (continuously) uniformly distributed within its bounds, we can approximately * understand the row activity \f$a(x)\f$ as a gaussian random variate with mean value \f$ \mu = E[a(x)] = \sum_i a_i\frac{l_i + u_i}{2}\f$ * and variance \f$ \sigma^2 = \sum_i a_i^2 \sigma_i^2 \f$, with \f$ \sigma_i^2 = \frac{(u_i - l_i)^2}{12}\f$ for * continuous and \f$ \sigma_i^2 = \frac{(u_i - l_i + 1)^2 - 1}{12}\f$ for discrete variables. * With these two parameters, we can calculate the probability to satisfy the row in terms of the cumulative distribution * of the standard normal distribution: \f$ P(a(x) \leq b) = \Phi(\frac{b - \mu}{\sigma})\f$. * * The impact of a variable on the probability to satisfy a constraint after branching can be estimated by altering * the variable contribution to the sums described above. In order to keep the tree size small, * variables are preferred which decrease the probability * to satisfy a row because it is more likely to create infeasible subproblems. * * The selection of the branching variable is subject to the parameter @p scoreparam. For both branching directions, * an individual score is calculated. Available options for scoring methods are: * - @b d: select a branching variable with largest difference in satisfaction probability after branching * - @b l: lowest cumulative probability amongst all variables on all rows (after branching). * - @b h: highest cumulative probability amongst all variables on all rows (after branching). * - @b v: highest number of votes for lowering row probability for all rows a variable appears in. * - @b w: highest number of votes for increasing row probability for all rows a variable appears in. * * If the parameter @p usescipscore is set to @a TRUE, a single branching score is calculated from the respective * up and down scores as defined by the SCIP branching score function (see advanced branching parameter @p scorefunc), * otherwise, the variable with the single highest score is selected, and the maximizing direction is assigned * higher branching priority. * * The original idea of probability based branching schemes appeared in: * * J. Pryor and J.W. Chinneck:@n * Faster Integer-Feasibility in Mixed-Integer Linear Programs by Branching to Force Change@n * Computers and Operations Research, vol. 38, 2011, p. 1143–1152@n * (http://www.sce.carleton.ca/faculty/chinneck/docs/PryorChinneck.pdf) */ /*---+----1----+----2----+----3----+----4----+----5----+----6----+----7----+----8----+----9----+----0----+----1----+----2*/ #ifndef __SCIP_BRANCH_DISTRIBUTION_H__ #define __SCIP_BRANCH_DISTRIBUTION_H__ #include "scip/def.h" #include "scip/type_lp.h" #include "scip/type_retcode.h" #include "scip/type_scip.h" #include "scip/type_var.h" #ifdef __cplusplus extern "C" { #endif /** creates the distribution branching rule and includes it in SCIP * * @ingroup BranchingRuleIncludes */ SCIP_EXPORT SCIP_RETCODE SCIPincludeBranchruleDistribution( SCIP* scip /**< SCIP data structure */ ); /**@addtogroup BRANCHINGRULES * * @{ */ /** calculate the variable's distribution parameters (mean and variance) for the bounds specified in the arguments. * special treatment of infinite bounds necessary */ SCIP_EXPORT void SCIPvarCalcDistributionParameters( SCIP* scip, /**< SCIP data structure */ SCIP_Real varlb, /**< variable lower bound */ SCIP_Real varub, /**< variable upper bound */ SCIP_VARTYPE vartype, /**< type of the variable */ SCIP_Real* mean, /**< pointer to store mean value */ SCIP_Real* variance /**< pointer to store the variance of the variable uniform distribution */ ); /** calculates the cumulative distribution P(-infinity <= x <= value) that a normally distributed * random variable x takes a value between -infinity and parameter \p value. * * The distribution is given by the respective mean and deviation. This implementation * uses the error function erf(). */ SCIP_EXPORT SCIP_Real SCIPcalcCumulativeDistribution( SCIP* scip, /**< current SCIP */ SCIP_Real mean, /**< the mean value of the distribution */ SCIP_Real variance, /**< the square of the deviation of the distribution */ SCIP_Real value /**< the upper limit of the calculated distribution integral */ ); /** calculates the probability of satisfying an LP-row under the assumption * of uniformly distributed variable values. * * For inequalities, we use the cumulative distribution function of the standard normal * distribution PHI(rhs - mu/sqrt(sigma2)) to calculate the probability * for a right hand side row with mean activity mu and variance sigma2 to be satisfied. * Similarly, 1 - PHI(lhs - mu/sqrt(sigma2)) is the probability to satisfy a left hand side row. * For equations (lhs==rhs), we use the centeredness measure p = min(PHI(lhs'), 1-PHI(lhs'))/max(PHI(lhs'), 1 - PHI(lhs')), * where lhs' = lhs - mu / sqrt(sigma2). */ SCIP_EXPORT SCIP_Real SCIProwCalcProbability( SCIP* scip, /**< current scip */ SCIP_ROW* row, /**< the row */ SCIP_Real mu, /**< the mean value of the row distribution */ SCIP_Real sigma2, /**< the variance of the row distribution */ int rowinfinitiesdown, /**< the number of variables with infinite bounds to DECREASE activity */ int rowinfinitiesup /**< the number of variables with infinite bounds to INCREASE activity */ ); /** update the up- and downscore of a single variable after calculating the impact of branching on a * particular row, depending on the chosen score parameter */ SCIP_EXPORT SCIP_RETCODE SCIPupdateDistributionScore( SCIP* scip, /**< current SCIP pointer */ SCIP_Real currentprob, /**< the current probability */ SCIP_Real newprobup, /**< the new probability if branched upwards */ SCIP_Real newprobdown, /**< the new probability if branched downwards */ SCIP_Real* upscore, /**< pointer to store the new score for branching up */ SCIP_Real* downscore, /**< pointer to store the new score for branching down */ char scoreparam /**< parameter to determine the way the score is calculated */ ); /** @} */ #ifdef __cplusplus } #endif #endif