linear_solver.h

// Copyright 2010-2022 Google LLC
// 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.

/**
 * \file
 * A C++ wrapper that provides a simple and unified interface to
 * several linear programming and mixed integer programming solvers:
 * GLOP, GLPK, CLP, CBC, and SCIP. The wrapper can also be used in Java, C#,
 * and Python via SWIG.
 *
 * What is Linear Programming?
 *
 *   In mathematics, linear programming (LP) is a technique for optimization of
 *   a linear objective function, subject to linear equality and linear
 *   inequality constraints. Informally, linear programming determines the way
 *   to achieve the best outcome (such as maximum profit or lowest cost) in a
 *   given mathematical model and given some list of requirements represented
 *   as linear equations.
 *
 *   The most widely used technique for solving a linear program is the Simplex
 *   algorithm, devised by George Dantzig in 1947. It performs very well on
 *   most instances, for which its running time is polynomial. A lot of effort
 *   has been put into improving the algorithm and its implementation. As a
 *   byproduct, it has however been shown that one can always construct
 *   problems that take exponential time for the Simplex algorithm to solve.
 *   Research has thus focused on trying to find a polynomial algorithm for
 *   linear programming, or to prove that linear programming is indeed
 *   polynomial.
 *
 *   Leonid Khachiyan first exhibited in 1979 a weakly polynomial algorithm for
 *   linear programming. "Weakly polynomial" means that the running time of the
 *   algorithm is in O(P(n) * 2^p) where P(n) is a polynomial of the size of the
 *   problem, and p is the precision of computations expressed in number of
 *   bits. With a fixed-precision, floating-point-based implementation, a weakly
 *   polynomial algorithm will  thus run in polynomial time. No implementation
 *   of Khachiyan's algorithm has proved efficient, but a larger breakthrough in
 *   the field came in 1984 when Narendra Karmarkar introduced a new interior
 *   point method for solving linear programming problems. Interior point
 *   algorithms have proved efficient on very large linear programs.
 *
 *   Check Wikipedia for more detail:
 *     http://en.wikipedia.org/wiki/Linear_programming
 *
 * -----------------------------------
 *
 * Example of a Linear Program
 *
 *   maximize:
 *     3x + y
 *   subject to:
 *     1.5 x + 2 y <= 12
 *     0 <= x <= 3
 *     0 <= y <= 5
 *
 *  A linear program has:
 *    1) a linear objective function
 *    2) linear constraints that can be equalities or inequalities
 *    3) bounds on variables that can be positive, negative, finite or
 *       infinite.
 *
 * -----------------------------------
 *
 * What is Mixed Integer Programming?
 *
 *   Here, the constraints and the objective are still linear but
 *   there are additional integrality requirements for variables. If
 *   all variables are required to take integer values, then the
 *   problem is called an integer program (IP). In most cases, only
 *   some variables are required to be integer and the rest of the
 *   variables are continuous: this is called a mixed integer program
 *   (MIP). IPs and MIPs are generally NP-hard.
 *
 *   Integer variables can be used to model discrete decisions (build a
 *   datacenter in city A or city B), logical relationships (only
 *   place machines in datacenter A if we have decided to build
 *   datacenter A) and approximate non-linear functions with piecewise
 *   linear functions (for example, the cost of machines as a function
 *   of how many machines are bought, or the latency of a server as a
 *   function of its load).
 *
 * -----------------------------------
 *
 * How to use the wrapper
 *
 *   The user builds the model and solves it through the MPSolver class,
 *   then queries the solution through the MPSolver, MPVariable and
 *   MPConstraint classes. To be able to query a solution, you need the
 *   following:
 *   - A solution exists: MPSolver::Solve has been called and a solution
 *     has been found.
 *   - The model has not been modified since the last time
 *     MPSolver::Solve was called. Otherwise, the solution obtained
 *     before the model modification may not longer be feasible or
 *     optimal.
 *
 * @see ../examples/linear_programming.cc for a simple LP example.
 *
 * @see ../examples/integer_programming.cc for a simple MIP example.
 *
 *   All methods cannot be called successfully in all cases. For
 *   example: you cannot query a solution when no solution exists, you
 *   cannot query a reduced cost value (which makes sense only on
 *   continuous problems) on a discrete problem. When a method is
 *   called in an unsuitable context, it aborts with a
 *   LOG(FATAL).
 * TODO(user): handle failures gracefully.
 *
 * -----------------------------------
 *
 * For developers: How the wrapper works
 *
 *   MPSolver stores a representation of the model (variables,
 *   constraints and objective) in its own data structures and a
 *   pointer to a MPSolverInterface that wraps the underlying solver
 *   (GLOP, CBC, CLP, GLPK, or SCIP) that does the actual work. The
 *   underlying solver also keeps a representation of the model in its
 *   own data structures. The model representations in MPSolver and in
 *   the underlying solver are kept in sync by the 'extraction'
 *   mechanism: synchronously for some changes and asynchronously
 *   (when MPSolver::Solve is called) for others. Synchronicity
 *   depends on the modification applied and on the underlying solver.
 */

#ifndef OR_TOOLS_LINEAR_SOLVER_LINEAR_SOLVER_H_
#define OR_TOOLS_LINEAR_SOLVER_LINEAR_SOLVER_H_

#include <atomic>
#include <cstdint>
#include <functional>
#include <limits>
#include <map>
#include <memory>
#include <optional>
#include <ostream>
#include <string>
#include <utility>
#include <vector>

#include "absl/base/attributes.h"
#include "absl/base/port.h"
#include "absl/base/thread_annotations.h"
#include "absl/container/flat_hash_map.h"
#include "absl/flags/declare.h"
#include "absl/log/check.h"
#include "absl/status/status.h"
#include "absl/strings/str_format.h"
#include "absl/strings/str_split.h"
#include "absl/strings/string_view.h"
#include "absl/time/clock.h"
#include "absl/time/time.h"
#include "absl/types/optional.h"
#include "ortools/base/logging.h"
#include "ortools/linear_solver/linear_expr.h"
#include "ortools/linear_solver/linear_solver.pb.h"
#include "ortools/linear_solver/linear_solver_callback.h"
#include "ortools/port/proto_utils.h"

ABSL_DECLARE_FLAG(bool, linear_solver_enable_verbose_output);
ABSL_DECLARE_FLAG(bool, log_verification_errors);
ABSL_DECLARE_FLAG(bool, verify_solution);

namespace operations_research {

constexpr double kDefaultPrimalTolerance = 1e-07;

class MPConstraint;
class MPObjective;
class MPSolverInterface;
class MPSolverParameters;
class MPVariable;

// There is a homonymous version taking a MPSolver::OptimizationProblemType.
bool SolverTypeIsMip(MPModelRequest::SolverType solver_type);

/**
 * This mathematical programming (MP) solver class is the main class
 * though which users build and solve problems.
 */
class MPSolver {
 public:
  /**
   * The type of problems (LP or MIP) that will be solved and the underlying
   *  solver (GLOP, GLPK, CLP, CBC or SCIP) that will solve them. This must
   * remain consistent with MPModelRequest::OptimizationProblemType
   *  (take particular care of the open-source version).
   */
  enum OptimizationProblemType {
    // Linear programming problems.
    // ----------------------------
    CLP_LINEAR_PROGRAMMING = 0,
    GLPK_LINEAR_PROGRAMMING = 1,
    GLOP_LINEAR_PROGRAMMING = 2,  // Recommended default value. Made in Google.
    // In-house linear programming solver based on the primal-dual hybrid
    // gradient method. Sometimes faster than Glop for medium-size problems and
    // scales to much larger problems than Glop.
    PDLP_LINEAR_PROGRAMMING = 8,
    HIGHS_LINEAR_PROGRAMMING = 15,

    // Integer programming problems.
    // -----------------------------
    // Recommended default value for MIP problems.
    SCIP_MIXED_INTEGER_PROGRAMMING = 3,
    GLPK_MIXED_INTEGER_PROGRAMMING = 4,
    CBC_MIXED_INTEGER_PROGRAMMING = 5,
    HIGHS_MIXED_INTEGER_PROGRAMMING = 16,

    // Boolean optimization problem (requires only integer variables and works
    // best with only Boolean variables).
    BOP_INTEGER_PROGRAMMING = 12,

    // SAT based solver (requires only integer and Boolean variables).
    // If you pass it mixed integer problems, it will scale coefficients to
    // integer values, and solver continuous variables as integral variables.
    //
    // Recommended default value for pure integral problems problems.
    SAT_INTEGER_PROGRAMMING = 14,

    // Dedicated knapsack solvers.
    KNAPSACK_MIXED_INTEGER_PROGRAMMING = 13,

    // Commercial software (need license).
    GUROBI_LINEAR_PROGRAMMING = 6,
    GUROBI_MIXED_INTEGER_PROGRAMMING = 7,
    CPLEX_LINEAR_PROGRAMMING = 10,
    CPLEX_MIXED_INTEGER_PROGRAMMING = 11,
    XPRESS_LINEAR_PROGRAMMING = 101,
    XPRESS_MIXED_INTEGER_PROGRAMMING = 102,
    COPT_LINEAR_PROGRAMMING = 103,
    COPT_MIXED_INTEGER_PROGRAMMING = 104,
  };

  /// Create a solver with the given name and underlying solver backend.
  MPSolver(const std::string& name, OptimizationProblemType problem_type);

#ifndef SWIG
  // This type is neither copyable nor movable.
  MPSolver(const MPSolver&) = delete;
  MPSolver& operator=(const MPSolver&) = delete;
#endif

  virtual ~MPSolver();

  /**
   * Recommended factory method to create a MPSolver instance, especially in
   * non C++ languages.
   *
   * It returns a newly created solver instance if successful, or a nullptr
   * otherwise. This can occur if the relevant interface is not linked in, or if
   * a needed license is not accessible for commercial solvers.
   *
   * Ownership of the solver is passed on to the caller of this method.
   * It will accept both string names of the OptimizationProblemType enum, as
   * well as a short version (i.e. "SCIP_MIXED_INTEGER_PROGRAMMING" or "SCIP").
   *
   * solver_id is case insensitive, and the following names are supported:
   *   - CLP_LINEAR_PROGRAMMING or CLP
   *   - CBC_MIXED_INTEGER_PROGRAMMING or CBC
   *   - GLOP_LINEAR_PROGRAMMING or GLOP
   *   - BOP_INTEGER_PROGRAMMING or BOP
   *   - SAT_INTEGER_PROGRAMMING or SAT or CP_SAT
   *   - SCIP_MIXED_INTEGER_PROGRAMMING or SCIP
   *   - GUROBI_LINEAR_PROGRAMMING or GUROBI_LP
   *   - GUROBI_MIXED_INTEGER_PROGRAMMING or GUROBI or GUROBI_MIP
   *   - CPLEX_LINEAR_PROGRAMMING or CPLEX_LP
   *   - CPLEX_MIXED_INTEGER_PROGRAMMING or CPLEX or CPLEX_MIP
   *   - XPRESS_LINEAR_PROGRAMMING or XPRESS_LP
   *   - XPRESS_MIXED_INTEGER_PROGRAMMING or XPRESS or XPRESS_MIP
   *   - GLPK_LINEAR_PROGRAMMING or GLPK_LP
   *   - GLPK_MIXED_INTEGER_PROGRAMMING or GLPK or GLPK_MIP
   */
  static MPSolver* CreateSolver(const std::string& solver_id);

  /**
   * Whether the given problem type is supported (this will depend on the
   * targets that you linked).
   */
  static bool SupportsProblemType(OptimizationProblemType problem_type);

  /**
   * Parses the name of the solver. Returns true if the solver type is
   * successfully parsed as one of the OptimizationProblemType.
   * See the documentation of CreateSolver() for the list of supported names.
   */
  static bool ParseSolverType(absl::string_view solver_id,
                              OptimizationProblemType* type);

  /**
   * Parses the name of the solver and returns the correct optimization type or
   * dies. Invariant: ParseSolverTypeOrDie(ToString(type)) = type.
   */
  static OptimizationProblemType ParseSolverTypeOrDie(
      const std::string& solver_id);

  bool IsMIP() const;

  /// Returns the name of the model set at construction.
  const std::string& Name() const {
    return name_;  // Set at construction.
  }

  /// Returns the optimization problem type set at construction.
  virtual OptimizationProblemType ProblemType() const {
    return problem_type_;  // Set at construction.
  }

  /**
   * Clears the objective (including the optimization direction), all variables
   * and constraints. All the other properties of the MPSolver (like the time
   * limit) are kept untouched.
   */
  void Clear();

  /// Returns the number of variables.
  int NumVariables() const { return variables_.size(); }

  /**
   * Returns the array of variables handled by the MPSolver. (They are listed in
   * the order in which they were created.)
   */
  const std::vector<MPVariable*>& variables() const { return variables_; }

  /**
   * Returns the variable at position index.
   */
  MPVariable* variable(int index) const { return variables_[index]; }

  /**
   * Looks up a variable by name, and returns nullptr if it does not exist. The
   * first call has a O(n) complexity, as the variable name index is lazily
   * created upon first use. Will crash if variable names are not unique.
   */
  MPVariable* LookupVariableOrNull(const std::string& var_name) const;

  /**
   * Creates a variable with the given bounds, integrality requirement and
   * name. Bounds can be finite or +/- MPSolver::infinity(). The MPSolver owns
   * the variable (i.e. the returned pointer is borrowed). Variable names are
   * optional. If you give an empty name, name() will auto-generate one for you
   * upon request.
   */
  MPVariable* MakeVar(double lb, double ub, bool integer,
                      const std::string& name);

  /// Creates a continuous variable.
  MPVariable* MakeNumVar(double lb, double ub, const std::string& name);

  /// Creates an integer variable.
  MPVariable* MakeIntVar(double lb, double ub, const std::string& name);

  /// Creates a boolean variable.
  MPVariable* MakeBoolVar(const std::string& name);

  /**
   * Creates an array of variables. All variables created have the same bounds
   * and integrality requirement. If nb <= 0, no variables are created, the
   * function crashes in non-opt mode.
   *
   * @param nb the number of variables to create.
   * @param lb the lower bound of created variables
   * @param ub the upper bound of created variables
   * @param integer controls whether the created variables are continuous or
   * integral.
   * @param name_prefix the prefix of the variable names. Variables are named
   * name_prefix0, name_prefix1, ...
   * @param[out] vars the vector of variables to fill with variables.
   */
  void MakeVarArray(int nb, double lb, double ub, bool integer,
                    const std::string& name_prefix,
                    std::vector<MPVariable*>* vars);

  /// Creates an array of continuous variables.
  void MakeNumVarArray(int nb, double lb, double ub, const std::string& name,
                       std::vector<MPVariable*>* vars);

  ///  Creates an array of integer variables.
  void MakeIntVarArray(int nb, double lb, double ub, const std::string& name,
                       std::vector<MPVariable*>* vars);

  /// Creates an array of boolean variables.
  void MakeBoolVarArray(int nb, const std::string& name,
                        std::vector<MPVariable*>* vars);

  /// Returns the number of constraints.
  int NumConstraints() const { return constraints_.size(); }

  /**
   * Returns the array of constraints handled by the MPSolver.
   *
   * They are listed in the order in which they were created.
   */
  const std::vector<MPConstraint*>& constraints() const { return constraints_; }

  /** Returns the constraint at the given index. */
  MPConstraint* constraint(int index) const { return constraints_[index]; }

  /**
   *  Looks up a constraint by name, and returns nullptr if it does not exist.
   *
   * The first call has a O(n) complexity, as the constraint name index is
   * lazily created upon first use. Will crash if constraint names are not
   * unique.
   */
  MPConstraint* LookupConstraintOrNull(
      const std::string& constraint_name) const;

  /**
   * Creates a linear constraint with given bounds.
   *
   * Bounds can be finite or +/- MPSolver::infinity(). The MPSolver class
   * assumes ownership of the constraint.
   *
   * @return a pointer to the newly created constraint.
   */
  MPConstraint* MakeRowConstraint(double lb, double ub);

  /// Creates a constraint with -infinity and +infinity bounds.
  MPConstraint* MakeRowConstraint();

  /// Creates a named constraint with given bounds.
  MPConstraint* MakeRowConstraint(double lb, double ub,
                                  const std::string& name);

  /// Creates a named constraint with -infinity and +infinity bounds.
  MPConstraint* MakeRowConstraint(const std::string& name);

  /**
   * Creates a constraint owned by MPSolver enforcing:
   *     range.lower_bound() <= range.linear_expr() <= range.upper_bound()
   */
  MPConstraint* MakeRowConstraint(const LinearRange& range);

  /// As above, but also names the constraint.
  MPConstraint* MakeRowConstraint(const LinearRange& range,
                                  const std::string& name);

  /**
   * Returns the objective object.
   *
   * Note that the objective is owned by the solver, and is initialized to its
   * default value (see the MPObjective class below) at construction.
   */
  const MPObjective& Objective() const { return *objective_; }

  /// Returns the mutable objective object.
  MPObjective* MutableObjective() { return objective_.get(); }

  /**
   * The status of solving the problem. The straightforward translation to
   * homonymous enum values of MPSolverResponseStatus (see
   * ./linear_solver.proto) is guaranteed by ./enum_consistency_test.cc, you may
   * rely on it.
   */
  enum ResultStatus {
    /// optimal.
    OPTIMAL,
    /// feasible, or stopped by limit.
    FEASIBLE,
    /// proven infeasible.
    INFEASIBLE,
    /// proven unbounded.
    UNBOUNDED,
    /// abnormal, i.e., error of some kind.
    ABNORMAL,
    /// the model is trivially invalid (NaN coefficients, etc).
    MODEL_INVALID,
    /// not been solved yet.
    NOT_SOLVED = 6
  };

  /// Solves the problem using the default parameter values.
  ResultStatus Solve();

  /// Solves the problem using the specified parameter values.
  ResultStatus Solve(const MPSolverParameters& param);

  /**
   * Writes the model using the solver internal write function.  Currently only
   * available for Gurobi.
   */
  void Write(const std::string& file_name);

  /**
   * Advanced usage: compute the "activities" of all constraints, which are the
   * sums of their linear terms. The activities are returned in the same order
   * as constraints(), which is the order in which constraints were added; but
   * you can also use MPConstraint::index() to get a constraint's index.
   */
  std::vector<double> ComputeConstraintActivities() const;

  /**
   * Advanced usage: Verifies the *correctness* of the solution.
   *
   * It verifies that all variables must be within their domains, all
   * constraints must be satisfied, and the reported objective value must be
   * accurate.
   *
   * Usage:
   * - This can only be called after Solve() was called.
   * - "tolerance" is interpreted as an absolute error threshold.
   * - For the objective value only, if the absolute error is too large,
   *   the tolerance is interpreted as a relative error threshold instead.
   * - If "log_errors" is true, every single violation will be logged.
   * - If "tolerance" is negative, it will be set to infinity().
   *
   * Most users should just set the --verify_solution flag and not bother using
   * this method directly.
   */
  bool VerifySolution(double tolerance, bool log_errors) const;

  /**
   * Advanced usage: resets extracted model to solve from scratch.
   *
   * This won't reset the parameters that were set with
   * SetSolverSpecificParametersAsString() or set_time_limit() or even clear the
   * linear program. It will just make sure that next Solve() will be as if
   * everything was reconstructed from scratch.
   */
  void Reset();

  /** Interrupts the Solve() execution to terminate processing if possible.
   *
   * If the underlying interface supports interruption; it does that and returns
   * true regardless of whether there's an ongoing Solve() or not. The Solve()
   * call may still linger for a while depending on the conditions.  If
   * interruption is not supported; returns false and does nothing.
   * MPSolver::SolverTypeSupportsInterruption can be used to check if
   * interruption is supported for a given solver type.
   */
  bool InterruptSolve();

  /**
   * Loads model from protocol buffer.
   *
   * Returns MPSOLVER_MODEL_IS_VALID if the model is valid, and another status
   * otherwise (currently only MPSOLVER_MODEL_INVALID and MPSOLVER_INFEASIBLE).
   * If the model isn't valid, populates "error_message".
   * If `clear_names` is true (the default), clears all names, otherwise returns
   * MPSOLVER_MODEL_INVALID if there are duplicates (non-empty) names.
   */
  MPSolverResponseStatus LoadModelFromProto(const MPModelProto& input_model,
                                            std::string* error_message,
                                            bool clear_names = true);
  /**
   * Loads model from protocol buffer.
   *
   * The same as above, except that the loading keeps original variable and
   * constraint names. Caller should make sure that all variable names and
   * constraint names are unique, respectively.
   */
  MPSolverResponseStatus LoadModelFromProtoWithUniqueNamesOrDie(
      const MPModelProto& input_model, std::string* error_message);

  /// Encodes the current solution in a solution response protocol buffer.
  void FillSolutionResponseProto(MPSolutionResponse* response) const;

  /**
   * Solves the model encoded by a MPModelRequest protocol buffer and fills the
   * solution encoded as a MPSolutionResponse. The solve is stopped prematurely
   * if interrupt is non-null at set to true during (or before) solving.
   * Interruption is only supported if SolverTypeSupportsInterruption() returns
   * true for the requested solver. Passing a non-null interruption with any
   * other solver type immediately returns an MPSOLVER_INCOMPATIBLE_OPTIONS
   * error.
   *
   * Note(user): This attempts to first use `DirectlySolveProto()` (if
   * implemented). Consequently, this most likely does *not* override any of
   * the default parameters of the underlying solver. This behavior *differs*
   * from `MPSolver::Solve()` which by default sets the feasibility tolerance
   * and the gap limit (as of 2020/02/11, to 1e-7 and 0.0001, respectively).
   */
  static void SolveWithProto(const MPModelRequest& model_request,
                             MPSolutionResponse* response,
                             // `interrupt` is non-const because the internal
                             // solver may set it to true itself, in some cases.
                             std::atomic<bool>* interrupt = nullptr);

  static bool SolverTypeSupportsInterruption(
      const MPModelRequest::SolverType solver) {
    // Interruption requires that MPSolver::InterruptSolve is supported for the
    // underlying solver. Interrupting requests using SCIP is also not supported
    // as of 2021/08/23, since InterruptSolve is not thread safe for SCIP.
    return solver == MPModelRequest::GLOP_LINEAR_PROGRAMMING ||
           solver == MPModelRequest::GUROBI_LINEAR_PROGRAMMING ||
           solver == MPModelRequest::GUROBI_MIXED_INTEGER_PROGRAMMING ||
           solver == MPModelRequest::SAT_INTEGER_PROGRAMMING ||
           solver == MPModelRequest::PDLP_LINEAR_PROGRAMMING;
  }

  /// Exports model to protocol buffer.
  void ExportModelToProto(MPModelProto* output_model) const;

  /**
   * Load a solution encoded in a protocol buffer onto this solver for easy
  access via the MPSolver interface.
   *
   * IMPORTANT: This may only be used in conjunction with ExportModel(),
  following this example:
   *
   \code
     MPSolver my_solver;
     ... add variables and constraints ...
     MPModelProto model_proto;
     my_solver.ExportModelToProto(&model_proto);
     MPSolutionResponse solver_response;
     MPSolver::SolveWithProto(model_proto, &solver_response);
     if (solver_response.result_status() == MPSolutionResponse::OPTIMAL) {
       CHECK_OK(my_solver.LoadSolutionFromProto(solver_response));
       ... inspect the solution using the usual API: solution_value(), etc...
     }
  \endcode
   *
   * The response must be in OPTIMAL or FEASIBLE status.
   *
   * Returns a non-OK status if a problem arised (typically, if it wasn't used
   *     like it should be):
   * - loading a solution whose variables don't correspond to the solver's
   *   current variables
   * - loading a dual solution whose constraints don't correspond to the
   *   solver's current constraints
   * - loading a solution with a status other than OPTIMAL / FEASIBLE.
   *
   * Note: the objective value isn't checked. You can use VerifySolution() for
   *       that.
   */
  absl::Status LoadSolutionFromProto(
      const MPSolutionResponse& response,
      double tolerance = std::numeric_limits<double>::infinity());

  /**
   * Resets values of out of bound variables to the corresponding bound and
   * returns an error if any of the variables have NaN value.
   */
  absl::Status ClampSolutionWithinBounds();

  /**
   * Shortcuts to the homonymous MPModelProtoExporter methods, via exporting to
   * a MPModelProto with ExportModelToProto() (see above).
   *
   * Produces empty std::string on portable platforms (e.g. android, ios).
   */
  bool ExportModelAsLpFormat(bool obfuscate, std::string* model_str) const;
  bool ExportModelAsMpsFormat(bool fixed_format, bool obfuscate,
                              std::string* model_str) const;

  /**
   *  Sets the number of threads to use by the underlying solver.
   *
   * Returns OkStatus if the operation was successful. num_threads must be equal
   * to or greater than 1. Note that the behaviour of this call depends on the
   * underlying solver. E.g., it may set the exact number of threads or the max
   * number of threads (check the solver's interface implementation for
   * details). Also, some solvers may not (yet) support this function, but still
   * enable multi-threading via SetSolverSpecificParametersAsString().
   */
  absl::Status SetNumThreads(int num_threads);

  /// Returns the number of threads to be used during solve.
  int GetNumThreads() const { return num_threads_; }

  /**
   * Advanced usage: pass solver specific parameters in text format.
   *
   * The format is solver-specific and is the same as the corresponding solver
   * configuration file format. Returns true if the operation was successful.
   */
  bool SetSolverSpecificParametersAsString(const std::string& parameters);
  std::string GetSolverSpecificParametersAsString() const {
    return solver_specific_parameter_string_;
  }

  /**
   * Sets a hint for solution.
   *
   * If a feasible or almost-feasible solution to the problem is already known,
   * it may be helpful to pass it to the solver so that it can be used. A solver
   * that supports this feature will try to use this information to create its
   * initial feasible solution.
   *
   * Note: It may not always be faster to give a hint like this to the
   * solver. There is also no guarantee that the solver will use this hint or
   * try to return a solution "close" to this assignment in case of multiple
   * optimal solutions.
   *
   * Calling SetHint clears all previous hints.
   */
  void SetHint(std::vector<std::pair<const MPVariable*, double> > hint);

  // Gives some brief (a few lines, at most) human-readable information about
  // the given request, suitable for debug logging.
  static std::string GetMPModelRequestLoggingInfo(
      const MPModelRequest& request);

  /**
   * Advanced usage: possible basis status values for a variable and the slack
   * variable of a linear constraint.
   */
  enum BasisStatus {
    FREE = 0,
    AT_LOWER_BOUND,
    AT_UPPER_BOUND,
    FIXED_VALUE,
    BASIC
  };

  /**
   * Advanced usage: Incrementality.
   *
   * This function takes a starting basis to be used in the next LP Solve()
   * call. The statuses of a current solution can be retrieved via the
   * basis_status() function of a MPVariable or a MPConstraint.
   *
   * WARNING: With Glop, you should disable presolve when using this because
   * this information will not be modified in sync with the presolve and will
   * likely not mean much on the presolved problem.
   */
  void SetStartingLpBasis(
      const std::vector<MPSolver::BasisStatus>& variable_statuses,
      const std::vector<MPSolver::BasisStatus>& constraint_statuses);

  /**
   * Infinity.
   *
   * You can use -MPSolver::infinity() for negative infinity.
   */
  static double infinity() { return std::numeric_limits<double>::infinity(); }
  double solver_infinity();

  /**
   * Controls (or queries) the amount of output produced by the underlying
   * solver. The output can surface to LOGs, or to stdout or stderr, depending
   * on the implementation. The amount of output will greatly vary with each
   * implementation and each problem.
   *
   * Output is suppressed by default.
   */
  bool OutputIsEnabled() const;

  /// Enables solver logging.
  void EnableOutput();

  /// Suppresses solver logging.
  void SuppressOutput();

  absl::Duration TimeLimit() const { return time_limit_; }
  void SetTimeLimit(absl::Duration time_limit) {
    DCHECK_GE(time_limit, absl::ZeroDuration());
    time_limit_ = time_limit;
  }

  absl::Duration DurationSinceConstruction() const {
    return absl::Now() - construction_time_;
  }

  /// Returns the number of simplex iterations.
  int64_t iterations() const;

  /**
   * Returns the number of branch-and-bound nodes evaluated during the solve.
   *
   * Only available for discrete problems.
   */
  int64_t nodes() const;

  /// Returns a string describing the underlying solver and its version.
  std::string SolverVersion() const;

  /**
   * Advanced usage: returns the underlying solver.
   *
   * Returns the underlying solver so that the user can use solver-specific
   * features or features that are not exposed in the simple API of MPSolver.
   * This method is for advanced users, use at your own risk! In particular, if
   * you modify the model or the solution by accessing the underlying solver
   * directly, then the underlying solver will be out of sync with the
   * information kept in the wrapper (MPSolver, MPVariable, MPConstraint,
   * MPObjective). You need to cast the void* returned back to its original type
   * that depends on the interface (CBC: OsiClpSolverInterface*, CLP:
   * ClpSimplex*, GLPK: glp_prob*, SCIP: SCIP*).
   */
  void* underlying_solver();

  /** Advanced usage: computes the exact condition number of the current scaled
   * basis: L1norm(B) * L1norm(inverse(B)), where B is the scaled basis.
   *
   * This method requires that a basis exists: it should be called after Solve.
   * It is only available for continuous problems. It is implemented for GLPK
   * but not CLP because CLP does not provide the API for doing it.
   *
   * The condition number measures how well the constraint matrix is conditioned
   * and can be used to predict whether numerical issues will arise during the
   * solve: the model is declared infeasible whereas it is feasible (or
   * vice-versa), the solution obtained is not optimal or violates some
   * constraints, the resolution is slow because of repeated singularities.
   *
   * The rule of thumb to interpret the condition number kappa is:
   *   - o kappa <= 1e7: virtually no chance of numerical issues
   *   - o 1e7 < kappa <= 1e10: small chance of numerical issues
   *   - o 1e10 < kappa <= 1e13: medium chance of numerical issues
   *   - o kappa > 1e13: high chance of numerical issues
   *
   * The computation of the condition number depends on the quality of the LU
   * decomposition, so it is not very accurate when the matrix is ill
   * conditioned.
   */
  double ComputeExactConditionNumber() const;

  /**
   * Some solvers (MIP only, not LP) can produce multiple solutions to the
   * problem. Returns true when another solution is available, and updates the
   * MPVariable* objects to make the new solution queryable. Call only after
   * calling solve.
   *
   * The optimality properties of the additional solutions found, and whether or
   * not the solver computes them ahead of time or when NextSolution() is called
   * is solver specific.
   *
   * As of 2020-02-10, only Gurobi and SCIP support NextSolution(), see
   * linear_solver_interfaces_test for an example of how to configure these
   * solvers for multiple solutions. Other solvers return false unconditionally.
   */
  ABSL_MUST_USE_RESULT bool NextSolution();

  // Does not take ownership of "mp_callback".
  //
  // As of 2019-10-22, only SCIP and Gurobi support Callbacks.
  // SCIP does not support suggesting a heuristic solution in the callback.
  void SetCallback(MPCallback* mp_callback);
  bool SupportsCallbacks() const;

  // Global counters of variables and constraints ever created across all
  // MPSolver instances. Those are only updated after the destruction
  // (or Clear()) of each MPSolver instance.
  static int64_t global_num_variables();
  static int64_t global_num_constraints();

  // DEPRECATED: Use TimeLimit() and SetTimeLimit(absl::Duration) instead.
  // NOTE: These deprecated functions used the convention time_limit = 0 to mean
  // "no limit", which now corresponds to time_limit_ = InfiniteDuration().
  int64_t time_limit() const {
    return time_limit_ == absl::InfiniteDuration()
               ? 0
               : absl::ToInt64Milliseconds(time_limit_);
  }
  void set_time_limit(int64_t time_limit_milliseconds) {
    SetTimeLimit(time_limit_milliseconds == 0
                     ? absl::InfiniteDuration()
                     : absl::Milliseconds(time_limit_milliseconds));
  }
  double time_limit_in_secs() const {
    return static_cast<double>(time_limit()) / 1000.0;
  }

  // DEPRECATED: Use DurationSinceConstruction() instead.
  int64_t wall_time() const {
    return absl::ToInt64Milliseconds(DurationSinceConstruction());
  }

  friend class GLPKInterface;
  friend class CLPInterface;
  friend class CBCInterface;
  friend class SCIPInterface;
  friend class GurobiInterface;
  friend class CplexInterface;
  friend class XpressInterface;
  friend class SLMInterface;
  friend class MPSolverInterface;
  friend class GLOPInterface;
  friend class BopInterface;
  friend class SatInterface;
  friend class PdlpInterface;
  friend class HighsInterface;
  friend class KnapsackInterface;

  // Debugging: verify that the given MPVariable* belongs to this solver.
  bool OwnsVariable(const MPVariable* var) const;

 private:
  // Computes the size of the constraint with the largest number of
  // coefficients with index in [min_constraint_index,
  // max_constraint_index)
  int ComputeMaxConstraintSize(int min_constraint_index,
                               int max_constraint_index) const;

  // Returns true if the model has constraints with lower bound > upper bound.
  bool HasInfeasibleConstraints() const;

  // Returns true if the model has at least 1 integer variable.
  bool HasIntegerVariables() const;

  // Generates the map from variable names to their indices.
  void GenerateVariableNameIndex() const;

  // Generates the map from constraint names to their indices.
  void GenerateConstraintNameIndex() const;

  // The name of the linear programming problem.
  const std::string name_;

  // The type of the linear programming problem.
  const OptimizationProblemType problem_type_;

  // The solver interface.
  std::unique_ptr<MPSolverInterface> interface_;

  // The vector of variables in the problem.
  std::vector<MPVariable*> variables_;
  // A map from a variable's name to its index in variables_.
  mutable std::optional<absl::flat_hash_map<std::string, int> >
      variable_name_to_index_;
  // Whether variables have been extracted to the underlying interface.
  std::vector<bool> variable_is_extracted_;

  // The vector of constraints in the problem.
  std::vector<MPConstraint*> constraints_;
  // A map from a constraint's name to its index in constraints_.
  mutable std::optional<absl::flat_hash_map<std::string, int> >
      constraint_name_to_index_;
  // Whether constraints have been extracted to the underlying interface.
  std::vector<bool> constraint_is_extracted_;

  // The linear objective function.
  std::unique_ptr<MPObjective> objective_;

  // Initial values for all or some of the problem variables that can be
  // exploited as a starting hint by a solver.
  //
  // Note(user): as of 05/05/2015, we can't use >> because of some SWIG errors.
  //
  // TODO(user): replace by two vectors, a std::vector<bool> to indicate if a
  // hint is provided and a std::vector<double> for the hint value.
  std::vector<std::pair<const MPVariable*, double> > solution_hint_;

  absl::Duration time_limit_ = absl::InfiniteDuration();  // Default = No limit.

  const absl::Time construction_time_;

  // Permanent storage for the number of threads.
  int num_threads_ = 1;

  // Permanent storage for SetSolverSpecificParametersAsString().
  std::string solver_specific_parameter_string_;

  static absl::Mutex global_count_mutex_;
#ifndef SWIG
  static int64_t global_num_variables_ ABSL_GUARDED_BY(global_count_mutex_);
  static int64_t global_num_constraints_ ABSL_GUARDED_BY(global_count_mutex_);
#endif

  enum ModelProtoNamesPolicy {
    DEFAULT_CLEAR_NAMES = 0,
    INVALID_MODEL_ON_DUPLICATE_NONEMPTY_NAMES = 1,
    DIE_ON_DUPLICATE_NONEMPTY_NAMES = 2,
  };
  MPSolverResponseStatus LoadModelFromProtoInternal(
      const MPModelProto& input_model, ModelProtoNamesPolicy name_policy,
      bool check_model_validity, std::string* error_message);
};

inline bool SolverTypeIsMip(MPSolver::OptimizationProblemType solver_type) {
  return SolverTypeIsMip(static_cast<MPModelRequest::SolverType>(solver_type));
}

absl::string_view ToString(
    MPSolver::OptimizationProblemType optimization_problem_type);

inline std::ostream& operator<<(
    std::ostream& os,
    MPSolver::OptimizationProblemType optimization_problem_type) {
  return os << ToString(optimization_problem_type);
}

inline std::ostream& operator<<(std::ostream& os,
                                MPSolver::ResultStatus status) {
  return os << ProtoEnumToString<MPSolverResponseStatus>(
             static_cast<MPSolverResponseStatus>(status));
}

bool AbslParseFlag(absl::string_view text,
                   MPSolver::OptimizationProblemType* solver_type,
                   std::string* error);

inline std::string AbslUnparseFlag(
    MPSolver::OptimizationProblemType solver_type) {
  return std::string(ToString(solver_type));
}

/// A class to express a linear objective.
class MPObjective {
 public:
#ifndef SWIG
  // This type is neither copyable nor movable.
  MPObjective(const MPObjective&) = delete;
  MPObjective& operator=(const MPObjective&) = delete;
#endif

  /**
   *  Clears the offset, all variables and coefficients, and the optimization
   * direction.
   */
  void Clear();

  /**
   * Sets the coefficient of the variable in the objective.
   *
   * If the variable does not belong to the solver, the function just returns,
   * or crashes in non-opt mode.
   */
  void SetCoefficient(const MPVariable* var, double coeff);

  /**
   *  Gets the coefficient of a given variable in the objective
   *
   * It returns 0 if the variable does not appear in the objective).
   */
  double GetCoefficient(const MPVariable* var) const;

  /**
   * Returns a map from variables to their coefficients in the objective.
   *
   * If a variable is not present in the map, then its coefficient is zero.
   */
  const absl::flat_hash_map<const MPVariable*, double>& terms() const {
    return coefficients_;
  }

  /// Sets the constant term in the objective.
  void SetOffset(double value);

  /// Gets the constant term in the objective.
  double offset() const { return offset_; }

  /**
   * Resets the current objective to take the value of linear_expr, and sets the
   * objective direction to maximize if "is_maximize", otherwise minimizes.
   */
  void OptimizeLinearExpr(const LinearExpr& linear_expr, bool is_maximization);

  /// Resets the current objective to maximize linear_expr.
  void MaximizeLinearExpr(const LinearExpr& linear_expr) {
    OptimizeLinearExpr(linear_expr, true);
  }
  /// Resets the current objective to minimize linear_expr.
  void MinimizeLinearExpr(const LinearExpr& linear_expr) {
    OptimizeLinearExpr(linear_expr, false);
  }

  /// Adds linear_expr to the current objective, does not change the direction.
  void AddLinearExpr(const LinearExpr& linear_expr);

  /// Sets the optimization direction (maximize: true or minimize: false).
  void SetOptimizationDirection(bool maximize);

  /// Sets the optimization direction to minimize.
  void SetMinimization() { SetOptimizationDirection(false); }

  /// Sets the optimization direction to maximize.
  void SetMaximization() { SetOptimizationDirection(true); }

  /// Is the optimization direction set to maximize?
  bool maximization() const;

  /// Is the optimization direction set to minimize?
  bool minimization() const;

  /**
   * Returns the objective value of the best solution found so far.
   *
   * It is the optimal objective value if the problem has been solved to
   * optimality.
   *
   * Note: the objective value may be slightly different than what you could
   * compute yourself using \c MPVariable::solution_value(); please use the
   * --verify_solution flag to gain confidence about the numerical stability of
   * your solution.
   */
  double Value() const;

  /**
   * Returns the best objective bound.
   *
   * In case of minimization, it is a lower bound on the objective value of the
   * optimal integer solution. Only available for discrete problems.
   */
  double BestBound() const;

 private:
  friend class MPSolver;
  friend class MPSolverInterface;
  friend class CBCInterface;
  friend class CLPInterface;
  friend class GLPKInterface;
  friend class SCIPInterface;
  friend class SLMInterface;
  friend class GurobiInterface;
  friend class CplexInterface;
  friend class XpressInterface;
  friend class GLOPInterface;
  friend class BopInterface;
  friend class SatInterface;
  friend class PdlpInterface;
  friend class HighsInterface;
  friend class KnapsackInterface;

  // Constructor. An objective points to a single MPSolverInterface
  // that is specified in the constructor. An objective cannot belong
  // to several models.
  // At construction, an MPObjective has no terms (which is equivalent
  // on having a coefficient of 0 for all variables), and an offset of 0.
  explicit MPObjective(MPSolverInterface* const interface_in)
      : interface_(interface_in), coefficients_(1), offset_(0.0) {}

  MPSolverInterface* const interface_;

  // Mapping var -> coefficient.
  absl::flat_hash_map<const MPVariable*, double> coefficients_;
  // Constant term.
  double offset_;
};

/// The class for variables of a Mathematical Programming (MP) model.
class MPVariable {
 public:
#ifndef SWIG
  // This type is neither copyable nor movable.
  MPVariable(const MPVariable&) = delete;
  MPVariable& operator=(const MPVariable&) = delete;
#endif

  /// Returns the name of the variable.
  const std::string& name() const { return name_; }

  /// Sets the integrality requirement of the variable.
  void SetInteger(bool integer);

  /// Returns the integrality requirement of the variable.
  bool integer() const { return integer_; }

  /**
   * Returns the value of the variable in the current solution.
   *
   * If the variable is integer, then the value will always be an integer (the
   * underlying solver handles floating-point values only, but this function
   * automatically rounds it to the nearest integer; see: man 3 round).
   */
  double solution_value() const;

  /// Returns the index of the variable in the MPSolver::variables_.
  int index() const { return index_; }

  /// Returns the lower bound.
  double lb() const { return lb_; }

  /// Returns the upper bound.
  double ub() const { return ub_; }

  /// Sets the lower bound.
  void SetLB(double lb) { SetBounds(lb, ub_); }

  /// Sets the upper bound.
  void SetUB(double ub) { SetBounds(lb_, ub); }

  /// Sets both the lower and upper bounds.
  void SetBounds(double lb, double ub);

  /**
   * Advanced usage: unrounded solution value.
   *
   * The returned value won't be rounded to the nearest integer even if the
   * variable is integer.
   */
  double unrounded_solution_value() const;

  /**
   * Advanced usage: returns the reduced cost of the variable in the current
   * solution (only available for continuous problems).
   */
  double reduced_cost() const;

  /**
   * Advanced usage: returns the basis status of the variable in the current
   * solution (only available for continuous problems).
   *
   * @see MPSolver::BasisStatus.
   */
  MPSolver::BasisStatus basis_status() const;

  /**
   * Advanced usage: Certain MIP solvers (e.g. Gurobi or SCIP) allow you to set
   * a per-variable priority for determining which variable to branch on.
   *
   * A value of 0 is treated as default, and is equivalent to not setting the
   * branching priority. The solver looks first to branch on fractional
   * variables in higher priority levels. As of 2019-05, only Gurobi and SCIP
   * support setting branching priority; all other solvers will simply ignore
   * this annotation.
   */
  int branching_priority() const { return branching_priority_; }
  void SetBranchingPriority(int priority);

 protected:
  friend class MPSolver;
  friend class MPSolverInterface;
  friend class CBCInterface;
  friend class CLPInterface;
  friend class GLPKInterface;
  friend class SCIPInterface;
  friend class SLMInterface;
  friend class GurobiInterface;
  friend class CplexInterface;
  friend class XpressInterface;
  friend class GLOPInterface;
  friend class MPVariableSolutionValueTest;
  friend class BopInterface;
  friend class SatInterface;
  friend class PdlpInterface;
  friend class HighsInterface;
  friend class KnapsackInterface;

  // Constructor. A variable points to a single MPSolverInterface that
  // is specified in the constructor. A variable cannot belong to
  // several models.
  MPVariable(int index, double lb, double ub, bool integer,
             const std::string& name, MPSolverInterface* const interface_in)
      : index_(index),
        lb_(lb),
        ub_(ub),
        integer_(integer),
        name_(name.empty() ? absl::StrFormat("auto_v_%09d", index) : name),
        solution_value_(0.0),
        reduced_cost_(0.0),
        interface_(interface_in) {}

  void set_solution_value(double value) { solution_value_ = value; }
  void set_reduced_cost(double reduced_cost) { reduced_cost_ = reduced_cost; }

 private:
  const int index_;
  double lb_;
  double ub_;
  bool integer_;
  const std::string name_;
  double solution_value_;
  double reduced_cost_;
  int branching_priority_ = 0;
  MPSolverInterface* const interface_;
};

/**
 * The class for constraints of a Mathematical Programming (MP) model.
 *
 * A constraint is represented as a linear equation or inequality.
 */
class MPConstraint {
 public:
#ifndef SWIG
  // This type is neither copyable nor movable.
  MPConstraint(const MPConstraint&) = delete;
  MPConstraint& operator=(const MPConstraint&) = delete;
#endif

  /// Returns the name of the constraint.
  const std::string& name() const { return name_; }

  /// Clears all variables and coefficients. Does not clear the bounds.
  void Clear();

  /**
   * Sets the coefficient of the variable on the constraint.
   *
   * If the variable does not belong to the solver, the function just returns,
   * or crashes in non-opt mode.
   */
  void SetCoefficient(const MPVariable* var, double coeff);

  /**
   * Gets the coefficient of a given variable on the constraint (which is 0 if
   * the variable does not appear in the constraint).
   */
  double GetCoefficient(const MPVariable* var) const;

  /**
   * Returns a map from variables to their coefficients in the constraint.
   *
   * If a variable is not present in the map, then its coefficient is zero.
   */
  const absl::flat_hash_map<const MPVariable*, double>& terms() const {
    return coefficients_;
  }

  /// Returns the lower bound.
  double lb() const { return lb_; }

  /// Returns the upper bound.
  double ub() const { return ub_; }

  /// Sets the lower bound.
  void SetLB(double lb) { SetBounds(lb, ub_); }

  /// Sets the upper bound.
  void SetUB(double ub) { SetBounds(lb_, ub); }

  /// Sets both the lower and upper bounds.
  void SetBounds(double lb, double ub);

  /// Advanced usage: returns true if the constraint is "lazy" (see below).
  bool is_lazy() const { return is_lazy_; }

  /**
   * Advanced usage: sets the constraint "laziness".
   *
   * <em>This is only supported for SCIP and has no effect on other
   * solvers.</em>
   *
   * When \b laziness is true, the constraint is only considered by the Linear
   * Programming solver if its current solution violates the constraint. In this
   * case, the constraint is definitively added to the problem. This may be
   * useful in some MIP problems, and may have a dramatic impact on performance.
   *
   * For more info see: http://tinyurl.com/lazy-constraints.
   */
  void set_is_lazy(bool laziness) { is_lazy_ = laziness; }

  const MPVariable* indicator_variable() const { return indicator_variable_; }
  bool indicator_value() const { return indicator_value_; }

  /// Returns the index of the constraint in the MPSolver::constraints_.
  int index() const { return index_; }

  /**
   * Advanced usage: returns the dual value of the constraint in the current
   * solution (only available for continuous problems).
   */
  double dual_value() const;

  /**
   * Advanced usage: returns the basis status of the constraint.
   *
   * It is only available for continuous problems).
   *
   * Note that if a constraint "linear_expression in [lb, ub]" is transformed
   * into "linear_expression + slack = 0" with slack in [-ub, -lb], then this
   * status is the same as the status of the slack variable with AT_UPPER_BOUND
   * and AT_LOWER_BOUND swapped.
   *
   * @see MPSolver::BasisStatus.
   */
  MPSolver::BasisStatus basis_status() const;

 protected:
  friend class MPSolver;
  friend class MPSolverInterface;
  friend class CBCInterface;
  friend class CLPInterface;
  friend class GLPKInterface;
  friend class SCIPInterface;
  friend class SLMInterface;
  friend class GurobiInterface;
  friend class CplexInterface;
  friend class XpressInterface;
  friend class GLOPInterface;
  friend class BopInterface;
  friend class SatInterface;
  friend class PdlpInterface;
  friend class HighsInterface;
  friend class KnapsackInterface;

  // Constructor. A constraint points to a single MPSolverInterface
  // that is specified in the constructor. A constraint cannot belong
  // to several models.
  MPConstraint(int index, double lb, double ub, const std::string& name,
               MPSolverInterface* const interface_in)
      : coefficients_(1),
        index_(index),
        lb_(lb),
        ub_(ub),
        name_(name.empty() ? absl::StrFormat("auto_c_%09d", index) : name),
        is_lazy_(false),
        indicator_variable_(nullptr),
        dual_value_(0.0),
        interface_(interface_in) {}

  void set_dual_value(double dual_value) { dual_value_ = dual_value; }

 private:
  // Returns true if the constraint contains variables that have not
  // been extracted yet.
  bool ContainsNewVariables();

  // Mapping var -> coefficient.
  absl::flat_hash_map<const MPVariable*, double> coefficients_;

  const int index_;  // See index().

  // The lower bound for the linear constraint.
  double lb_;

  // The upper bound for the linear constraint.
  double ub_;

  // Name.
  const std::string name_;

  // True if the constraint is "lazy", i.e. the constraint is added to the
  // underlying Linear Programming solver only if it is violated.
  // By default this parameter is 'false'.
  bool is_lazy_;

  // If given, this constraint is only active if `indicator_variable_`'s value
  // is equal to `indicator_value_`.
  const MPVariable* indicator_variable_;
  bool indicator_value_;

  double dual_value_;
  MPSolverInterface* const interface_;
};

/**
 * This class stores parameter settings for LP and MIP solvers. Some parameters
 * are marked as advanced: do not change their values unless you know what you
 * are doing!
 *
 * For developers: how to add a new parameter:
 * - Add the new Foo parameter in the DoubleParam or IntegerParam enum.
 * - If it is a categorical param, add a FooValues enum.
 * - Decide if the wrapper should define a default value for it: yes
 *   if it controls the properties of the solution (example:
 *   tolerances) or if it consistently improves performance, no
 *   otherwise. If yes, define kDefaultFoo.
 * - Add a foo_value_ member and, if no default value is defined, a
 *   foo_is_default_ member.
 * - Add code to handle Foo in Set...Param, Reset...Param,
 *   Get...Param, Reset and the constructor.
 * - In class MPSolverInterface, add a virtual method SetFoo, add it
 *   to SetCommonParameters or SetMIPParameters, and implement it for
 *   each solver. Sometimes, parameters need to be implemented
 *   differently, see for example the INCREMENTALITY implementation.
 * - Add a test in linear_solver_test.cc.
 *
 * TODO(user): store the parameter values in a protocol buffer
 * instead. We need to figure out how to deal with the subtleties of
 * the default values.
 */
class MPSolverParameters {
 public:
  /// Enumeration of parameters that take continuous values.
  enum DoubleParam {
    /// Limit for relative MIP gap.
    RELATIVE_MIP_GAP = 0,

    /**
     * Advanced usage: tolerance for primal feasibility of basic solutions.
     *
     * This does not control the integer feasibility tolerance of integer
     * solutions for MIP or the tolerance used during presolve.
     */
    PRIMAL_TOLERANCE = 1,
    /// Advanced usage: tolerance for dual feasibility of basic solutions.
    DUAL_TOLERANCE = 2
  };

  /// Enumeration of parameters that take integer or categorical values.
  enum IntegerParam {
    /// Advanced usage: presolve mode.
    PRESOLVE = 1000,
    /// Algorithm to solve linear programs.
    LP_ALGORITHM = 1001,
    /// Advanced usage: incrementality from one solve to the next.
    INCREMENTALITY = 1002,
    /// Advanced usage: enable or disable matrix scaling.
    SCALING = 1003
  };

  /// For each categorical parameter, enumeration of possible values.
  enum PresolveValues {
    /// Presolve is off.
    PRESOLVE_OFF = 0,
    /// Presolve is on.
    PRESOLVE_ON = 1
  };

  /// LP algorithm to use.
  enum LpAlgorithmValues {
    /// Dual simplex.
    DUAL = 10,
    /// Primal simplex.
    PRIMAL = 11,
    /// Barrier algorithm.
    BARRIER = 12
  };

  /// Advanced usage: Incrementality options.
  enum IncrementalityValues {
    /// Start solve from scratch.
    INCREMENTALITY_OFF = 0,

    /**
     * Reuse results from previous solve as much as the underlying solver
     * allows.
     */
    INCREMENTALITY_ON = 1
  };

  /// Advanced usage: Scaling options.
  enum ScalingValues {
    /// Scaling is off.
    SCALING_OFF = 0,
    /// Scaling is on.
    SCALING_ON = 1
  };

  // Placeholder value to indicate that a parameter is set to
  // the default value defined in the wrapper.
  static const double kDefaultDoubleParamValue;
  static const int kDefaultIntegerParamValue;

  // Placeholder value to indicate that a parameter is unknown.
  static const double kUnknownDoubleParamValue;
  static const int kUnknownIntegerParamValue;

  // Default values for parameters. Only parameters that define the
  // properties of the solution returned need to have a default value
  // (that is the same for all solvers). You can also define a default
  // value for performance parameters when you are confident it is a
  // good choice (example: always turn presolve on).
  static const double kDefaultRelativeMipGap;
  static const double kDefaultPrimalTolerance;
  static const double kDefaultDualTolerance;
  static const PresolveValues kDefaultPresolve;
  static const IncrementalityValues kDefaultIncrementality;

  /// The constructor sets all parameters to their default value.
  MPSolverParameters();

#ifndef SWIG
  // This type is neither copyable nor movable.
  MPSolverParameters(const MPSolverParameters&) = delete;
  MPSolverParameters& operator=(const MPSolverParameters&) = delete;
#endif

  /// Sets a double parameter to a specific value.
  void SetDoubleParam(MPSolverParameters::DoubleParam param, double value);

  /// Sets a integer parameter to a specific value.
  void SetIntegerParam(MPSolverParameters::IntegerParam param, int value);

  /**
   * Sets a double parameter to its default value (default value defined in
   * MPSolverParameters if it exists, otherwise the default value defined in
   * the underlying solver).
   */
  void ResetDoubleParam(MPSolverParameters::DoubleParam param);

  /**
   * Sets an integer parameter to its default value (default value defined in
   * MPSolverParameters if it exists, otherwise the default value defined in
   * the underlying solver).
   */
  void ResetIntegerParam(MPSolverParameters::IntegerParam param);

  /// Sets all parameters to their default value.
  void Reset();

  /// Returns the value of a double parameter.
  double GetDoubleParam(MPSolverParameters::DoubleParam param) const;

  /// Returns the value of an integer parameter.
  int GetIntegerParam(MPSolverParameters::IntegerParam param) const;

 private:
  // Parameter value for each parameter.
  // @see DoubleParam
  // @see IntegerParam
  double relative_mip_gap_value_;
  double primal_tolerance_value_;
  double dual_tolerance_value_;
  int presolve_value_;
  int scaling_value_;
  int lp_algorithm_value_;
  int incrementality_value_;

  // Boolean value indicating whether each parameter is set to the
  // solver's default value. Only parameters for which the wrapper
  // does not define a default value need such an indicator.
  bool lp_algorithm_is_default_;
};

// Whether the given MPSolverResponseStatus (of a solve) would yield an RPC
// error when happening on the linear solver stubby server, see
// ./linear_solver_service.proto.
// Note that RPC errors forbid to carry a response to the client, who can only
// see the RPC error itself (error code + error message).
bool MPSolverResponseStatusIsRpcError(MPSolverResponseStatus status);

// This class wraps the actual mathematical programming solvers. Each
// solver (GLOP, CLP, CBC, GLPK, SCIP) has its own interface class that
// derives from this abstract class. This class is never directly
// accessed by the user.
// @see glop_interface.cc
// @see cbc_interface.cc
// @see clp_interface.cc
// @see glpk_interface.cc
// @see scip_interface.cc
class MPSolverInterface {
 public:
  enum SynchronizationStatus {
    // The underlying solver (CLP, GLPK, ...) and MPSolver are not in
    // sync for the model nor for the solution.
    MUST_RELOAD,
    // The underlying solver and MPSolver are in sync for the model
    // but not for the solution: the model has changed since the
    // solution was computed last.
    MODEL_SYNCHRONIZED,
    // The underlying solver and MPSolver are in sync for the model and
    // the solution.
    SOLUTION_SYNCHRONIZED
  };

  // When the underlying solver does not provide the number of simplex
  // iterations.
  static constexpr int64_t kUnknownNumberOfIterations = -1;
  // When the underlying solver does not provide the number of
  // branch-and-bound nodes.
  static constexpr int64_t kUnknownNumberOfNodes = -1;

  // Constructor. The user will access the MPSolverInterface through the
  // MPSolver passed as argument.
  explicit MPSolverInterface(MPSolver* solver);
  virtual ~MPSolverInterface();

  // ----- Solve -----
  // Solves problem with specified parameter values. Returns true if the
  // solution is optimal.
  virtual MPSolver::ResultStatus Solve(const MPSolverParameters& param) = 0;

  // Attempts to directly solve a MPModelRequest, bypassing the MPSolver data
  // structures entirely. Like MPSolver::SolveWithProto(), optionally takes in
  // an 'interrupt' boolean.
  // Returns {} (eg. absl::nullopt) if direct-solve is not supported by the
  // underlying solver (possibly because interrupt != nullptr), in which case
  // the user should fall back to using MPSolver.
  virtual std::optional<MPSolutionResponse> DirectlySolveProto(
      const MPModelRequest& /*request*/,
      // `interrupt` is non-const because the internal
      // solver may set it to true itself, in some cases.
      std::atomic<bool>* /*interrupt*/) {
    return std::nullopt;
  }

  // Writes the model using the solver internal write function.  Currently only
  // available for GurobiInterface.
  virtual void Write(const std::string& filename);

  // ----- Model modifications and extraction -----
  // Resets extracted model.
  virtual void Reset() = 0;

  // Sets the optimization direction (min/max).
  virtual void SetOptimizationDirection(bool maximize) = 0;

  // Modifies bounds of an extracted variable.
  virtual void SetVariableBounds(int index, double lb, double ub) = 0;

  // Modifies integrality of an extracted variable.
  virtual void SetVariableInteger(int index, bool integer) = 0;

  // Modify bounds of an extracted variable.
  virtual void SetConstraintBounds(int index, double lb, double ub) = 0;

  // Adds a linear constraint.
  virtual void AddRowConstraint(MPConstraint* ct) = 0;

  // Adds an indicator constraint. Returns true if the feature is supported by
  // the underlying solver.
  virtual bool AddIndicatorConstraint(MPConstraint* const /*ct*/) {
    LOG(ERROR) << "Solver doesn't support indicator constraints.";
    return false;
  }

  // Add a variable.
  virtual void AddVariable(MPVariable* var) = 0;

  // Changes a coefficient in a constraint.
  virtual void SetCoefficient(MPConstraint* constraint,
                              const MPVariable* variable, double new_value,
                              double old_value) = 0;

  // Clears a constraint from all its terms.
  virtual void ClearConstraint(MPConstraint* constraint) = 0;

  // Changes a coefficient in the linear objective.
  virtual void SetObjectiveCoefficient(const MPVariable* variable,
                                       double coefficient) = 0;

  // Changes the constant term in the linear objective.
  virtual void SetObjectiveOffset(double value) = 0;

  // Clears the objective from all its terms.
  virtual void ClearObjective() = 0;

  virtual void BranchingPriorityChangedForVariable(int /*var_index*/) {}
  // ------ Query statistics on the solution and the solve ------
  // Returns the number of simplex iterations. The problem must be discrete,
  // otherwise it crashes, or returns kUnknownNumberOfIterations in NDEBUG mode.
  virtual int64_t iterations() const = 0;
  // Returns the number of branch-and-bound nodes. The problem must be discrete,
  // otherwise it crashes, or returns kUnknownNumberOfNodes in NDEBUG mode.
  virtual int64_t nodes() const = 0;
  // Returns the best objective bound. The problem must be discrete, otherwise
  // it crashes, or returns trivial bound (+/- inf) in NDEBUG mode.
  double best_objective_bound() const;
  // Returns the objective value of the best solution found so far.
  double objective_value() const;

  // Returns the basis status of a row.
  virtual MPSolver::BasisStatus row_status(int constraint_index) const = 0;
  // Returns the basis status of a constraint.
  virtual MPSolver::BasisStatus column_status(int variable_index) const = 0;

  // Checks whether the solution is synchronized with the model, i.e. whether
  // the model has changed since the solution was computed last.
  // If it isn't, it crashes in NDEBUG, and returns false otherwise.
  bool CheckSolutionIsSynchronized() const;
  // Checks whether a feasible solution exists. The behavior is similar to
  // CheckSolutionIsSynchronized() above.
  virtual bool CheckSolutionExists() const;
  // Handy shortcut to do both checks above (it is often used).
  bool CheckSolutionIsSynchronizedAndExists() const {
    return CheckSolutionIsSynchronized() && CheckSolutionExists();
  }

  // ----- Misc -----
  // Queries problem type. For simplicity, the distinction between
  // continuous and discrete is based on the declaration of the user
  // when the solver is created (example: GLPK_LINEAR_PROGRAMMING
  // vs. GLPK_MIXED_INTEGER_PROGRAMMING), not on the actual content of
  // the model.
  // Returns true if the problem is continuous.
  virtual bool IsContinuous() const = 0;
  // Returns true if the problem is continuous and linear.
  virtual bool IsLP() const = 0;
  // Returns true if the problem is discrete and linear.
  virtual bool IsMIP() const = 0;

  // Returns the index of the last variable extracted.
  int last_variable_index() const { return last_variable_index_; }

  bool variable_is_extracted(int var_index) const {
    return solver_->variable_is_extracted_[var_index];
  }
  void set_variable_as_extracted(int var_index, bool extracted) {
    solver_->variable_is_extracted_[var_index] = extracted;
  }
  bool constraint_is_extracted(int ct_index) const {
    return solver_->constraint_is_extracted_[ct_index];
  }
  void set_constraint_as_extracted(int ct_index, bool extracted) {
    solver_->constraint_is_extracted_[ct_index] = extracted;
  }

  // Returns the boolean indicating the verbosity of the solver output.
  bool quiet() const { return quiet_; }
  // Sets the boolean indicating the verbosity of the solver output.
  void set_quiet(bool quiet_value) { quiet_ = quiet_value; }

  // Returns the result status of the last solve.
  MPSolver::ResultStatus result_status() const {
    CheckSolutionIsSynchronized();
    return result_status_;
  }

  // Returns a string describing the underlying solver and its version.
  virtual std::string SolverVersion() const = 0;

  // Returns the underlying solver.
  virtual void* underlying_solver() = 0;

  // Computes exact condition number. Only available for continuous
  // problems and only implemented in GLPK.
  virtual double ComputeExactConditionNumber() const;

  // See MPSolver::SetStartingLpBasis().
  virtual void SetStartingLpBasis(
      const std::vector<MPSolver::BasisStatus>& /*variable_statuses*/,
      const std::vector<MPSolver::BasisStatus>& /*constraint_statuses*/) {
    LOG(FATAL) << "Not supported by this solver.";
  }

  virtual double infinity() { return std::numeric_limits<double>::infinity(); }

  virtual bool InterruptSolve() { return false; }

  // See MPSolver::NextSolution() for contract.
  virtual bool NextSolution() { return false; }

  // See MPSolver::SetCallback() for details.
  virtual void SetCallback(MPCallback* /*mp_callback*/) {
    LOG(FATAL) << "Callbacks not supported for this solver.";
  }

  virtual bool SupportsCallbacks() const { return false; }

  friend class MPSolver;

  // To access the maximize_ bool and the MPSolver.
  friend class MPConstraint;
  friend class MPObjective;

 protected:
  MPSolver* const solver_;
  // Indicates whether the model and the solution are synchronized.
  SynchronizationStatus sync_status_;
  // Indicates whether the solve has reached optimality,
  // infeasibility, a limit, etc.
  MPSolver::ResultStatus result_status_;
  // Optimization direction.
  bool maximize_;

  // Index in MPSolver::variables_ of last constraint extracted.
  int last_constraint_index_;
  // Index in MPSolver::constraints_ of last variable extracted.
  int last_variable_index_;

  // The value of the objective function.
  double objective_value_;

  // The value of the best objective bound. Used only for MIP solvers.
  double best_objective_bound_;

  // Boolean indicator for the verbosity of the solver output.
  bool quiet_;

  // Index of dummy variable created for empty constraints or the
  // objective offset.
  static const int kDummyVariableIndex;

  // Extracts model stored in MPSolver.
  void ExtractModel();
  // Extracts the variables that have not been extracted yet.
  virtual void ExtractNewVariables() = 0;
  // Extracts the constraints that have not been extracted yet.
  virtual void ExtractNewConstraints() = 0;
  // Extracts the objective.
  virtual void ExtractObjective() = 0;
  // Resets the extraction information.
  void ResetExtractionInformation();
  // Change synchronization status from SOLUTION_SYNCHRONIZED to
  // MODEL_SYNCHRONIZED. To be used for model changes.
  void InvalidateSolutionSynchronization();

  // Sets parameters common to LP and MIP in the underlying solver.
  void SetCommonParameters(const MPSolverParameters& param);
  // Sets MIP specific parameters in the underlying solver.
  void SetMIPParameters(const MPSolverParameters& param);
  // Sets all parameters in the underlying solver.
  virtual void SetParameters(const MPSolverParameters& param) = 0;
  // Sets an unsupported double parameter.
  void SetUnsupportedDoubleParam(MPSolverParameters::DoubleParam param);
  // Sets an unsupported integer parameter.
  virtual void SetUnsupportedIntegerParam(
      MPSolverParameters::IntegerParam param);
  // Sets a supported double parameter to an unsupported value.
  void SetDoubleParamToUnsupportedValue(MPSolverParameters::DoubleParam param,
                                        double value);
  // Sets a supported integer parameter to an unsupported value.
  virtual void SetIntegerParamToUnsupportedValue(
      MPSolverParameters::IntegerParam param, int value);
  // Sets each parameter in the underlying solver.
  virtual void SetRelativeMipGap(double value) = 0;
  virtual void SetPrimalTolerance(double value) = 0;
  virtual void SetDualTolerance(double value) = 0;
  virtual void SetPresolveMode(int value) = 0;

  // Sets the number of threads to be used by the solver.
  virtual absl::Status SetNumThreads(int num_threads);

  // Pass solver specific parameters in text format. The format is
  // solver-specific and is the same as the corresponding solver configuration
  // file format. Returns true if the operation was successful.
  //
  // Default implementation returns true if the input is empty. It returns false
  // and logs a WARNING if the input is not empty.
  virtual bool SetSolverSpecificParametersAsString(
      const std::string& parameters);

  // Sets the scaling mode.
  virtual void SetScalingMode(int value) = 0;
  virtual void SetLpAlgorithm(int value) = 0;
};

}  // namespace operations_research

#endif  // OR_TOOLS_LINEAR_SOLVER_LINEAR_SOLVER_H_