Package operations_research.pdlp

Interface Solvers.PrimalDualHybridGradientParamsOrBuilder

    • Method Detail

      • hasTerminationCriteria

        boolean hasTerminationCriteria()
        optional .operations_research.pdlp.TerminationCriteria termination_criteria = 1;
        Returns:
        Whether the terminationCriteria field is set.

      • getTerminationCriteria

        Solvers.TerminationCriteria getTerminationCriteria()
        optional .operations_research.pdlp.TerminationCriteria termination_criteria = 1;
        Returns:
        The terminationCriteria.

      • getTerminationCriteriaOrBuilder

        Solvers.TerminationCriteriaOrBuilder getTerminationCriteriaOrBuilder()
        optional .operations_research.pdlp.TerminationCriteria termination_criteria = 1;

      • hasNumThreads

        boolean hasNumThreads()
         The number of threads to use. Must be positive.
 Try various values of num_threads, up to the number of physical cores.
 Performance may not be monotonically increasing with the number of threads
 because of memory bandwidth limitations.
        optional int32 num_threads = 2 [default = 1];
        Returns:
        Whether the numThreads field is set.

      • getNumThreads

        int getNumThreads()
         The number of threads to use. Must be positive.
 Try various values of num_threads, up to the number of physical cores.
 Performance may not be monotonically increasing with the number of threads
 because of memory bandwidth limitations.
        optional int32 num_threads = 2 [default = 1];
        Returns:
        The numThreads.

      • hasNumShards

        boolean hasNumShards()
         For more efficient parallel computation, the matrices and vectors are
 divided (virtually) into num_shards shards. Results are computed
 independently for each shard and then combined. As a consequence, the order
 of computation, and hence floating point roundoff, depends on the number of
 shards so reproducible results require using the same value for num_shards.
 However, for efficiency num_shards should a be at least num_threads, and
 preferably at least 4*num_threads to allow better load balancing. If
 num_shards is positive, the computation will use that many shards.
 Otherwise a default that depends on num_threads will be used.
        optional int32 num_shards = 27 [default = 0];
        Returns:
        Whether the numShards field is set.

      • getNumShards

        int getNumShards()
         For more efficient parallel computation, the matrices and vectors are
 divided (virtually) into num_shards shards. Results are computed
 independently for each shard and then combined. As a consequence, the order
 of computation, and hence floating point roundoff, depends on the number of
 shards so reproducible results require using the same value for num_shards.
 However, for efficiency num_shards should a be at least num_threads, and
 preferably at least 4*num_threads to allow better load balancing. If
 num_shards is positive, the computation will use that many shards.
 Otherwise a default that depends on num_threads will be used.
        optional int32 num_shards = 27 [default = 0];
        Returns:
        The numShards.

      • hasRecordIterationStats

        boolean hasRecordIterationStats()
         If true, the iteration_stats field of the SolveLog output will be populated
 at every iteration. Note that we only compute solution statistics at
 termination checks. Setting this parameter to true may substantially
 increase the size of the output.
        optional bool record_iteration_stats = 3;
        Returns:
        Whether the recordIterationStats field is set.

      • getRecordIterationStats

        boolean getRecordIterationStats()
         If true, the iteration_stats field of the SolveLog output will be populated
 at every iteration. Note that we only compute solution statistics at
 termination checks. Setting this parameter to true may substantially
 increase the size of the output.
        optional bool record_iteration_stats = 3;
        Returns:
        The recordIterationStats.

      • hasVerbosityLevel

        boolean hasVerbosityLevel()
         The verbosity of logging.
 0: No informational logging. (Errors are logged.)
 1: Summary statistics only. No iteration-level details.
 2: A table of iteration-level statistics is logged.
    (See ToShortString() in primal_dual_hybrid_gradient.cc).
 3: A more detailed table of iteration-level statistics is logged.
    (See ToString() in primal_dual_hybrid_gradient.cc).
 4: For iteration-level details, prints the statistics of both the average
    (prefixed with A) and the current iterate (prefixed with C). Also prints
    internal algorithmic state and details.
 Logging at levels 2-4 also includes messages from level 1.
        optional int32 verbosity_level = 26 [default = 0];
        Returns:
        Whether the verbosityLevel field is set.

      • getVerbosityLevel

        int getVerbosityLevel()
         The verbosity of logging.
 0: No informational logging. (Errors are logged.)
 1: Summary statistics only. No iteration-level details.
 2: A table of iteration-level statistics is logged.
    (See ToShortString() in primal_dual_hybrid_gradient.cc).
 3: A more detailed table of iteration-level statistics is logged.
    (See ToString() in primal_dual_hybrid_gradient.cc).
 4: For iteration-level details, prints the statistics of both the average
    (prefixed with A) and the current iterate (prefixed with C). Also prints
    internal algorithmic state and details.
 Logging at levels 2-4 also includes messages from level 1.
        optional int32 verbosity_level = 26 [default = 0];
        Returns:
        The verbosityLevel.

      • hasLogIntervalSeconds

        boolean hasLogIntervalSeconds()
         Time between iteration-level statistics logging (if `verbosity_level > 1`).
 Since iteration-level statistics are only generated when performing
 termination checks, logs will be generated from next termination check
 after `log_interval_seconds` have elapsed. Should be >= 0.0. 0.0 (the
 default) means log statistics at every termination check.
        optional double log_interval_seconds = 31 [default = 0];
        Returns:
        Whether the logIntervalSeconds field is set.

      • getLogIntervalSeconds

        double getLogIntervalSeconds()
         Time between iteration-level statistics logging (if `verbosity_level > 1`).
 Since iteration-level statistics are only generated when performing
 termination checks, logs will be generated from next termination check
 after `log_interval_seconds` have elapsed. Should be >= 0.0. 0.0 (the
 default) means log statistics at every termination check.
        optional double log_interval_seconds = 31 [default = 0];
        Returns:
        The logIntervalSeconds.

      • hasMajorIterationFrequency

        boolean hasMajorIterationFrequency()
         The frequency at which extra work is performed to make major algorithmic
 decisions, e.g., performing restarts and updating the primal weight. Major
 iterations also trigger a termination check. For best performance using the
 NO_RESTARTS or EVERY_MAJOR_ITERATION rule, one should perform a log-scale
 grid search over this parameter, for example, over powers of two.
 ADAPTIVE_HEURISTIC is mostly insensitive to this value.
        optional int32 major_iteration_frequency = 4 [default = 64];
        Returns:
        Whether the majorIterationFrequency field is set.

      • getMajorIterationFrequency

        int getMajorIterationFrequency()
         The frequency at which extra work is performed to make major algorithmic
 decisions, e.g., performing restarts and updating the primal weight. Major
 iterations also trigger a termination check. For best performance using the
 NO_RESTARTS or EVERY_MAJOR_ITERATION rule, one should perform a log-scale
 grid search over this parameter, for example, over powers of two.
 ADAPTIVE_HEURISTIC is mostly insensitive to this value.
        optional int32 major_iteration_frequency = 4 [default = 64];
        Returns:
        The majorIterationFrequency.

      • hasTerminationCheckFrequency

        boolean hasTerminationCheckFrequency()
         The frequency (based on a counter reset every major iteration) to check for
 termination (involves extra work) and log iteration stats. Termination
 checks do not affect algorithmic progress unless termination is triggered.
        optional int32 termination_check_frequency = 5 [default = 64];
        Returns:
        Whether the terminationCheckFrequency field is set.

      • getTerminationCheckFrequency

        int getTerminationCheckFrequency()
         The frequency (based on a counter reset every major iteration) to check for
 termination (involves extra work) and log iteration stats. Termination
 checks do not affect algorithmic progress unless termination is triggered.
        optional int32 termination_check_frequency = 5 [default = 64];
        Returns:
        The terminationCheckFrequency.

      • hasRestartStrategy

        boolean hasRestartStrategy()
         NO_RESTARTS and EVERY_MAJOR_ITERATION occasionally outperform the default.
 If using a strategy other than ADAPTIVE_HEURISTIC, you must also tune
 major_iteration_frequency.
        optional .operations_research.pdlp.PrimalDualHybridGradientParams.RestartStrategy restart_strategy = 6 [default = ADAPTIVE_HEURISTIC];
        Returns:
        Whether the restartStrategy field is set.

      • getRestartStrategy

        Solvers.PrimalDualHybridGradientParams.RestartStrategy getRestartStrategy()
         NO_RESTARTS and EVERY_MAJOR_ITERATION occasionally outperform the default.
 If using a strategy other than ADAPTIVE_HEURISTIC, you must also tune
 major_iteration_frequency.
        optional .operations_research.pdlp.PrimalDualHybridGradientParams.RestartStrategy restart_strategy = 6 [default = ADAPTIVE_HEURISTIC];
        Returns:
        The restartStrategy.

      • hasPrimalWeightUpdateSmoothing

        boolean hasPrimalWeightUpdateSmoothing()
         This parameter controls exponential smoothing of log(primal_weight) when a
 primal weight update occurs (i.e., when the ratio of primal and dual step
 sizes is adjusted). At 0.0, the primal weight will be frozen at its initial
 value and there will be no dynamic updates in the algorithm. At 1.0, there
 is no smoothing in the updates. The default of 0.5 generally performs well,
 but has been observed on occasion to trigger unstable swings in the primal
 weight. We recommend also trying 0.0 (disabling primal weight updates), in
 which case you must also tune initial_primal_weight.
        optional double primal_weight_update_smoothing = 7 [default = 0.5];
        Returns:
        Whether the primalWeightUpdateSmoothing field is set.

      • getPrimalWeightUpdateSmoothing

        double getPrimalWeightUpdateSmoothing()
         This parameter controls exponential smoothing of log(primal_weight) when a
 primal weight update occurs (i.e., when the ratio of primal and dual step
 sizes is adjusted). At 0.0, the primal weight will be frozen at its initial
 value and there will be no dynamic updates in the algorithm. At 1.0, there
 is no smoothing in the updates. The default of 0.5 generally performs well,
 but has been observed on occasion to trigger unstable swings in the primal
 weight. We recommend also trying 0.0 (disabling primal weight updates), in
 which case you must also tune initial_primal_weight.
        optional double primal_weight_update_smoothing = 7 [default = 0.5];
        Returns:
        The primalWeightUpdateSmoothing.

      • hasInitialPrimalWeight

        boolean hasInitialPrimalWeight()
         The initial value of the primal weight (i.e., the ratio of primal and dual
 step sizes). The primal weight remains fixed throughout the solve if
 primal_weight_update_smoothing = 0.0. If unset, the default is the ratio of
 the norm of the objective vector to the L2 norm of the combined constraint
 bounds vector (as defined above). If this ratio is not finite and positive,
 then the default is 1.0 instead. For tuning, try powers of 10, for example,
 from 10^{-6} to 10^6.
        optional double initial_primal_weight = 8;
        Returns:
        Whether the initialPrimalWeight field is set.

      • getInitialPrimalWeight

        double getInitialPrimalWeight()
         The initial value of the primal weight (i.e., the ratio of primal and dual
 step sizes). The primal weight remains fixed throughout the solve if
 primal_weight_update_smoothing = 0.0. If unset, the default is the ratio of
 the norm of the objective vector to the L2 norm of the combined constraint
 bounds vector (as defined above). If this ratio is not finite and positive,
 then the default is 1.0 instead. For tuning, try powers of 10, for example,
 from 10^{-6} to 10^6.
        optional double initial_primal_weight = 8;
        Returns:
        The initialPrimalWeight.

      • hasPresolveOptions

        boolean hasPresolveOptions()
        optional .operations_research.pdlp.PrimalDualHybridGradientParams.PresolveOptions presolve_options = 16;
        Returns:
        Whether the presolveOptions field is set.

      • hasLInfRuizIterations

        boolean hasLInfRuizIterations()
         Number of L_infinity Ruiz rescaling iterations to apply to the constraint
 matrix. Zero disables this rescaling pass. Recommended values to try when
 tuning are 0, 5, and 10.
        optional int32 l_inf_ruiz_iterations = 9 [default = 5];
        Returns:
        Whether the lInfRuizIterations field is set.

      • getLInfRuizIterations

        int getLInfRuizIterations()
         Number of L_infinity Ruiz rescaling iterations to apply to the constraint
 matrix. Zero disables this rescaling pass. Recommended values to try when
 tuning are 0, 5, and 10.
        optional int32 l_inf_ruiz_iterations = 9 [default = 5];
        Returns:
        The lInfRuizIterations.

      • hasL2NormRescaling

        boolean hasL2NormRescaling()
         If true, applies L_2 norm rescaling after the Ruiz rescaling. Heuristically
 this has been found to help convergence.
        optional bool l2_norm_rescaling = 10 [default = true];
        Returns:
        Whether the l2NormRescaling field is set.

      • getL2NormRescaling

        boolean getL2NormRescaling()
         If true, applies L_2 norm rescaling after the Ruiz rescaling. Heuristically
 this has been found to help convergence.
        optional bool l2_norm_rescaling = 10 [default = true];
        Returns:
        The l2NormRescaling.

      • hasSufficientReductionForRestart

        boolean hasSufficientReductionForRestart()
         For ADAPTIVE_HEURISTIC and ADAPTIVE_DISTANCE_BASED only: A relative
 reduction in the potential function by this amount always triggers a
 restart. Must be between 0.0 and 1.0.
        optional double sufficient_reduction_for_restart = 11 [default = 0.1];
        Returns:
        Whether the sufficientReductionForRestart field is set.

      • getSufficientReductionForRestart

        double getSufficientReductionForRestart()
         For ADAPTIVE_HEURISTIC and ADAPTIVE_DISTANCE_BASED only: A relative
 reduction in the potential function by this amount always triggers a
 restart. Must be between 0.0 and 1.0.
        optional double sufficient_reduction_for_restart = 11 [default = 0.1];
        Returns:
        The sufficientReductionForRestart.

      • hasNecessaryReductionForRestart

        boolean hasNecessaryReductionForRestart()
         For ADAPTIVE_HEURISTIC only: A relative reduction in the potential function
 by this amount triggers a restart if, additionally, the quality of the
 iterates appears to be getting worse. The value must be in the interval
 [sufficient_reduction_for_restart, 1). Smaller values make restarts less
 frequent, and larger values make them more frequent.
        optional double necessary_reduction_for_restart = 17 [default = 0.9];
        Returns:
        Whether the necessaryReductionForRestart field is set.

      • getNecessaryReductionForRestart

        double getNecessaryReductionForRestart()
         For ADAPTIVE_HEURISTIC only: A relative reduction in the potential function
 by this amount triggers a restart if, additionally, the quality of the
 iterates appears to be getting worse. The value must be in the interval
 [sufficient_reduction_for_restart, 1). Smaller values make restarts less
 frequent, and larger values make them more frequent.
        optional double necessary_reduction_for_restart = 17 [default = 0.9];
        Returns:
        The necessaryReductionForRestart.

      • hasLinesearchRule

        boolean hasLinesearchRule()
         Linesearch rule applied at each major iteration.
        optional .operations_research.pdlp.PrimalDualHybridGradientParams.LinesearchRule linesearch_rule = 12 [default = ADAPTIVE_LINESEARCH_RULE];
        Returns:
        Whether the linesearchRule field is set.

      • getLinesearchRule

        Solvers.PrimalDualHybridGradientParams.LinesearchRule getLinesearchRule()
         Linesearch rule applied at each major iteration.
        optional .operations_research.pdlp.PrimalDualHybridGradientParams.LinesearchRule linesearch_rule = 12 [default = ADAPTIVE_LINESEARCH_RULE];
        Returns:
        The linesearchRule.

      • hasAdaptiveLinesearchParameters

        boolean hasAdaptiveLinesearchParameters()
        optional .operations_research.pdlp.AdaptiveLinesearchParams adaptive_linesearch_parameters = 18;
        Returns:
        Whether the adaptiveLinesearchParameters field is set.

      • getAdaptiveLinesearchParameters

        Solvers.AdaptiveLinesearchParams getAdaptiveLinesearchParameters()
        optional .operations_research.pdlp.AdaptiveLinesearchParams adaptive_linesearch_parameters = 18;
        Returns:
        The adaptiveLinesearchParameters.

      • getAdaptiveLinesearchParametersOrBuilder

        Solvers.AdaptiveLinesearchParamsOrBuilder getAdaptiveLinesearchParametersOrBuilder()
        optional .operations_research.pdlp.AdaptiveLinesearchParams adaptive_linesearch_parameters = 18;

      • hasMalitskyPockParameters

        boolean hasMalitskyPockParameters()
        optional .operations_research.pdlp.MalitskyPockParams malitsky_pock_parameters = 19;
        Returns:
        Whether the malitskyPockParameters field is set.

      • getMalitskyPockParameters

        Solvers.MalitskyPockParams getMalitskyPockParameters()
        optional .operations_research.pdlp.MalitskyPockParams malitsky_pock_parameters = 19;
        Returns:
        The malitskyPockParameters.

      • getMalitskyPockParametersOrBuilder

        Solvers.MalitskyPockParamsOrBuilder getMalitskyPockParametersOrBuilder()
        optional .operations_research.pdlp.MalitskyPockParams malitsky_pock_parameters = 19;

      • hasInitialStepSizeScaling

        boolean hasInitialStepSizeScaling()
         Scaling factor applied to the initial step size (all step sizes if
 linesearch_rule == CONSTANT_STEP_SIZE_RULE).
        optional double initial_step_size_scaling = 25 [default = 1];
        Returns:
        Whether the initialStepSizeScaling field is set.

      • getInitialStepSizeScaling

        double getInitialStepSizeScaling()
         Scaling factor applied to the initial step size (all step sizes if
 linesearch_rule == CONSTANT_STEP_SIZE_RULE).
        optional double initial_step_size_scaling = 25 [default = 1];
        Returns:
        The initialStepSizeScaling.

      • getRandomProjectionSeedsList

        java.util.List<java.lang.Integer> getRandomProjectionSeedsList()
         Seeds for generating (pseudo-)random projections of iterates during
 termination checks. For each seed, the projection of the primal and dual
 solutions onto random planes in primal and dual space will be computed and
 added the IterationStats if record_iteration_stats is true. The random
 planes generated will be determined by the seeds, the primal and dual
 dimensions, and num_threads.
        repeated int32 random_projection_seeds = 28 [packed = true];
        Returns:
        A list containing the randomProjectionSeeds.

      • getRandomProjectionSeedsCount

        int getRandomProjectionSeedsCount()
         Seeds for generating (pseudo-)random projections of iterates during
 termination checks. For each seed, the projection of the primal and dual
 solutions onto random planes in primal and dual space will be computed and
 added the IterationStats if record_iteration_stats is true. The random
 planes generated will be determined by the seeds, the primal and dual
 dimensions, and num_threads.
        repeated int32 random_projection_seeds = 28 [packed = true];
        Returns:
        The count of randomProjectionSeeds.

      • getRandomProjectionSeeds

        int getRandomProjectionSeeds​(int index)
         Seeds for generating (pseudo-)random projections of iterates during
 termination checks. For each seed, the projection of the primal and dual
 solutions onto random planes in primal and dual space will be computed and
 added the IterationStats if record_iteration_stats is true. The random
 planes generated will be determined by the seeds, the primal and dual
 dimensions, and num_threads.
        repeated int32 random_projection_seeds = 28 [packed = true];
        Parameters:
        index - The index of the element to return.
        Returns:
        The randomProjectionSeeds at the given index.

      • hasInfiniteConstraintBoundThreshold

        boolean hasInfiniteConstraintBoundThreshold()
         Constraint bounds with absolute value at least this threshold are replaced
 with infinities.
 NOTE: This primarily affects the relative convergence criteria. A smaller
 value makes the relative convergence criteria stronger. It also affects the
 problem statistics LOG()ed at the start of the run, and the default initial
 primal weight, since that is based on the norm of the bounds.
        optional double infinite_constraint_bound_threshold = 22 [default = inf];
        Returns:
        Whether the infiniteConstraintBoundThreshold field is set.

      • getInfiniteConstraintBoundThreshold

        double getInfiniteConstraintBoundThreshold()
         Constraint bounds with absolute value at least this threshold are replaced
 with infinities.
 NOTE: This primarily affects the relative convergence criteria. A smaller
 value makes the relative convergence criteria stronger. It also affects the
 problem statistics LOG()ed at the start of the run, and the default initial
 primal weight, since that is based on the norm of the bounds.
        optional double infinite_constraint_bound_threshold = 22 [default = inf];
        Returns:
        The infiniteConstraintBoundThreshold.

      • hasHandleSomePrimalGradientsOnFiniteBoundsAsResiduals

        boolean hasHandleSomePrimalGradientsOnFiniteBoundsAsResiduals()
         See
 https://developers.google.com/optimization/lp/pdlp_math#treating_some_variable_bounds_as_infinite
 for a description of this flag.
        optional bool handle_some_primal_gradients_on_finite_bounds_as_residuals = 29 [default = true];
        Returns:
        Whether the handleSomePrimalGradientsOnFiniteBoundsAsResiduals field is set.

      • getHandleSomePrimalGradientsOnFiniteBoundsAsResiduals

        boolean getHandleSomePrimalGradientsOnFiniteBoundsAsResiduals()
         See
 https://developers.google.com/optimization/lp/pdlp_math#treating_some_variable_bounds_as_infinite
 for a description of this flag.
        optional bool handle_some_primal_gradients_on_finite_bounds_as_residuals = 29 [default = true];
        Returns:
        The handleSomePrimalGradientsOnFiniteBoundsAsResiduals.

      • hasUseDiagonalQpTrustRegionSolver

        boolean hasUseDiagonalQpTrustRegionSolver()
         When solving QPs with diagonal objective matrices, this option can be
 turned on to enable an experimental solver that avoids linearization of the
 quadratic term. The `diagonal_qp_solver_accuracy` parameter controls the
 solve accuracy.
 TODO(user): Turn this option on by default for quadratic
 programs after numerical evaluation.
        optional bool use_diagonal_qp_trust_region_solver = 23 [default = false];
        Returns:
        Whether the useDiagonalQpTrustRegionSolver field is set.

      • getUseDiagonalQpTrustRegionSolver

        boolean getUseDiagonalQpTrustRegionSolver()
         When solving QPs with diagonal objective matrices, this option can be
 turned on to enable an experimental solver that avoids linearization of the
 quadratic term. The `diagonal_qp_solver_accuracy` parameter controls the
 solve accuracy.
 TODO(user): Turn this option on by default for quadratic
 programs after numerical evaluation.
        optional bool use_diagonal_qp_trust_region_solver = 23 [default = false];
        Returns:
        The useDiagonalQpTrustRegionSolver.

      • hasDiagonalQpTrustRegionSolverTolerance

        boolean hasDiagonalQpTrustRegionSolverTolerance()
         The solve tolerance of the experimental trust region solver for diagonal
 QPs, controlling the accuracy of binary search over a one-dimensional
 scaling parameter. Smaller values imply smaller relative error of the final
 solution vector.
 TODO(user): Find an expression for the final relative error.
        optional double diagonal_qp_trust_region_solver_tolerance = 24 [default = 1e-08];
        Returns:
        Whether the diagonalQpTrustRegionSolverTolerance field is set.

      • getDiagonalQpTrustRegionSolverTolerance

        double getDiagonalQpTrustRegionSolverTolerance()
         The solve tolerance of the experimental trust region solver for diagonal
 QPs, controlling the accuracy of binary search over a one-dimensional
 scaling parameter. Smaller values imply smaller relative error of the final
 solution vector.
 TODO(user): Find an expression for the final relative error.
        optional double diagonal_qp_trust_region_solver_tolerance = 24 [default = 1e-08];
        Returns:
        The diagonalQpTrustRegionSolverTolerance.

      • hasUseFeasibilityPolishing

        boolean hasUseFeasibilityPolishing()
         If true, periodically runs feasibility polishing, which attempts to move
 from latest average iterate to one that is closer to feasibility (i.e., has
 smaller primal and dual residuals) while probably increasing the objective
 gap. This is useful primarily when the feasibility tolerances are fairly
 tight and the objective gap tolerance is somewhat looser. Note that this
 does not change the termination criteria, but rather can help achieve the
 termination criteria more quickly when the objective gap is not as
 important as feasibility.

 `use_feasibility_polishing` cannot be used with glop presolve, and requires
 `handle_some_primal_gradients_on_finite_bounds_as_residuals == false`.
 `use_feasibility_polishing` can only be used with linear programs.

 Feasibility polishing runs two separate phases, primal feasibility and dual
 feasibility. The primal feasibility phase runs PDHG on the primal
 feasibility problem (obtained by changing the objective vector to all
 zeros), using the average primal iterate and zero dual (which is optimal
 for the primal feasibility problem) as the initial solution. The dual
 feasibility phase runs PDHG on the dual feasibility problem (obtained by
 changing all finite variable and constraint bounds to zero), using the
 average dual iterate and zero primal (which is optimal for the dual
 feasibility problem) as the initial solution. The primal solution from the
 primal feasibility phase and dual solution from the dual feasibility phase
 are then combined (forming a solution of type
 `POINT_TYPE_FEASIBILITY_POLISHING_SOLUTION`) and checked against the
 termination criteria.
        optional bool use_feasibility_polishing = 30 [default = false];
        Returns:
        Whether the useFeasibilityPolishing field is set.

      • getUseFeasibilityPolishing

        boolean getUseFeasibilityPolishing()
         If true, periodically runs feasibility polishing, which attempts to move
 from latest average iterate to one that is closer to feasibility (i.e., has
 smaller primal and dual residuals) while probably increasing the objective
 gap. This is useful primarily when the feasibility tolerances are fairly
 tight and the objective gap tolerance is somewhat looser. Note that this
 does not change the termination criteria, but rather can help achieve the
 termination criteria more quickly when the objective gap is not as
 important as feasibility.

 `use_feasibility_polishing` cannot be used with glop presolve, and requires
 `handle_some_primal_gradients_on_finite_bounds_as_residuals == false`.
 `use_feasibility_polishing` can only be used with linear programs.

 Feasibility polishing runs two separate phases, primal feasibility and dual
 feasibility. The primal feasibility phase runs PDHG on the primal
 feasibility problem (obtained by changing the objective vector to all
 zeros), using the average primal iterate and zero dual (which is optimal
 for the primal feasibility problem) as the initial solution. The dual
 feasibility phase runs PDHG on the dual feasibility problem (obtained by
 changing all finite variable and constraint bounds to zero), using the
 average dual iterate and zero primal (which is optimal for the dual
 feasibility problem) as the initial solution. The primal solution from the
 primal feasibility phase and dual solution from the dual feasibility phase
 are then combined (forming a solution of type
 `POINT_TYPE_FEASIBILITY_POLISHING_SOLUTION`) and checked against the
 termination criteria.
        optional bool use_feasibility_polishing = 30 [default = false];
        Returns:
        The useFeasibilityPolishing.