|
MoochoPack : Framework for Large-Scale Optimization Algorithms
Version of the Day
|
|
MoochoPack : Framework for Large-Scale Optimization Algorithms
Version of the Day
|
Below are all of the options that MOOCHO will accept for the the "MamaJama" algorithm configuration with full documentation. This is the file that is returned by generate-opt-file.pl with no options. To view these same options stripped of most of the comments see here.
*** Automatically generated options file
begin_options
*** Begin Moocho.opt.MoochoSolver
*************************************************************************************
*** All of these options can be used with the class MoochoSolver.
***
*** This file will be maintained and will include every option that
*** users can set. Most of these options the user will want to
*** leave alone but they are there in any case.
***
*** For options specific to the NLPAlgoConfigMamaJama configuration
*** class see the file 'Moocho.opt.NLPAlgoConfigMamaJama'.
***
**********************************************************
*** Options specific for the MoochoSolver class.
***
*** These options work on the highest level in determining
*** what output files are allowed, workspace requirements,
*** objective function scaling etc.
***
options_group MoochoSolver {
* workspace_MB = -1.0; *** [default]
*** (+-dbl) If > 0, givens the number of megabytes that are allocated for
*** temporary workspace for automatic arrays. If < 0 then the
*** this will be determined internally. This value should be set by the
*** user for whatever is appropriate for the computing environment. See
*** the summary output for statistics on memory allocation usage when the
*** algorithm finishes. Example values:
*** -1.0 : (< 0) Allow the algorithm to decide how much to allocate.
*** 100 : Allocate 100 MB of ram for this purpose
* obj_scale = 1.0; *** [default]
*** (+-dbl) Scale for the objective function. This can have a dramatic impact
*** on the behavior of the algorithm in some cases. Changing this value
*** does a lot to change the weight between minimizing the objective and converging
*** the constraints. Example values:
*** 1e-8 : Really scale the objective down a lot!
*** 1.0 : Leave the objective unscaled [default].
*** 1e+8 : Really scale the objective up a lot!
* test_nlp = true; *** [default]
* test_nlp = false;
*** If true then the NLP will be tested at the initial point. The
*** vector spaces (see options_group VectorSpaceTester),
*** the NLP interface (see options_group NLPTester), and the gradients
*** (see options_group NLPDirectTester and NLPFirstOrderInfoTester)
*** will all be tested. With all of the default options in place
*** these tests are fairly cheap so it is recommended that you perform
*** these tests when first getting started.
* console_outputting = true; *** [default]
* console_outputting = false;
*** If true, then output from MoochoTrackerConsoleStd is sent to the
*** console_out stream (which is std::cout by default)
* summary_outputting = true; *** [default]
* summary_outputting = false;
*** If true, then output from MoochoTrackerSummaryStd is sent to the
*** summary_out stream (which is the file 'MoochoSummary.out' by default)
* journal_outputting = true; *** [default]
* journal_outputting = false;
*** If true, then output from the algorithm steps an other detailed testing
*** output is set to the journal_out stream (which is the file
*** 'MoochoJournal.out' by default)
* algo_outputting = true; *** [default]
* algo_outputting = false;
*** If true, then an algorithm description is sent to the algo_out stream
*** (which is the file 'MoochoAlgo.out' by default)
* print_algo = true; *** [default]
* print_algo = false;
*** [algo_outputting == true]
*** If true then the algorithm will be printed to the algo_out stream
*** (the file 'NLPAlgo.out' by default). In order to get more insight into
*** what all of the options do it is a good idea to print the algorithm
*** description out and search for the options you are curious about.
* algo_timing = true; *** [default]
* algo_timing = false;
*** [summary_outputting == true]
*** If true, then the steps in the algorithm will be timed and a table of the
*** algorithm and step times will be sent to the summary_out stream (the file
*** 'MoochoSummary.out' by default). This feature is very useful in examining
*** performance of the algorithm and can give more detailed information than you
*** get from a profiler in may ways.
* generate_stats_file = true;
* generate_stats_file = false; *** [default]
*** If true, then a MoochoTrackerStatsStd object will be used to generate
*** statistics about the solution process to an NLP. The track object
*** will overwrite the file 'MoochoStats.out' in the current directory.
* print_opt_grp_not_accessed = true; *** [default]
* print_opt_grp_not_accessed = false;
*** [algo_outputting == true]
*** If true, then the options groups that are specified but are not read by
*** some software entity are printed to the algo_out stream (the file 'MoochoAlgo.out'
*** by default). This may help you catch problems associated with spelling the name of
*** an options group improperly and then having its default options used instead of the
*** options that you set. Note that some options groups are only looked for depending
*** on option values from other options groups.
* configuration = mama_jama; *** [default]
* configuration = interior_point;
*** decides which configuration object will be used:
*** mama_jama : standard reduced-space SQP configuration
*** interior_point : configuration for a simple reduced-space interior point method
}
***********************************************************
*** Options for NLPSolverClientInterface
***
*** These are options that are used by every algorithm
*** configuration and they are available to the
*** optimization steps.
***
*** These include basic algorithmic control parameters.
***
*** Note that an algorithm can be interrupted at any time
*** by pressing Ctrl+C.
***
options_group NLPSolverClientInterface {
* max_iter = 1000; *** [default]
*** (+int) Maximum number of SQP iterations allowed.
* max_run_time = 1e+10; *** In minutes [default]
*** (+dbl) Maximum runtime for the SQP algorithm (in minutes).
*** The default is to run forever.
* opt_tol = 1e-6; *** [default]
*** (+dbl) Convergence tolerance for (scaled) KKT linear dependence of gradients.
*** This is usually the hardest error to reduce. The exact definition of this tolerance
*** depends on the algorithms used and may use different scalings (see other options and
*** outputted algorithm description for details). Example values:
*** 0.0 : The algorithm will never converge except in trivial cases.
*** 1e-6 : Only converge when opt_kkt_err is below this value [default].
*** 1e+50 : (big number) Converged when any of the other tolerances are satisfied.
* feas_tol = 1e-6; *** [default]
*** (+dbl) Convergence tolerance for (scaled) feasibility of the equality constraints
*** ||c(x)||inf. The norm of the constraints ||c(x)||inf may be scaled (see other
*** options and the outputted algorithm description). Example values:
*** 0.0 : Never converge the algorithm except in trivial cases.
*** 1e-6 : Only converge when feas_kkt_err is below this value [default].
*** 1e+50 : (big number) Converged when any of the other tolerances are satisfied.
* step_tol = 1e-2; *** [default]
*** (+dbl) Convergence tolerance for (scaled) step size ||x_k(+1)-x_k||inf.
*** This tolerance is usually scaled by x is some way (see other output algorithm
*** description). Example values:
*** 0.0 : Never converge the algorithm except in trivial cases.
*** 1e-2 : Only converge when the max (scaled) step size is below this value [default].
*** 1e+50 : (big number) Converged when any of the other tolerances are satisfied.
* journal_output_level = PRINT_NOTHING; * No output to journal from algorithm
* journal_output_level = PRINT_BASIC_ALGORITHM_INFO; * O(1) information usually
* journal_output_level = PRINT_ALGORITHM_STEPS; * O(iter) output to journal [default]
* journal_output_level = PRINT_ACTIVE_SET; * O(iter*nact) output to journal
* journal_output_level = PRINT_VECTORS; * O(iter*n) output to journal (lots!)
* journal_output_level = PRINT_ITERATION_QUANTITIES; * O(iter*n*m) output to journal (big lots!)
*** [MoochoSolver::journal_outputting == true]
*** This option determines the type and amount of output to the journal_out stream
*** (the file 'MoochoJournal.out' by default) that is generated while the algorithm runs.
*** In general, each increasing value includes the same output from the lower options
*** (i.e. PRINT_VECTORS includes all the output for PRINT_ACTIVE_SET and more). Above,
*** the identifier 'iter' is the number of total rSQP iterations (see max_iter above), 'nact'
*** is the total number of active inequality constraints, 'n' is the total number
*** of NLP variables, and 'm' is the total number of equality constraints. The higher output
*** values are generally used for debugging. For most problems the value
*** PRINT_ALGORITHM_STEPS is usually the most appropriate and will give a great deal
*** of information about the algorithm without generating excessive output.
*** For the fastest possible execution you should set this to PRINT_NOTHING.
* null_space_journal_output_level = DEFAULT; * Set to journal_output_level [default]
* null_space_journal_output_level = PRINT_ACTIVE_SET; * O(iter*nact) output to journal
* null_space_journal_output_level = PRINT_VECTORS; * O(iter*(n-m)) output to journal (lots!)
* null_space_journal_output_level = PRINT_ITERATION_QUANTITIES; * O(iter*(n-m)^2) output to journal (big lots!)
*** [MoochoSolver::journal_outputting == true]
*** This option determines the type and amount of output to the journal_out stream
*** (the file 'MoochoJournal.out' by default) that is generated for quantities in the
*** null space while the algorithm runs. If null_space_journal_output_level is
*** set to DEFAULT then it will default to the value of journal_output_level.
*** If set to some other value then this value overrides journal_output_level
*** for quantities in the null space. For problems where the null space is small but
*** the full space is much larger, setting the value of null_space_journal_output_level higher
*** than journal_output_level can yield significantly more information while not generating
*** too much output or impacting runtime to any great extent.
* journal_print_digits = 6; *** [default]
*** [MoochoSolver::journal_outputting == true]
*** (+int) Number of decimal significant figures to print to journal_out stream.
*** With a higher number more significant figures will be printed. This may be useful
*** for debugging or in seeing the effects of subtle rounding differences. For IEEE double
*** precision, 18 is usually the maximum number of unique decimal significant figures.
* check_results = true; *** (costly?)
* check_results = false; *** [default]
*** If true then all computation that can be reasonably checked will be checked at runtime.
*** When all of the other default testing options are used, this overhead usually will
*** not dominate the cost of the algorithm so if speed is not critical then it is a
*** good idea to turning testing on. If your problem is not solving then you should
*** definitely try turning this on and try to see if it will catch any errors. However,
*** for the fastest possible execution you should set this to 'false'.
* calc_conditioning = true; *** (costly?)
* calc_conditioning = false; *** [default]
*** If true then estimates of the condition numbers of all of the important nonsingular
*** matrices used in the algorithm will be computed and printed. Note that this can be
*** a fairly expensive operation (especially when iterative solvers are being used)
*** so it should be used with care. Warning! see the option calc_matrix_info_null_space_only
*** as it affects the behavior of this option.
* calc_matrix_norms = true; *** (costly?)
* calc_matrix_norms = false; *** [default]
*** If true, then estimates of the matrix norms of all of the important
*** matrices used in the algorithm will be computed and printed. Note that this can be
*** a fairly expensive operation (especially if iterative solvers are being used) so
*** it should be used with care. Warning! see the option calc_matrix_info_null_space_only.
* calc_matrix_info_null_space_only = true; *** (costly?)
* calc_matrix_info_null_space_only = false; *** [default]
*** If true, then the options calc_conditioning and calc_matrix_norms will only
*** apply to quantities in the null space and not the quasi-range space
*** or the full space for which these options will be considered to be false.
}
************************************************************
*** Options for testing the NLP interface
***
*** [MoochoSolver::test_nlp == true]
***
options_group NLPTester {
* print_all = true;
* print_all = false; *** [default]
*** If true, then everything about the NLP will be printed
*** to journal_out (i.e. the file 'MoochoJournal.out').
*** This is useful for initial debugging but not recommended
*** for larger problems.
}
*************************************************************
*** Options for testing the vector spaces from the NLP object
***
*** [MoochoSolver::test_nlp == true]
***
options_group VectorSpaceTester {
* print_all_tests = true;
* print_all_tests = false;
* print_vectors = true;
* print_vectors = false;
* throw_exception = true;
* throw_exception = false;
* num_random_tests = 4; *** [default]
* warning_tol = 1e-14; *** [default]
* error_tol = 1e-10; *** [default]
}
***********************************************************
*** Options for the finite derivative testing for a
*** standard NLP.
***
*** See options_group NLPFirstDerivTester in
*** Moocho.opt.NLPAlgoConfigMamaJama
***
****************************************************************
*** Options for the finite difference derivative tester for a
*** direct sensitivity NLP.
***
*** See options_group NLPDirectTester in
*** Moocho.opt.NLPAlgoConfigMamaJama
***
****************************************************************
*** Options for the BasisSystem tester used to validate the
*** basis of the constraints Jacobian.
***
*** See options_group BasisSystemTester in
*** Moocho.opt.DecompositionSystemStateStepBuilderStd
***
****************************************************************
*** Options for the default BasisSystem factory object for
*** the constraints of the Jacobian used by
*** NLPSerialPreprocessExplJac
***
options_group BasisSystemFactoryStd {
* direct_linear_solver = DENSE; *** Use LAPACK xGETRF()
* direct_linear_solver = MA28; *** Use Harwell MA28 (see options_group DirectSparseSolverMA28)
* direct_linear_solver = MA48; *** Not supported yet
* direct_linear_solver = SUPERLU; *** Use SuperLU (see options_group DirectSparseSolverSuperLU)
*** Direct fortran-compatible linear solver for the basis of the Jacobian.
*** When a general NLP is being solved this selects the sparse linear solver used.
*** If the user specializes the BasisSystem object this option might be meaningless.
}
*****************************************************************
*** Set options for the MA28 solver.
***
*** [BasisSystemFactoryStd::direct_linear_solver == MA28]
***
options_group DirectSparseSolverMA28 {
* estimated_fillin_ratio = 10.0; *** [default]
*** (+dbl) Estimated amount of fillin ( > 1.0 ) for the
*** the sparse LU factorization. If this is too little
*** then more storage will automatically be allocated
*** on the fly (at the cost some wasted computations).
*** This parameter is mostly problem dependent and can
*** be adjusted to a proper size to reduce memory requirements.
*** Example values:
*** 1.0 : No fill-in?
*** 10.0 : LU factors have three times the number of nonzeros
* u = 0.1; *** [default]
*** (+dbl) Control parameter (0 <= u <= 1.0) that us used
*** to balance sparsity and accuracy.
*** Example values:
*** 0.0 : Pivot for sparsity only
*** 0.1 : Balance sparsity and stability
*** 1.0 : Pivot for stability only
* grow = true;
* grow = false; *** [default]
*** See MA28 documentation.
* nsrch = 4; *** [default]
*** (+int) Number of columns that MA28 will search to find
*** pivots to try to reduce fill-in. Warning, setting a large
*** value for 'nsrch' can greatly slow down the initial
*** rSQP iteration.
*** Example values:
*** 0 : No columns are searched
*** 4 : Four columns are searched
*** 1000 : A thousand columns are searched
* lbig = true;
* lbig = false; *** [default]
*** See MA28 documentation.
* print_ma28_outputs = true;
* print_ma28_outputs = false; *** [default]
*** If true, then the values of the MA28 output will
*** be dumped to the journal output stream
*** (if journal_output_level >= PRINT_ALGORITHM_STEPS).
* output_file_name = NONE; *** [default]
*** Gives the file name that MA28 Fortran output is set to (from LP and MP).
*** Example values:
*** NONE : No output file
*** any_other_name : Output from MA28 will be sent to this file in the
*** current directory
}
*** End Moocho.opt.MoochoSolver
*** Moocho.opt.DecompositionSystemStateStepBuilderStd
**********************************************************************************
*** All of these options can be used with any NLPAlgoConfig
*** subclass that uses the standard DecompositionSystemStateStepBuilderStd
*** class.
***
*** This file will be maintained and will include every option that
*** users can set. Most of these options the user will want to leave
*** alone but they are there in any case.
***
***********************************************************
*** Options for IterationPack::Algorithm
***
*** These are options that are used by every IterationPack::Algorithm
*** object that gets created and used.
***
options_group IterationPack_Algorithm {
* interrupt_file_name = ""; *** Does not check for interrupt file [default]
* interrupt_file_name = "interrupt.in"; *** checks for this file in current directory
*** This specifies a file name that is looked for at the end of every
*** step in a running algorithm. If this file exists, it is read for
*** algorithm termination criteria (see the class IterationPack::Algorithm
*** and the option interrupt_file_name). Using a file to interrupt an running
*** algorithm allows the algorithm to be gracefully terminated when run in batch
*** mode or when access to STDOUT or STDIN is not possible.
}
*****************************************************************
*** Options specific for the shared rSQP algorithm builder
*** class DecompositionSystemStateStepBuilderStd.
***
options_group DecompositionSystemStateStepBuilderStd {
*** Variable Reduction range/null space decomposition
* null_space_matrix = AUTO; *** Let the solver decide [default]
* null_space_matrix = EXPLICIT; *** Compute and store D = -inv(C)*N explicitly
* null_space_matrix = IMPLICIT; *** Perform operations implicitly with C, N (requires adjoints)
*** This option is used to determine the type of implementation to use for
*** the variable reduction null space matrix Z = [ -inv(C)*N; I ].
*** AUTO : Let the algorithm decide. The algorithm will take into
*** account the number of degrees of freedom in the problem
*** (n-r), the number of active inequality constraints and
*** other issues when deciding what implementation to use within
*** each iteration. Warning! These automatic tests are only
*** affective when a direct solver for C is used.
*** EXPLICIT : The matrix D = -inv(C)*N is computed and formed explicitly and is used
*** to form Z = [ D; I ].
*** IMPLICIT : The matrix D = -inv(C)*N is not formed explicitly but is
*** instead used implicitly in matrix-vector and other
*** related operations.
* range_space_matrix = AUTO; *** Let the algorithm decide dynamically [default]
* range_space_matrix = COORDINATE; *** Y = [ I; 0 ] (Cheaper computationally)
* range_space_matrix = ORTHOGONAL; *** Y = [ I; -N'*inv(C') ] (more stable)
*** This option is used to determine the selection of the range space matrix Y.
*** AUTO : Let the algorithm decide. The algorithm will take into
*** account the number of degrees of freedom in the problem
*** (n-r) and other issues when deciding what representation to use.
*** Warning! These automatic tests are only affective when a direct
*** solver for C is used.
*** COORDINATE : Use the coordinate decomposition Y = [ I; 0 ]
*** ORTHOGONAL : Use the orthogonal decomposition Y = [ I; N'*inv(C') ].
*** Warning! For general NLPs this option costs approximately
*** O((n-r)^2*r) dense flops per rSQP iteration and will dominate
*** the runtime for large (n-r). In addition, this option
*** assumes that you will be using null_space_matrix=EXPLICIT.
*** Reduced Hessian Approximations
* max_dof_quasi_newton_dense = -1; *** [default]
*** [quasi_newton == AUTO] (+-int) This option is used to
*** determine when the algorithm will switch from quasi_newton=BFGS
*** to quasi_newton=LBFGS and from range_space_matrix=ORTHOGONAL
*** to range_space_matrix=COORDINATE.
*** Example values:
*** -1 : (< 0) Let the solver decide dynamically [default]
*** 0 : Always use limited memory LBFGS and COORDINATE.
*** 500 : Use LBFGS when n-r >= 500 and dense BFGS when n-m < 500
*** Use COORDINATE when (n-m)*r >= 500 and ORTHOGONAL when
*** (n-r)*r <= 500.
*** 1e10: Always use the dense BFGS and the orthogonal decomposition.
}
***************************************************************
*** Options group for CalcFiniteDiffProd class
***
*** These options control how finite differences are computed
*** for testing and for other purposes.
***
options_group CalcFiniteDiffProd {
* fd_method_order = FD_ORDER_ONE; *** Use O(eps) one sided finite differences
* fd_method_order = FD_ORDER_TWO; *** Use O(eps^2) one sided finite differences
* fd_method_order = FD_ORDER_TWO_CENTRAL; *** Use O(eps^2) two sided central finite differences
* fd_method_order = FD_ORDER_TWO_AUTO; *** Uses FD_ORDER_TWO_CENTRAL or FD_ORDER_TWO
* fd_method_order = FD_ORDER_FOUR; *** Use O(eps^4) one sided finite differences
* fd_method_order = FD_ORDER_FOUR_CENTRAL; *** Use O(eps^4) two sided central finite differences
* fd_method_order = FD_ORDER_FOUR_AUTO; *** [default] Use FD_ORDER_FOUR_CENTRAL or FD_ORDER_FOUR
*** Selects the finite differencing method to use. Several different
*** methods of different orders are available. For more accuracy use a higher order
*** method, for faster execution time use a lower order method.
* fd_step_select = FD_STEP_ABSOLUTE; *** [default] Use absolute step size fd_step_size
* fd_step_select = FD_STEP_RELATIVE; *** Use relative step size fd_step_size * ||x||inf
*** Determines how the actual finite difference step size that is used is selected.
*** FD_STEP_ABSOLUTE : The actual step size used is taken from fd_step_size or
*** is determined by the implementation if fd_step_size < 0.0.
*** Taking an absolute step size can result in inaccurate gradients
*** for badly scaled NLPs.
*** FD_STEP_RELATIVE : The actual step size used is taken from fd_step_size or
*** is determined by the implementation if fd_step_size < 0.0
*** and then multiplied by ||x||inf. Taking a relative step
*** will not always result in accurate gradients and the user
*** may have to play with fd_step_size some.
* fd_step_size = -1.0; *** [default] Let the implementation decide
*** Determines what finite difference step size to use. If fd_step_select=FD_STEP_ABSOLUTE
*** then this is the absolute step size that is used. If fd_step_select=FD_STEP_RELATIVE
*** the actual step size used in fd_step_size * ||x||inf.
*** Some common values are:
*** < 0.0 : let the implementation decide.
*** 1e-8 : Optimal step size for FD_ORDER_ONE for IEEE double and perfect scaling?
*** 1e-5 : Optimal step size for FD_ORDER_TWOx for IEEE double and perfect scaling?
*** 1e-3 : Optimal step size for FD_ORDER_FOURx for IEEE double and perfect scaling?
* fd_step_size_min = -1.0; *** [default] Let the implementation decide.
*** Determines the minimum step size that will be taken to compute the finite differences.
*** This option is used to forbid the computation of a finite difference with a very small
*** step size as required by the variable bounds. Computing finite difference derivatives
*** for such small step sizes generally result in a lot of roundoff error. If
*** fd_step_size_min < 0.0, then the implementation will pick a default value that is
*** smaller than the default value for fd_step_size.
* fd_step_size_f = -1.0; *** [default] Let the implementation decide
*** Determines what finite difference step size to use for the objective function
*** f(x). If fd_step_size_f < 0.0, then the selected value for fd_step_size will be
*** used (see the options fd_step_size and fd_step_select). This option allows
*** fine-tuning of the finite difference computations.
* fd_step_size_c = -1.0; *** [default] Let the implementation decide
*** Determines what finite difference step size to use for the equality constraints
*** c(x). If fd_step_size_c < 0.0, then the selected value for fd_step_size will be
*** used (see the options fd_step_size and fd_step_select). This option allows
*** fine-tuning of the finite difference computations.
* fd_step_size_h = -1.0; *** [default] Let the implementation decide
*** Determines what finite difference step size to use for the inequality constraints
*** h(x). If fd_step_size_h < 0.0, then the selected value for fd_step_size will be
*** used (see the options fd_step_size and fd_step_select). This option allows
*** fine-tuning of the finite difference computations.
}
***************************************************************
*** Options for EvalNewPoint for NLPFirstOrder.
*** See options_group NLPFirstDerivTester
***
options_group EvalNewPointStd {
* fd_deriv_testing = FD_DEFAULT; *** [default] Test if check_results==true (see above)
* fd_deriv_testing = FD_TEST; *** Always test
* fd_deriv_testing = FD_NO_TEST; *** never test
*** Determines if the derivatives of the NLP returned from the NLPFirstOrder interface
*** are correct using finite differences (see the options_group NLPFirstDerivTester).
*** Valid options include:
*** FD_DEFAULT : Perform the finite difference tests if check_results==true.
*** FD_TEST : Always test, regardless of the value of check_results.
*** FD_NO_TEST : Never test, regardless of the value of check_results.
* decomp_sys_testing = DST_DEFAULT; *** [default] Test if check_results==true (see above)
* decomp_sys_testing = DST_TEST; *** Always test
* decomp_sys_testing = DST_NO_TEST; *** never test
*** Determines if the range/null decomposition matrices from DecompositionSystem are
*** tested or not (see the options_group DecompositionSystemTester).
*** Valid options include:
*** FD_DEFAULT : Perform the tests if check_results==true.
*** FD_TEST : Always test, regardless of the value of check_results.
*** FD_NO_TEST : Never test, regardless of the value of check_results.
* decomp_sys_testing_print_level = DSPL_USE_GLOBAL; *** [default] Use the value in journal_print_level (see above).
* decomp_sys_testing_print_level = DSPL_LEAVE_DEFAULT; *** Leave whatever setting in already in use.
*** This option allows the user to determine how the testing print level is determined.
*** Valid options include:
*** DSLP_USE_GLOBAL : The value of journal_print_level will be used to determine reasonable values for
*** print_tests and dump_all.
*** DSLP_LEAVE_DEFAULT : Whatever values are currently set for DecompositionSystemTester::print_tests and
*** DecompositionSystemTester::dump_all will be used (see the options_group
*** DecompositionSystemTester).
}
***************************************************************
*** Options for determining if variable bounds
*** xL <= x <= xU are violated by
*** more than an acceptable tolerance.
***
options_group VariableBoundsTester {
* warning_tol = 1e-10; *** [default]
* error_tol = 1e-5; *** [default]
}
***********************************************************
*** Options for the finite difference testing of derivatives for a
*** standard NLP.
***
options_group NLPFirstDerivTester {
* fd_testing_method = FD_COMPUTE_ALL; *** Compute all of the derivatives (O(m))
* fd_testing_method = FD_DIRECTIONAL; *** [default] Only compute along random directions (O(1))
* num_fd_directions = 1; *** [fd_testing_method == DIRECTIONAL]
* num_fd_directions = -1; *** [fd_testing_method == DIRECTIONAL] Use single direction y=1.0
* warning_tol = 1e-8; *** [default]
* warning_tol = 0.0; *** Show me all comparisons.
* error_tol = 1e-3; *** [default]
}
***************************************************************
*** Options for EvalNewPoint for a "Tailored Approach" NLP.
*** See options_group NLPDirectTester
***
options_group EvalNewPointTailoredApproach {
* fd_deriv_testing = FD_DEFAULT; *** [default] Test if check_results==true (see above)
* fd_deriv_testing = FD_TEST; *** Always test
* fd_deriv_testing = FD_NO_TEST; *** never test
}
****************************************************************
*** Options for the finite difference derivative tester for a
*** direct sensitivity NLP.
***
options_group NLPDirectTester {
* Gf_testing_method = FD_COMPUTE_ALL; *** Compute all of the derivatives (O(n))
* Gf_testing_method = FD_DIRECTIONAL; *** [default] Only compute along random directions (O(1))
* Gf_warning_tol = 1e-10;
* Gf_error_tol = 1e-5;
* Gc_testing_method = FD_COMPUTE_ALL; *** Compute all of the derivatives (O(n-m))
* Gc_testing_method = FD_DIRECTIONAL; *** [default] Only compute along random directions (O(1))
* Gc_warning_tol = 1e-10;
* Gc_error_tol = 1e-5;
* num_fd_directions = 1; *** [testing_method == DIRECTIONAL]
* dump_all = true;
* dump_all = false; *** [default]
}
****************************************************************
*** Options for the BasisSystem tester used to validate the
*** basis of the constraints Jacobian.
***
options_group BasisSystemTester {
* print_tests = PRINT_NONE; *** [default]
* print_tests = PRINT_BASIC;
* print_tests = PRINT_MORE;
* print_tests = PRINT_ALL;
* dump_all = true;
* dump_all = false; *** [default]
* num_random_tests = 1; *** (+int) Number of sets of random tests to perform
* warning_tol = 1e-15; *** (+dbl) Warning tolerance
* error_tol = 1e-12; *** (+dbl) Error tolerance
}
****************************************************************
*** Options for the DecompositionSystem tester used to validate
*** range/null decomposition matrices (NLPFirstOrder only).
***
options_group DecompositionSystemTester {
* print_tests = PRINT_NONE; *** [default]
* print_tests = PRINT_BASIC;
* print_tests = PRINT_MORE;
* print_tests = PRINT_ALL;
* dump_all = true; *** (costly)
* dump_all = false; *** [default]
* num_random_tests = 1; *** (+int) Number of sets of random test to perform
* mult_warning_tol = 1e-14; *** (+dbl) Warning tolerance for checking matrix-vector multiplication
* mult_error_tol = 1e-8; *** (+dbl) Error tolerance for checking matrix-vector multiplication
* solve_warning_tol = 1e-14; *** (+dbl) Warning tolerance for checking linear solves
* solve_error_tol = 1e-8; *** (+dbl) Error tolerance for checking linear solves
}
*** End Moocho.opt.DecompositionSystemStateStepBuilderStd
*** Begin Moocho.opt.NLPAlgoConfigMamaJama
*************************************************************************
*** All of these options can be used with the NLPAlgoConfigMamaJama
*** algorithm configuration class.
***
*** See the file Moocho.opt.DecompositionSystemStateStepBuilderStd
*** for more options that are used by this class.
***
*** This file will be maintained and will include every option that
*** users can set. Most of these options the user will want to leave
*** alone but they are there in any case.
***
**********************************************************
*** Options specific for the rSQP algorithm configuration
*** class NLPAlgoConfigMamaJama.
***
options_group NLPAlgoConfigMamaJama {
*** Variable Reduction range/null space decomposition
* max_basis_cond_change_frac = -1.0; *** [default]
*** (+-dbl) If < 0 then the solver will decide what value to use.
*** Otherwise this is the change in a very inexact condition number estimate
*** between iterations (see printed algorithm description) which triggers the
*** selection of a new basis.
*** Example values:
*** -1 : Allow solver to decide [default]
*** 0 : Switch to a new basis every iteration (not a good idea)
*** 100 : Switch to a new basis when change is more that 100?
*** 1e+50 : (big number) Never switch to a new basis.
*** Reduced Hessian Approximations
* exact_reduced_hessian = true; *** Use NLP Hessian info if available
* exact_reduced_hessian = false; *** Use quasi_newton [default]
*** If true and if the NLP supports second order information
*** (hessian of the lagrangian HL) then the exact reduced hessian
*** rHL = Z'*HL*Z will be computed at each iteration.
* quasi_newton = AUTO; *** Let solver decide dynamically [default]
* quasi_newton = BFGS; *** Dense BFGS
* quasi_newton = LBFGS; *** Limited memory BFGS
*** [exact_reduced_hessian == false]
*** ToDo: Finish documentation!
* num_lbfgs_updates_stored = -1; *** [default]
*** [quasi_newton == LBFGS] (+-int) If < 0 then let solver decide
*** otherwise this is the maximum number of update vectors stored
*** for limited memory LBFGS.
* lbfgs_auto_scaling = true; *** (default)
* lbfgs_auto_scaling = false;
*** [quasi_newton == LBFGS] If true then auto scaling of initial
*** hessian approximation will be use for LBFGS.
* hessian_initialization = AUTO; *** Let the solver decide dynamically [default]
* hessian_initialization = SERIALIZE; *** rHL_(0) read from file (see ReducedHessianSerialization)
* hessian_initialization = IDENTITY; *** rHL_(0) = I
* hessian_initialization = FINITE_DIFF_SCALE_IDENTITY; *** rHL_(0) = ||fd|| * I
* hessian_initialization = FINITE_DIFF_DIAGONAL; *** rHL_(0) = diag(max(fd(i),small),i)
* hessian_initialization = FINITE_DIFF_DIAGONAL_ABS; *** rHL_(0) = diag(abs(fd(i))
*** [exact_reduced_hessian == false] Determines how the quasi-newton hessian is initialized.
*** ToDo: Finis documentation!
*** QP solvers
* qp_solver = AUTO; *** Let the solver decide dynamically
* qp_solver = QPKWIK; *** Primal-dual, active set, QR
* qp_solver = QPOPT; *** Primal, active set, null space, Gill et. al.
* qp_solver = QPSOL; *** Primal, active set, null space, Gill et. al.
* qp_solver = QPSCHUR; *** [default] Primal-dual, active set, schur complement
*** QP solver to use to solve the reduced space QP subproblem (null
*** space step). Note that only QPSCHUR ships with MOOCHO by default.
* reinit_hessian_on_qp_fail = true; *** [default]
* reinit_hessian_on_qp_fail = false;
*** If true, then if a QPFailure exception is thrown (see printed algorithm)
*** then the Hessian approximation will be reinitialized and the QP solver will
*** attempt to solve the QP again.
*** Line search methods
* line_search_method = AUTO; *** Let the solver decide dynamically [default]
* line_search_method = NONE; *** Take full steps at every iteration
* line_search_method = DIRECT; *** Use standard Armijo backtracking
* line_search_method = 2ND_ORDER_CORRECT; *** Like DIRECT except computes corrections for
* *** c(x) before backtracking line search
* line_search_method = WATCHDOG; *** Like DIRECT except uses watchdog type trial steps
* line_search_method = FILTER; *** [default] Use the Filter line search method
*** Options:
*** AUTO : Let the solver decide dynamically what line search method to use (if any)
*** NONE : Take full steps at every iteration. For most problems this is a bad idea.
*** However, for some problems this can help when starting close to the solution usually.
*** DIRECT : Use a standard Armijo backtracking line search at every iteration.
*** 2ND_ORDER_CORRECT : Like DIRECT except computes corrections for before applying
*** the backtracking line search (see options_group LineSearch2ndOrderCorrect).
*** This can help greatly on some problems and can counter act the Maritos effect.
*** FILTER : Use the filter line search. Here we accept either a decrease in the
*** objective function for the constraints. See "Global and Local Convergence of
*** Line Search Filter Methods for Nonlinear Programming" by Waechter and Biegler.
* merit_function_type = AUTO; *** [line_search_method != NONE] Let solver decide
* merit_function_type = L1; *** [line_search_method != NONE] phi(x) = f(x) + mu*||c(x)||1
* merit_function_type = MODIFIED_L1; *** [line_search_method != NONE] phi(x) = f(x) + sum(mu(j),|cj(x)|,j)
* merit_function_type = MODIFIED_L1_INCR; *** [line_search_method != NONE] Like MODIFIED_L1 except mu(j) are altered in order to take larger steps
*** Determines the type of merit function used when the line search
*** method uses a merit function.
* l1_penalty_parameter_update = AUTO; *** [merit_function_type == L1] let solver decide
* l1_penalty_parameter_update = WITH_MULT; *** [merit_function_type == L1] Use Lagrange multipliers to update mu
* l1_penalty_parameter_update = MULT_FREE; *** [merit_function_type == L1] Don't use Lagrange multipliers to update mu
*** Determines how the penalty parameter is updated for the L1 merit function.
}
*********************************************************************
*** Options for serialization of the reduced Hessian
***
*** [NLPAlgoConfigMamaJama::hessian_initialization == SERIALIZE]
***
options_group ReducedHessianSerialization {
* reduced_hessian_input_file_name = "reduced_hessian.in"; *** [default]
* reduced_hessian_input_file_name = ""; *** Does not read from file
*** The name of a file that will be used to read in the reduced Hessian
*** in a format that is compatible with the internal implementation.
* reduced_hessian_output_file_name = "reduced_hessian.out"; *** [default]
* reduced_hessian_output_file_name = ""; *** Does not write to file
*** The name of a file that will be used to write in the reduced Hessian.
*** This reduced Hessian can then be read back in using the
*** reduced_hessian_input_file_name option.
}
*********************************************************************
*** Options for finite difference initialization of reduced hessian.
***
*** [NLPAlgoConfigMamaJama::hessian_initialization == FINITE_DIFF_*]
***
options_group InitFinDiffReducedHessian {
* initialization_method = SCALE_IDENTITY;
* initialization_method = SCALE_DIAGONAL;
* initialization_method = SCALE_DIAGONAL_ABS;
* max_cond = 1e+1;
* min_diag = 1e-8;
* step_scale = 1e-1;
}
*********************************************************************
*** Options for checking for skipping the BFGS update.
***
*** [NLPAlgoConfigMamaJama::exact_hessian == false]
***
options_group CheckSkipBFGSUpdateStd {
* skip_bfgs_prop_const = 10.0; *** (+dbl)
}
*********************************************************************
*** Options for BFGS updating (dense or limited memory)
***
*** [NLPAlgoConfigMamaJama::exact_hessian == false]
***
options_group BFGSUpdate {
* rescale_init_identity = true; *** [default]
* rescale_init_identity = false;
*** If true, then rescale the initial identity matrix at 2nd iteration
* use_dampening = true; *** [default]
* use_dampening = false;
*** Use dampened BFGS update
* secant_testing = DEFAULT; *** Test secant condition if check_results==true (see above) [default]
* secant_testing = TEST; *** Always test secant condition
* secant_testing = NO_TEST; *** Never test secant condition
* secant_warning_tol = 1e-6; *** [default]
* secant_error_tol = 1e-1; *** [default]
}
*********************************************************************
*** Options for the convergence test.
***
*** See the printed step description (i.e. 'MoochoAlgo.out') for a
*** description of what these options do.
***
options_group CheckConvergenceStd {
* scale_opt_error_by = SCALE_BY_NORM_2_X;
* scale_opt_error_by = SCALE_BY_NORM_INF_X;
* scale_opt_error_by = SCALE_BY_ONE; *** [default]
* scale_feas_error_by = SCALE_BY_NORM_2_X;
* scale_feas_error_by = SCALE_BY_NORM_INF_X;
* scale_feas_error_by = SCALE_BY_ONE; *** [default]
* scale_comp_error_by = SCALE_BY_NORM_2_X;
* scale_comp_error_by = SCALE_BY_NORM_INF_X;
* scale_comp_error_by = SCALE_BY_ONE; *** [default]
*** Determines what all of the error measures are scaled by when checking convergence
*** SCALE_BY_NORM_2_X : Scale the optimality conditions by 1/(1+||x_k||2)
*** SCALE_BY_NORM_INF_X : Scale the optimality conditions by 1/(1+||x_k||inf)
*** SCALE_BY_ONE : Scale the optimality conditions by 1
* scale_opt_error_by_Gf = true; *** [default]
* scale_opt_error_by_Gf = false;
*** Determines if the linear dependence of gradients (i.e. ||rGL_k||inf or ||GL_k||inf)
*** is scaled by the gradient of the objective function or not.
*** true : Scale ||rGL_k||inf or ||GL_k|| by an additional 1/(1+||Gf||inf)
*** false : Scale ||rGL_k||inf or ||GL_k|| by an additional 1 (i.e. no extra scaling)
}
****************************************************************
*** Options for the TangentalStepWithInequStd_Step
***
*** This used for NLPs that have bounds.
***
options_group TangentialStepWithInequStd {
* warm_start_frac = 0.8; *** [default]
* warm_start_frac = 0.0; *** Never do a warm start
* warm_start_frac = 1.0; *** Do a warm start as soon a possible
*** (+dbl) Determines the number of inequality constraints that
*** must be the same any two rSQP iterations in order for
*** a warm start to be used on the next rSQP iteration.
* qp_testing = QP_TEST_DEFAULT; *** [default] Test if check_results==true
* qp_testing = QP_TEST; *** Always test
* qp_testing = QP_NO_TEST; *** Never test
*** Determines if the postconditions for the QP solve are checked
*** or not.
* primal_feasible_point_error = true; *** [default] Throw exception on PRIMAL_FEASIBLE_POINT
* primal_feasible_point_error = false; *** No throw exception on PRIMAL_FEASIBLE_POINT
* dual_feasible_point_error = true; *** [default] Throw exception on DUAL_FEASIBLE_POINT
* dual_feasible_point_error = false; *** No throw exception on DUAL_FEASIBLE_POINT
}
********************************************************************
*** Options for the QPSolverRelaxedTester object that is used
*** to test the QP solution.
***
*** This is only used when the NLP has bounds and a QP solver is used.
*** This sets testing options for the TangentalStepWithInequalStd_Step
*** object.
***
*** See the MoochoAlgo.opt file for details.
***
options_group QPSolverRelaxedTester {
* opt_warning_tol = 1e-10; *** [default] Tolerances for optimality conditions
* opt_error_tol = 1e-5; *** [default]
* feas_warning_tol = 1e-10; *** [default] Tolerances for feasibility
* feas_error_tol = 1e-5; *** [default]
* comp_warning_tol = 1e-10; *** [default] Tolerances for complementarity
* comp_error_tol = 1e-5; *** [default]
}
****************************************************************
*** Options for the primal-dual, active-set, schur-complement
*** QP solver QPSchur.
***
*** [NLPAlgoConfigMamaJama::qp_solver == QPSCHUR]
***
options_group QPSolverRelaxedQPSchur {
*** Convergence criteria and algorithm control options
* max_qp_iter_frac = 10.0; *** (+dbl) max_qp_itr = max_qp_itr_frac * (# variables)
* bounds_tol = 1e-10; *** (+dbl) feasibility tolerance for bound constraints
* inequality_tol = 1e-10; *** (+dbl) feasibility tolerance for general inequality constraints
* equality_tol = 1e-10; *** (+dbl) feasibility tolerance for general equality constraints
* loose_feas_tol = 1e-9; *** (+dbl) (Expert use only)
* dual_infeas_tol = 1e-12; *** (+dbl) allowable dual infeasibility before reporting an error
* huge_primal_step = 1e+20; *** (+dbl) value of a near infinite primal step
* huge_dual_step = 1e+20; *** (+dbl) value of a near infinite dual step
* bigM = 1e+10; *** (+dbl) value or relaxation penalty in objective
* iter_refine_at_solution = true; *** [default]
* iter_refine_at_solution = false;
*** If true then iterative refinement will be performed at the solution of
*** the QP in every case.
* iter_refine_min_iter = 1; *** [default]
*** Minimum number of iterative refinement iterations to perform when
*** using iterative refinement.
*** Example values:
*** 0 : Don't perform any iterations of iterative refinement if you don't have to
*** 1 : Perform at least one step if iterative refinement no matter what the
*** residual is.
* iter_refine_max_iter = 3; *** [default]
*** Maximum number of steps of iterative refinement to perform.
*** This helps to keep down the cost of iterative refinement but
*** can cause the QP method to fail due to ill-conditioning and roundoff.
*** This number should be kept small since the target residuals may not
*** be obtainable.
*** Example values:
*** 0 : Never perform any steps of iterative refinement
*** 1 : Never perform more than one step of iterative refinement
*** 10 : Never perform more than 10 steps of iterative refinement.
* iter_refine_opt_tol = 1e-12; *** [default]
*** Iterative refinement convergence tolerance for the optimality
*** conditions of the QP. Specifically this is compared to the
*** scaled linear dependence of gradients condition.
*** Example values:
*** 1e-50 : (very small number) Do iterative refinement until
*** iter_refine_max_iter is exceeded.
*** 1e-12 : [default]
*** 1e+50 : (very big number) Allow convergence of iterative refinement at
*** any time.
* iter_refine_feas_tol = 1e-12; *** [default]
*** Iterative refinement convergence tolerance for the feasibility
*** conditions of the QP. Specifically this is compared to the
*** scaled residual of the active constraints (equalities and inequalities)
*** Example values:
*** 1e-50 : (very small number) Do iterative refinement until
*** iter_refine_max_iter is exceeded.
*** 1e-12 : [default]
*** 1e+50 : (very big number) Allow convergence of iterative refinement at
*** any time.
* inequality_pick_policy = ADD_BOUNDS_THEN_MOST_VIOLATED_INEQUALITY; *** [default]
* inequality_pick_policy = ADD_BOUNDS_THEN_FIRST_VIOLATED_INEQUALITY; *** not supported yet!
* inequality_pick_policy = ADD_MOST_VIOLATED_BOUNDS_AND_INEQUALITY;
*** Warning and error tolerances
* warning_tol = 1e-10; *** General testing warning tolerance
* error_tol = 1e-5; *** General testing error tolerance
* pivot_warning_tol = 1e-8; *** [default]
*** (+dbl) Minimum relative tolerance for a pivot element in the schur complement
*** under which a warning message for near singularity will be printed
*** (see MatrixSymAddDelUpdateable).
*** Example values:
*** 0.0: Don't print any warning messages about near singularity.
*** 1e-6: default
*** 2.0: ( > 1 ) Show the pivot tolerance of every update!
* pivot_singular_tol = 1e-11; *** [default]
*** (+dbl) Minimum relative tolerance for a pivot element in the schur complement
*** under which the matrix is considered singular and an error message will be printed
*** (see MatrixSymAddDelUpdateable).
*** Example values:
*** 0.0: Allow any numerically nonsingular matrix.
*** 1e-8: default
*** 2.0: ( > 1 ) Every matrix is singular (makes no sense!)
* pivot_wrong_inertia_tol = 1e-11; *** [default]
*** (+dbl) Minimum relative tolerance for a pivot element in the schur complement
*** over which the matrix is considered to have the wrong inertia rather than
*** being singular and an error message will be printed
*** (see MatrixSymAddDelUpdateable).
*** Example values:
*** 0.0: Any pivot with the wrong sign will be considered to have the wrong inertia
*** 1e-8: default
*** 2.0: ( > 1 ) Every matrix has the wrong inertia (makes no sense!)
*** Output control
* print_level = USE_INPUT_ARG; *** [default] Use the input argument to solve_qp(...)
* print_level = NO_OUTPUT;
* print_level = OUTPUT_BASIC_INFO;
* print_level = OUTPUT_ITER_SUMMARY;
* print_level = OUTPUT_ITER_STEPS;
* print_level = OUTPUT_ACT_SET;
* print_level = OUTPUT_ITER_QUANTITIES;
}
********************************************************************
*** Options for the direct line search object that is used in all the
*** line search methods for the SQP step.
***
*** [NLPAlgoConfigMamaJama::line_search_method != NONE]
***
*** See the MoochoAlgo.opt file for details.
***
options_group DirectLineSearchArmQuadSQPStep {
* slope_frac = 1.0e-4;
* min_frac_step = 0.1:
* max_frac_step = 0.5;
* max_ls_iter = 20;
}
*******************************************************************
*** Options for the watchdog line search.
***
*** [NLPAlgoConfigMamaJama::line_search_method = WATCHDOG]
***
*** Warning! The watchdog option is not currently supported!
***
options_group LineSearchWatchDog {
* opt_kkt_err_threshold = 1e-3; *** (+dbl)
* feas_kkt_err_threshold = 1e-3; *** (+dbl)
*** Start the watchdog linesearch when opt_kkt_err_k < opt_kkt_err_threshold and
*** feas_kkt_err_k < feas_kkt_err_threshold
}
*******************************************************************
*** Options for the second order correction line search.
***
*** [NLPAlgoConfigMamaJama::line_search_method == 2ND_ORDER_CORRECT]
***
*** Warning! The 2nd order correction option is not currently supported!
***
options_group LineSearch2ndOrderCorrect {
* newton_olevel = PRINT_USE_DEFAULT; *** O(?) output [default]
* newton_olevel = PRINT_NOTHING; *** No output
* newton_olevel = PRINT_SUMMARY_INFO; *** O(max_newton_iter) output
* newton_olevel = PRINT_STEPS; *** O(max_newton_iter) output
* newton_olevel = PRINT_VECTORS; *** O(max_newton_iter*n) output
*** Determines the amount of output printed to the journal output stream.
*** PRINT_USE_DEFAULT: Use the output level from the overall algorithm
*** to print comparable output (see journal_output_level).
*** PRINT_NOTHING: Don't print anything (overrides default print level).
*** PRINT_SUMMARY_INFO: Print a nice little summary table showing the sizes of the
*** newton steps used to compute a feasibility correction as well as what
*** progress is being made.
*** PRINT_STEPS: Don't print a summary table and instead print some more detail
*** as to what computations are being performed etc.
*** PRINT_VECTORS: Print out relevant vectors as well for the feasibility Newton
*** iterations.
* constr_norm_threshold = 1.0; *** [default]
*** (+dbl) Tolerance for ||c_k||inf below which a 2nd order correction step
*** will be considered (see printed description). Example values:
*** 0.0: Never consider computing a correction.
*** 1e-3: Consider a correction if and only if ||c_k||inf <= 1e-3
*** 1e+50: (big number) Consider a correction regardless of ||c_k||inf.
* constr_incr_ratio = 10.0; *** [default]
*** (+dbl) Tolerance for ||c_kp1||inf/(1.0+||c_k||inf) below which a 2nd order
*** correction step will be considered (see printed description). Example values:
*** 0.0: Consider computing a correction only if ||c_kp1||inf is zero.
*** 10.0: Consider computing a correction if and only if
*** ||c_kp1||inf/(1.0+||c_k||inf) < 10.0.
*** 1e+50: (big number) Consider a correction regardless how big ||c_kp1||inf is.
* after_k_iter = 0; *** [default]
*** (+int) Number of SQP iterations before a 2nd order correction will be considered
*** (see printed description). Example values:
*** 0: Consider computing a correction right away at the first iteration.
*** 2: Consider computing a correction when k >= 2.
*** 999999: (big number) Never consider a 2nd order correction.
* forced_constr_reduction = LESS_X_D;
* forced_constr_reduction = LESS_X; *** [default]
*** Determine the amount of reduction required for c(x_k+d+w).
*** LESS_X_D: phi(c(x_k+d+w)) < forced_reduct_ratio * phi(c(x_k+d)) is all that is required.
*** As long as a feasible step can be computed, only one newton
*** iteration should be required for this.
*** LESS_X: phi(c(x_k+d+w)) < forced_reduct_ratio * phi(c(x_k)) is required.
*** In general, this may require several feasibility step calculations
*** and several newton iterations. Of course the maximum number of
*** newton iterations may be exceeded before this is achieved.
* forced_reduct_ratio = 1.0; *** [default]
*** (+dbl) (< 1) Fraction of reduction in phi(c(x)) for required reduction.
*** Example values:
*** 0.0: The constraints must be fully converged and newton iterations will
*** performed until max_newton_itr is exceeded.
*** 0.5: Require an extra 50% of the required reduction.
*** 1.0: Don't require any extra reduction.
* max_step_ratio = 1.0; *** [default]
*** (+dbl) Maximum ratio of ||w^p||inf/||d||inf allowed for correction step w^p before
*** a line search along phi(c(x_k+d+b*w^p)) is performed. The purpose of this parameter
*** is to limit the number of line search iterations needed for each feasibility
*** step and to keep the full w = sum(w^p,p=1...) from getting too big. Example values:
*** 0.0: Don't allow any correction step.
*** 0.1: Allow ||w^p||inf/||d||inf <= 0.1.
*** 1.0: Allow ||w^p||inf/||d||inf <= 1.0.
*** 1e+50: (big number) Allow ||w^p||inf/||d||inf to be a big as possible.
* max_newton_iter = 3; *** [default]
*** (+int) Limit the number of newton feasibility iterations (with line searches)
*** allowed. Example values:
*** 0: Don't allow any newton iterations (no 2nd order correction).
*** 3: Allow 3 newton iterations
*** 999999: Allow any number of newton iterations (not a good idea)
}
*******************************************************************
*** Options for the filter search.
***
*** [NLPAlgoConfigMamaJama::line_search_method = FILTER]
***
*** See the MoochoAlgo.out file for details
***
options_group LineSearchFilter {
* gamma_theta = 1e-5; *** [default]
* gamma_f = 1e-5; *** [default]
* f_min = UNBOUNDED; *** [default]
* f_min = 0.0; *** If 0 is minimum ...
* gamma_alpha = 5e-2; *** [default]
* delta = 1e-4; *** [default]
* s_theta = 1.1; *** [default]
* s_f = 2.3; *** [default]
* theta_small_fact = 1e-4; *** [default]
* theta_max = 1e10; *** [default]
* eta_f = 1e-4; *** [default]
* back_track_frac = 0.5; *** [default]
}
*******************************************************************
*** Options for generating feasibility steps for reduced space.
***
*** [NLPAlgoConfigMamaJama::line_search_method == 2ND_ORDER_CORRECT]
***
*** Warning! The 2nd order correction option is not currently supported!
***
options_group FeasibilityStepReducedStd {
* qp_objective = OBJ_MIN_FULL_STEP;
* qp_objective = OBJ_MIN_NULL_SPACE_STEP;
* qp_objective = OBJ_RSQP;
* qp_testing = QP_TEST_DEFAULT;
* qp_testing = QP_TEST;
* qp_testing = QP_NO_TEST;
}
******************************************************************
*** Options for the direct line search object for
*** the newton steps of the 2nd order correction (see above).
***
*** [NLPAlgoConfigMamaJama::line_search_method == 2ND_ORDER_CORRECT]
***
*** Warning! The 2nd order correction option is not currently supported!
***
options_group DirectLineSearchArmQuad2ndOrderCorrectNewton {
* slope_frac = 1.0e-4;
* min_frac_step = 0.1:
* max_frac_step = 0.5;
* max_ls_iter = 20;
}
******************************************************************
*** Change how the penalty parameters for the merit function
*** are adjusted.
***
*** [NLPAlgoConfigMamaJama::line_search_method != NONE]
***
options_group MeritFuncPenaltyParamUpdate {
* small_mu = 1e-6;
* min_mu_ratio = 1e-8
* mult_factor = 7.5e-4;
* kkt_near_sol = 1e-1;
}
*****************************************************************
*** Change how the penalty parameters for the modified
*** L1 merit function are increased.
***
*** [NLPAlgoConfigMamaJama::line_search_method != NONE]
*** [NLPAlgoConfigMamaJama::merit_function_type == MODIFIED_L1_INCR]
***
options_group MeritFuncModifiedL1LargerSteps {
* after_k_iter = 3;
*** (+int) Number of SQP iterations before considering increasing penalties.
*** Set to 0 to start at the first iteration.
* obj_increase_threshold = 1e-4;
*** (+-dbl) Consider increasing penalty parameters when the relative
*** increase in the objective function is greater than this value.
*** Set to a very large negative number (i.e. -1e+100) to always
*** allow increasing the penalty parameters.
* max_pos_penalty_increase = 1.0;
*** (+dbl) Ratio the multipliers are allowed to be increased.
*** Set to a very large number (1e+100) to allow increasing the penalties
*** to any value if it will help in taking a larger step. This will
*** in effect put all of the weight of the constraints and will force
*** the algorithm to only minimize the infeasibility and ignore
*** optimality.
* pos_to_neg_penalty_increase = 1.0; *** (+dbl)
* incr_mult_factor = 1e-4; *** (+dbl)
}
*** End Moocho.opt.NLPAlgoConfigMamaJama
end_options
1.8.6