|
MoochoPack : Framework for Large-Scale Optimization Algorithms
Version of the Day
|
|
MoochoPack : Framework for Large-Scale Optimization Algorithms
Version of the Day
|
The following is the contents of the file Moocho.opt.NLPAlgoConfigIP which are options specific to the class MoochoPack::NLPAlgoConfigIP and the class objects that it configures.
*** Begin Moocho.opt.NLPAlgoConfigIP
********************************************************************
*** All of these options can be used with the NLPAlgoConfigIP
*** algorithm configuration 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 IP algorithm configuration
*** class NLPAlgoConfigIP
***
options_group NLPAlgoConfigIP {
*** 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 (see printed algorithm description) above
*** which a new basis will be selected. Example values:
*** -1 : Allow solver to decide [default]
*** 0 : Switch to a new basis every iteration (not a good idea)
*** 100 : Seitch 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!
* 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=LBFGS
*** to quasi_newton=BFGS 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 LBFG 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.
* num_lbfgs_updates_stored = -1; *** [default]
*** [quasi_newton == L*BFGS] (+-int) If < 0 then let solver decide
*** otherwise this is the maximum number of update vectors stored
*** for limite memory LBFGS.
* lbfgs_auto_scaling = true; *** (defalut)
* lbfgs_auto_scaling = false;
*** [quasi_newton == LBFGS] If true then auto scaling of initial
*** hessian approximation will be use for LBFGS.
*** 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 = FILTER; *** 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 constriants. See "Global and Local Convergence of
*** Line Search Filter Methods for Nonliner 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
* 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
}
***************************************************************
*** Options group for UpdateBarrierParameter_Step
***
*** These options control how the barrier parameter (mu)
*** is updated
***
options_group UpdateBarrierParameter {
* tau_mu = 0.2; *** linear decrease fraction for mu
* theta_mu = 1.5; *** superlinear decrease power for mu
* tau_epsilon = 10; *** error tolerance fraction
* theta_epsilon = 1.1; *** error tolerance power
* e_tol_max = 1000; *** maximum error tolerance
}
***************************************************************
*** Options group for CalcFiniteDiffProd class
***
*** These options control how finite differences are computed
*** when 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] Uses 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 errors. 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-tunning 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-tunning 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-tunning of the finite difference computations.
}
***************************************************************
*** Options for EvalNewPointBarrier.
*** See options_group NLPFirstDerivTester
***
options_group EvalNewPointBarrier {
* relative_bounds_push = 0.01; ***
* absolute_bounds_push = 0.001; ***
*** determines how to push x within bounds if initial values are specified
*** either outside the bounds or too close to the bounds
*** * xl_sb = min(xl+relative_bound_push*(xu-xl),
* xl + absolute_bound_push)
* xu_sb = max(xu-relative_bound_push*(xu-xl),
* xu - absolute_bound_push)
* if (xl_sb > xu_sb) then
* x = (xl + (xu-xl)/2
* else if (x < xl_sb) then
* x = xl_sb
* else if (x > xu_sb) then
* x = xu_sb
}
*********************************************************************
*** Options for the check for if to skip the BFGS update.
***
*** [NLPAlgoConfigMamaJama::exact_hessian == false]
***
options_group CheckSkipBFGSUpdateStd {
* skip_bfgs_prop_const = 10.0; *** (+dbl)
}
*********************************************************************
*** Options the 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)
* secant_testing = TEST; *** Always test secant condition
* secant_testing = NO_TEST; *** Never test secant condition
* secant_warning_tol = 1e-14;
* secant_error_tol = 1e-10;
}
*********************************************************************
*** Options the updating of the Reduced Sigma term
***
***
options_group UpdateReducedSigma {
* update_method = always_explicit;
* update_method = BFGS_primal;
* update_method = BFGS_dual_no_correction;
* update_method = BFGS_dual_explicit_correction; *** [default]
* update_method = BFGS_dual_scaling_correction;
*** These options determine exactly how the reduced sigma
*** term will be updated.
***
*** always_explicit : the full Z_kT*Sigma*Zk at each step (expensive)
*** BFGS_primal : a BFGS update of mu*X^-2 (exact at solution)
*** BFGS_dual_no_correction : update with Z_kT*Sigma*Z_k*pz
*** (no correction when mu changes)
*** BFGS_dual_explicit_correction : same as above
*** (do an explicit calculation when mu changes)
*** BFGS_dual_scaling_correction : same as above
*** (scale by mu_kp1/mu_k when mu changes)
}
*********************************************************************
*** 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 optimiality 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 ir ||GL_k|| by an additional 1/(1+||Gf||inf)
*** false : Scale ||rGL_k||inf ir ||GL_k|| by an addiational 1 (i.e. no extra scaling)
}
********************************************************************
*** 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]
***
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]
***
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]
***
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 algortihm
*** to print comperable output (see journal_output_level).
*** PRINT_NOTHING: Don't print anything (overriddes 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 relavant 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 correrction 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 corrrection requardless 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 maxinum 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) Maxumum 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 feasibilty 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 iteations (not a good idea)
}
*******************************************************************
*** Options for the filter search.
***
*** [NLPAlgoConfigMamaJama::line_search_method = FILTER]
***
options_group LineSearchFilter {
* gamma_theta = 1e-5;
* gamma_f = 1e-5;
* gamma_alpha = 5e-2;
* delta = 1e-4;
* s_theta = 1.1;
* s_f = 2.3;
* theta_small_fact = 1e-4;
* theta_max = 1e10;
* eta_f = 1e-4;
* back_track_frac = 0.5;
}
*******************************************************************
*** Options for generating feasibility steps for reduced space.
***
*** [NLPAlgoConfigMamaJama::line_search_method == 2ND_ORDER_CORRECT]
***
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]
***
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 modifed
*** 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 wieght of the constraints and will force
*** the algorithm to only minimize the infeasibilities and ignore
*** optimality.
* pos_to_neg_penalty_increase = 1.0; *** (+dbl)
* incr_mult_factor = 1e-4; *** (+dbl)
}
*** End Moocho.opt.NLPAlgoConfigIP
1.7.6.1