NAG CL Interface
e04ggc (handle_​solve_​bxnl)

Note: this function uses optional parameters to define choices in the problem specification and in the details of the algorithm. If you wish to use default settings for all of the optional parameters, you need only read Sections 1 to 10 of this document. If, however, you wish to reset some or all of the settings please refer to Section 11 for a detailed description of the algorithm and to Section 12 for a detailed description of the specification of the optional parameters.
Settings help

CL Name Style:


1 Purpose

e04ggc is a bound-constrained nonlinear least squares trust region solver (BXNL) from the NAG optimization modelling suite aimed for small to medium-scale problems.

2 Specification

#include <nag.h>
void  e04ggc (void *handle,
void (*lsqfun)(Integer nvar, const double x[], Integer nres, double rx[], Integer *inform, Nag_Comm *comm),
void (*lsqgrd)(Integer nvar, const double x[], Integer nres, Integer nnzrd, double rdx[], Integer *inform, Nag_Comm *comm),
void (*lsqhes)(Integer nvar, const double x[], Integer nres, const double lambda[], double hx[], Integer *inform, Nag_Comm *comm),
void (*lsqhprd)(Integer nvar, const double x[], const double y[], Integer nres, double hxy[], Integer *inform, Nag_Comm *comm),
void (*monit)(Integer nvar, const double x[], Integer *inform, const double rinfo[], const double stats[], Nag_Comm *comm),
Integer nvar, double x[], Integer nres, double rx[], double rinfo[], double stats[], Nag_Comm *comm, NagError *fail)
The function may be called by the names: e04ggc or nag_opt_handle_solve_bxnl.

3 Description

e04ggc computes a solution x to the nonlinear least squares problem
minimize x nvar f(x)= 12 i=1 nres [wiri(x)] 2 + σp x2p subject to lx x ux , (1)
where ri(x),i=1,,nres, are smooth nonlinear functions called residuals, wi ,i=1,,nres are weights (by default they are all defined to 1, see Section 9.2 on how to change them), and the rightmost element represents the regularization term with parameter σ0 and power p>0 (by default the regularization term is not used, see Section 11 on how to enable it). The constraint elements lx and ux are nvar-dimensional vectors defining the bounds on the variables.
Typically in a calibration or data fitting context, the residuals will be defined as the difference between the observed values yi at ti and the values provided by a nonlinear model ϕ(t;x), i.e., ri(x)yi-ϕ(ti;x). If these residuals (errors) follow a Gaussian distribution, then the values of the optimal parameter vector x* are the maximum likelihood estimates. For a description of the various algorithms implemented for solving this problem see Section 11. It is also recommended that you read Section 2.2.3 in the E04 Chapter Introduction.
e04ggc serves as a solver for problems stored as a handle. The handle points to an internal data structure which defines the problem and serves as a means of communication for functions in the NAG optimization modelling suite. First, the problem handle is initialized by calling e04rac. A nonlinear least square residual objective can be added by calling e04rmc and, optionally, (simple) box constraints can be defined by calling e04rhc. It should be noted that e04ggc internally works with a dense representation of the residual Jacobian even if a sparse structure was defined in the call to e04rmc. Once the problem is fully described, the handle may be passed to the solver e04ggc. When the handle is not needed anymore, e04rzc should be called to destroy it and deallocate the memory held within. For more information refer to the NAG optimization modelling suite in Section 4.1 in the E04 Chapter Introduction.
The algorithm is based on the trust region framework and its behaviour can be modified by various optional parameters (see Section 12) which can be set by e04zmc and e04zpc anytime between the initialization of the handle by e04rac and a call to the solver. Once the solver has finished, options may be modified for the next solve. The solver may be called repeatedly with various starting points and/or optional parameters. The option getter e04znc can be called to retrieve the current value of any option.
Several options might have significant impact on the performance of the solver. Even though the defaults were chosen to suit the majority of anticipated problems, it is recommended that you experiment with the option settings to find the most suitable set of options for a particular problem, see Sections 11 and 12 for further details.

4 References

Adachi S, Iwata S, Nakatsukasa Y, and Takeda A (2015) Solving the trust region subproblem by a generalized eigenvalue problem Technical report, METR 2015-14. Mathematical Engineering, The University of Tokyo https://www.keisu.t.u-tokyo.ac.jp/data/2015/METR15-14.pdf
Conn A R, Gould N I M and Toint Ph L (2000) Trust Region Methods SIAM, Philadephia
Gould N I M, Orban D, and Toint Ph L (2003) GALAHAD, a library of thread-safe Fortran 90 packages for large-scale nonlinear optimization ACM Transactions on Mathematical Software (TOMS) 29(4) 353–372
Gould N I M, Rees T, and Scott J A (2017) A higher order method for solving nonlinear least-squares problems Technical report, RAL-P-1027-010 RAL Library. STFC Rutherford Appleton Laboratory http://www.numerical.rl.ac.uk/people/rees/pdf/RAL-P-2017-010.pdf
Kanzow C, Yamashita N, and Fukushima M (2004) Levenberg-Marquardt methods with strong local convergence properties for solving nonlinear equations with convex constraints Journal of Computational and Applied Mathematics 174 375–397
Lanczos C (1956) Applied Analysis 272–280 Prentice Hall, Englewood Cliffs, NJ, USA
Nielsen H B (1999) Damping parameter in Marquadt’s Method Technical report TR IMM-REP-1999-05. Department of Mathematical Modelling, Technical University of Denmark http://www2.imm.dtu.dk/documents/ftp/tr99/tr05_99.pdf
Nocedal J and Wright S J (2006) Numerical Optimization (2nd Edition) Springer Series in Operations Research, Springer, New York

5 Arguments

1: handle void * Input
On entry: the handle to the problem. It needs to be initialized (e.g., by e04rac) and to hold a problem formulation compatible with e04ggc. It must not be changed between calls to the NAG optimization modelling suite.
2: lsqfun function, supplied by the user External Function
lsqfun must evaluate the value of the nonlinear residuals, ri(x)yi-ϕ(ti;x) ,i=1,,nres , at a specified point x.
The specification of lsqfun is:
void  lsqfun (Integer nvar, const double x[], Integer nres, double rx[], Integer *inform, Nag_Comm *comm)
1: nvar Integer Input
On entry: nvar, the current number of decision variables, x, in the model.
2: x[nvar] const double Input
On entry: x, the vector of variable values at which the residuals, ri, are to be evaluated.
3: nres Integer Input
On entry: nres, the current number of residuals in the model.
4: rx[nres] double Output
On exit: the value of the residual vector, r(x), evaluated at x.
5: inform Integer * Input/Output
On entry: a non-negative value.
On exit: may be used to indicate that some residuals could not be computed at the requested point. This can be done by setting inform to a negative value. The solver will attempt a rescue procedure and request an alternative point. If the rescue procedure fails, the solver will exit with fail.code= NE_USER_NAN.
6: comm Nag_Comm *
Pointer to structure of type Nag_Comm; the following members are relevant to lsqfun.
userdouble *
iuserInteger *
pPointer 
The type Pointer will be void *. Before calling e04ggc you may allocate memory and initialize these pointers with various quantities for use by lsqfun when called from e04ggc (see Section 3.1.1 in the Introduction to the NAG Library CL Interface).
3: lsqgrd function, supplied by the user External Function
lsqgrd evaluates the residual gradients, ri(x), at a specified point x.
The specification of lsqgrd is:
void  lsqgrd (Integer nvar, const double x[], Integer nres, Integer nnzrd, double rdx[], Integer *inform, Nag_Comm *comm)
1: nvar Integer Input
On entry: nvar, the current number of decision variables, x, in the model.
2: x[nvar] const double Input
On entry: x, the vector of variable values at which the residual gradients, ri(x), are to be evaluated.
3: nres Integer Input
On entry: nres, the current number of residuals in the model.
4: nnzrd Integer Input
On entry: the number of nonzeros in the first derivative matrix. If isparse =0 in the call to e04rmc (recommended use for e04ggc) then nnzrd=nvar*nres.
5: rdx[dim] double Input/Output
On entry: the elements should only be assigned and not referenced.
On exit: the vector containing the nonzero residual gradients evaluated at x,
r(x) = [r1(x),r2(x),,r nres (x)],  
where
ri(x)= [ ri(x) x1 ,, ri(x) x nvar ]T.  
The elements must be stored in the same order as the defined sparsity pattern provided in the call to e04rmc.
6: inform Integer * Input/Output
On entry: a non-negative value.
On exit: may be used to indicate that the residual gradients could not be computed at the requested point. This can be done by setting inform to a negative value. The solver will attempt a rescue procedure and request an alternative point. If the rescue procedure fails, the solver will exit with fail.code= NE_USER_NAN.
7: comm Nag_Comm *
Pointer to structure of type Nag_Comm; the following members are relevant to lsqgrd.
userdouble *
iuserInteger *
pPointer 
The type Pointer will be void *. Before calling e04ggc you may allocate memory and initialize these pointers with various quantities for use by lsqgrd when called from e04ggc (see Section 3.1.1 in the Introduction to the NAG Library CL Interface).
4: lsqhes function, supplied by the user External Function
lsqhes evaluates the residual Hessians, 2 ri (x) , at a specified point x. By default, the optional parameter Bxnl Use Second Derivatives=NO and lsqhes is never called. lsqhes may be specified as NULLFN.
This function will only be called if the optional parameter Bxnl Use Second Derivatives=YES and if the model (see Section 11.2) requires second order information. Under these circumstances, if you do not provide a valid lsqhes the solver will terminate with either fail.code= NE_DERIV_ERRORS or fail.code= NE_FAILED_START.
The specification of lsqhes is:
void  lsqhes (Integer nvar, const double x[], Integer nres, const double lambda[], double hx[], Integer *inform, Nag_Comm *comm)
1: nvar Integer Input
On entry: nvar, the current number of decision variables, x, in the model.
2: x[nvar] const double Input
On entry: x, the vector of decision variables at the current iteration.
3: nres Integer Input
On entry: nres, the current number of residuals in the model.
4: lambda[nres] const double Input
On entry: λ, the vector containing the (weighted) residuals at x, λi= wi ri(x). See (1) and Section 9.2.
5: hx[nvar×nvar] double Input/Output
On entry: the elements should only be assigned and not referenced.
On exit: a dense square (symmetric) matrix containing the weighted sum of residual Hessians,
H(x)= i=1 nres λi 2 ri (x),  
where
2 ri(x) = ( 2 x1 x1 ri (x) 2 x1 x2 ri (x) 2 x1 x nvar ri (x) 2 x2 x1 ri (x) 2 x2 x2 ri (x) 2 x2 x nvar ri (x) 2 x nvar x1 ri (x) 2 x nvar x2 ri (x) 2 x nvar x nvar ri (x) ) ,  
is also a dense square (symmetric) matrix containing the ith residual Hessian evaluated at the point x. All matrix elements must be provided: both upper and lower triangular parts.
6: inform Integer * Input/Output
On entry: a non-negative value.
On exit: may be used to indicate that one or more elements of the residual Hessian could not be computed at the requested point. This can be done by setting inform to a negative value. The solver will attempt a rescue procedure and if the rescue procedure fails, the solver will exit with fail.code= NE_USER_NAN.
7: comm Nag_Comm *
Pointer to structure of type Nag_Comm; the following members are relevant to lsqhes.
userdouble *
iuserInteger *
pPointer 
The type Pointer will be void *. Before calling e04ggc you may allocate memory and initialize these pointers with various quantities for use by lsqhes when called from e04ggc (see Section 3.1.1 in the Introduction to the NAG Library CL Interface).
5: lsqhprd function, supplied by the user External Function
lsqhprd evaluates the residual Hessians, 2 ri (x) , at a specified point, x, and performs matrix-vector products with a given vector, y, returning the dense matrix [ 2 r1(x)y, 2 r2(x)y,, 2 rnres(x)y] . If you do not supply this function, it may be specified as NULLFN.
The specification of lsqhprd is:
void  lsqhprd (Integer nvar, const double x[], const double y[], Integer nres, double hxy[], Integer *inform, Nag_Comm *comm)
1: nvar Integer Input
On entry: nvar, the current number of decision variables, x, in the model.
2: x[nvar] const double Input
On entry: x, the vector of decision variables at the current iteration.
3: y[nvar] const double Input
On entry: y, the vector used to perform the required matrix-vector products.
4: nres Integer Input
On entry: nres, the current number of residuals in the model.
5: hxy[nvar×nres] double Input/Output
Note: the (i,j)th element of the matrix is stored in hxy[(j-1)×nvar+i-1].
On entry: the elements should only be assigned and not referenced.
On exit: a dense matrix of size nvar×nres containing the following matrix-vector products,
H(x,y)= [ 2 r1(x)y, 2 r2(x)y,, 2 rnres(x)y] .  
6: inform Integer * Input/Output
On entry: The first call to lsqhprd will have a non-zero value and can be used to optimize your code in order to avoid recalculations of common quantities when evaluating the Hessians. For all other instances inform will have a value of zero. This notification parameter may be safely ignored if such optimization is not required.
On exit: may be used to indicate that one or more elements of the residual Hessian could not be computed at the requested point. This can be done by setting inform to a negative value. The solver will attempt a rescue procedure and if the rescue procedure fails, the solver will exit with fail.code= NE_USER_NAN. The value of inform returned on the first call is ignored.
7: comm Nag_Comm *
Pointer to structure of type Nag_Comm; the following members are relevant to lsqhprd.
userdouble *
iuserInteger *
pPointer 
The type Pointer will be void *. Before calling e04ggc you may allocate memory and initialize these pointers with various quantities for use by lsqhprd when called from e04ggc (see Section 3.1.1 in the Introduction to the NAG Library CL Interface).
6: monit function, supplied by the user External Function
monit is provided to enable monitoring of the progress of the optimization and, if necessary, to halt the optimization process.
If no monitoring is required, monit may be specified as NULLFN.
monit is called at the end of every ith step where i is controlled by the optional parameter Bxnl Monitor Frequency (the default value is 0, monit is not called).
The specification of monit is:
void  monit (Integer nvar, const double x[], Integer *inform, const double rinfo[], const double stats[], Nag_Comm *comm)
1: nvar Integer Input
On entry: nvar, the current number of decision variables, x, in the model.
2: x[nvar] const double Input
On entry: the current best point.
3: inform Integer * Input/Output
On entry: a non-negative value.
On exit: may be used to request the solver to stop immediately by setting inform to a non-zero value in which case it will terminate with fail.code= NE_USER_STOP; otherwise, the solver will proceed normally.
4: rinfo[100] const double Input
On entry: best objective value computed and various indicators (the values are as described in the main argument rinfo).
5: stats[100] const double Input
On entry: solver statistics at monitoring steps or at the end of the current iteration (the values are as described in the main argument stats).
6: comm Nag_Comm *
Pointer to structure of type Nag_Comm; the following members are relevant to monit.
userdouble *
iuserInteger *
pPointer 
The type Pointer will be void *. Before calling e04ggc you may allocate memory and initialize these pointers with various quantities for use by monit when called from e04ggc (see Section 3.1.1 in the Introduction to the NAG Library CL Interface).
7: nvar Integer Input
On entry: nvar, the current number of decision variables, x, in the model.
8: x[nvar] double Input/Output
On entry: x0, the initial estimates of the variables, x.
On exit: the final values of the variables, x.
9: nres Integer Input
On entry: nres, the current number of residuals in the model.
10: rx[nres] double Output
On exit: the values of the residuals at the final point given in x.
11: rinfo[100] double Output
On exit: objective value and various indicators at monitoring steps or at the end of the final iteration. The measures are given in the table below:
0 Objective function value, f(x).
1 Norm of the projected gradient at the current iterate, see PG STEP in Section 11.4 and (8) in Section 11.5.
2 Norm of the scaled projected gradient at the current iterate, see (8) in Section 11.5
3 Norm of the step between the current and previous iterate.
4 Convergence tests result. A scalar value between 0-7 indicates whether a convergence test has passed. Specifically, 1 indicates small objective test passed, 2 indicates small (scaled) gradient test passed, 4 indicates small step test passed. In the case where two or more tests passed, they are accumulated.
5 Norm of the current iterate x. If regularization is requested, then this value was used in the regularization and it might differ from x if x has fixed or disabled elements.
6-99 Reserved for future use.
12: stats[100] double Output
On exit: solver statistics at monitoring steps or at the end of the final iteration as given in the table below:
0 Number of iterations performed.
1 Total number of calls to the objective function lsqfun.
2 Total number of calls to the objective gradient function lsqgrd.
3 Total number of calls to the objective Hessian function lsqhes.
4 Total time in seconds spent in the solver. It includes time spent in user-supplied subroutines.
5 Number of calls to the objective function lsqfun required by linesearch steps.
6 Number of calls to the objective gradient function lsqgrd required by linesearch steps.
7 Number of calls to the objective function lsqfun required by projected gradient steps.
8 Number of calls to the objective gradient function lsqgrd required by projected gradient steps.
9 Number of inner iterations performed, see option Bxnl Model=TENSOR-NEWTON.
10 Number of linesearch iterations performed.
11 Number of projected gradient iterations performed.
12 Total number of calls to the objective auxiliary Hessian function lsqhprd.
13-99 Reserved for future use.
13: comm Nag_Comm *
The NAG communication argument (see Section 3.1.1 in the Introduction to the NAG Library CL Interface).
14: fail NagError * Input/Output
The NAG error argument (see Section 7 in the Introduction to the NAG Library CL Interface).
e04ggc returns with fail.code= NE_NOERROR if the iterates have converged to a point x that satisfies the convergence criteria described in Section 11.5.

6 Error Indicators and Warnings

NE_ALLOC_FAIL
Dynamic memory allocation failed.
See Section 3.1.2 in the Introduction to the NAG Library CL Interface for further information.
NE_BAD_PARAM
On entry, argument value had an illegal value.
NE_DC_MATCH
Data for residual weights not found or is invalid.
Custom residual weights are required, optional parameter Bxnl Use Weights=YES, but the weights data is missing, of the wrong expected size or has invalid values. Please refer to Section 9.2.
NE_DERIV_ERRORS
Exact second derivatives needed for tensor model.
Model in the optional parameter Bxnl Model=TENSOR-NEWTON requires exact second derivatives but Bxnl Use Second Derivatives=NO.
Provide second derivatives via lsqhes and optionally lsqhprd functions, and set optional parameter Bxnl Use Second Derivatives=YES.
NE_FAILED_START
The current starting point is unusable.
While trying to evaluate the starting point x0, either inform was set to a non-zero value within the user-supplied functions, lsqfun, lsqgrd or lsqhes, or an Infinity or NaN was detected in values returned from them.
NE_HANDLE
The supplied handle does not define a valid handle to the data structure for the NAG optimization modelling suite. It has not been properly initialized or it has been corrupted.
NE_INTERNAL_ERROR
An internal error has occurred in this function. Check the function call and any array sizes. If the call is correct then please contact NAG for assistance.
See Section 7.5 in the Introduction to the NAG Library CL Interface for further information.
NE_NO_IMPROVEMENT
The solver was terminated because no further progress could be achieved.
This can indicate that the solver is calculating very small step sizes and is making very little progress. It could also indicate that the problem has been solved to the best numerical accuracy possible given the current scaling.
It can also indicate that a recovery procedure was interrupted due to the user-supplied function lsqgrd being incorrect.
NE_NO_LICENCE
Your licence key may have expired or may not have been installed correctly.
See Section 8 in the Introduction to the NAG Library CL Interface for further information.
NE_NUM_DIFFICULTIES
Numerical difficulties encountered and solver was terminated.
This error can be caused by an ill-posed or poorly scaled problem.
NE_PHASE
The problem is already being solved.
Unsupported model and method chosen. The specified model in optional parameter Bxnl Model is not supported by the method specified by the optional parameter Bxnl Nlls Method=POWELL-DOGLEG.
Unsupported option combinations. The specified combination of values for optional parameters Bxnl Nlls Method and Bxnl Glob Method is not supported.
NE_REF_MATCH
On entry, nvar=value, expected value=value.
Constraint: nvar must match the current number of variables of the model in the handle.
The information supplied does not match with that previously stored.
On entry, nres=value must match that given during the definition of the objective in the handle, i.e., value.
There are no decision variables. nvar must be greater than zero.
NE_SETUP_ERROR
This solver does not support the model defined in the handle.
NE_TIME_LIMIT
The solver terminated after the maximum time allowed was exceeded.
Maximum number of seconds exceeded. Use optional parameter Time Limit to reset the limit.
NE_TOO_MANY_ITER
Maximum number of iterations reached.
Use optional parameter Bxnl Iteration Limit to reset the limit.
NE_TOO_MANY_MINOR_ITER
Iteration limit reached while solving a subproblem.
Maximum number of iterations reached while trying to solve an auxiliary subproblem.
Line Search failed.
Line Search in the projected gradient direction did not find an acceptable new iterate.
NE_USER_NAN
Invalid number detected in user-supplied function and recovery failed.
Either inform was set to a non-zero value within one of the user-supplied functions, lsqfun, lsqgrd, lsqhes, or lsqhprd, or an Infinity or NaN was detected in values returned from them and the recovery attempt failed.
NE_USER_STOP
User requested termination during a monitoring step. inform was set to a non-zero value within the user-supplied function monit.

7 Accuracy

The accuracy of the solution is determined by optional parameters Bxnl Stop Abs Tol Fun, Bxnl Stop Abs Tol Grd, Bxnl Stop Rel Tol Fun, Bxnl Stop Rel Tol Grd, and Bxnl Stop Step Tol. If fail.code= NE_NOERROR on exit, the returned point satisfies (7), (8) or (9) to the defined accuracies.
Please refer to Section 11.5 and the description of the particular options in Section 12.

8 Parallelism and Performance

e04ggc is threaded by NAG for parallel execution in multithreaded implementations of the NAG Library.
e04ggc makes calls to BLAS and/or LAPACK routines, which may be threaded within the vendor library used by this implementation. Consult the documentation for the vendor library for further information.
Please consult the X06 Chapter Introduction for information on how to control and interrogate the OpenMP environment used within this function. Please also consult the Users' Note for your implementation for any additional implementation-specific information.

9 Further Comments

9.1 Description of the Printed Output

The solver can print information to give an overview of the problem and the progress of the computation. The output may be sent to two independent file ID which are set by optional parameters Print File and Monitoring File. Optional parameters Print Level, Print Options, Monitoring Level and Print Solution determine the exposed level of detail. This allows, for example, a detailed log file to be generated while the condensed information is displayed on the screen.
By default (Print File=6, Print Level=2), the following sections are printed to the standard output:
Header
The header is a message indicating the start of the solver. It should look like:
----------------------------------------------------------------------
 E04GG, Nonlinear least squares method for bound-constrained problems
----------------------------------------------------------------------
Optional parameters list
If Print Options=YES, a list of the optional parameters and their values is printed before the problem statistics. The list shows all options of the solver, each displayed on one line. Each line contains the option name, its current value and an indicator for how it was set. The options unchanged from their defaults are noted by ‘d’ and the ones you have set are noted by ‘U’. Note that the output format is compatible with the file format expected by e04zpc. The output looks similar to:
Begin of Options
    Bxnl Model                    =        Gauss-newton     * U
    Bxnl Nlls Method              =             Galahad     * d
    Bxnl Glob Method              =                 Reg     * U
    Bxnl Reg Order                =                Auto     * d
    Bxnl Tn Method                =           Min-1-var     * d
    Bxnl Basereg Pow              =         2.00000E+00     * d
    Bxnl Basereg Term             =         1.00000E-02     * d
    Bxnl Iteration Limit          =                1000     * d
    Bxnl Use Second Derivatives   =                 Yes     * U
End of Options
Problem statistics
If Print Level2, statistics on the problem are printed, for example:
Problem Statistics
  No of variables                  4 (+2 disabled, +0 fixed)
    free (unconstrained)           3
    bounded                        1
  Objective function    LeastSquares
    No of residuals               16 (+8 disabled)
Iteration log
If Print Level=2, the solver will print a summary line for each step. An iteration is considered successful when it yields a decrease of the objective, either sufficiently close to the decrease predicted by the model or to a given relative threshold. Each line shows the iteration number (Iter), the value of the objective function (error), the absolute and relative norms for the projected gradient (optim) and (rel optim), this last one is used in the convergence test of equation (8). The output looks as follows:
--------------------------------------------
  Iter |   error  |   optim    |  rel optim
--------------------------------------------
      0 3.6953E+01  1.30501E+01  1.51801E+00
      1 5.6464E-01  1.61396E+00  1.51877E+00
      2 7.7340E-02  7.06213E-01  1.79563E+00
      3 2.2354E-02  5.21619E-01  2.46694E+00
      4 7.0758E-03  3.11695E-01  2.62016E+00
      5 2.2392E-03  1.76206E-01  2.63307E+00
If Print Level3, each line additionally shows the current value of the trust region radius (Delta), quality of the model (rho), some flags relating to the iteration (S2IF), inner iteration counter (inn it) for the tensor Newton model, the step length taken (step), trust region loop exit status (loop), performed line search type (LS), as well as the projection factor over the constraints (tau). It might look as follows:
---------------------------------------------------------------------------------------------------
  Iter |   error  |   optim    |  rel optim |  Delta  |   rho   |S2IF| inn it|  step | loop|LS| tau
---------------------------------------------------------------------------------------------------
     65 2.2217E-06  3.04184E-05  1.44305E-02  4.00E+02  9.22E-01 SNR-       0 1.3E-03 gamma    1.00
     66 2.2203E-06  1.44158E-05  6.84102E-03  8.00E+02  9.82E-01 SNR-       0 2.6E-03 gamma    1.00
     67 2.2177E-06  1.49444E-06  7.09599E-04  1.60E+03  9.84E-01 SNR-       0 4.9E-03 TR Ok    1.00
     68 2.2132E-06  3.44970E-06  1.63968E-03  3.20E+03  9.71E-01 SNR-       0 9.2E-03 TR Ok    1.00
     69 2.2054E-06  1.01207E-05  4.81891E-03  6.40E+03  9.47E-01 SNR-       0 1.7E-02 TR Ok    1.00
Iteration flags column (S2IF) contains four flags related to the iteration. Flag ‘S’ indicates if the trust region iteration was successful (S) or unsuccessful (U). Flag ‘2’ shows if iteration used second-order information: yes (Y), no (N), tensor (T), or approximate (A). Flag ‘I’ indicates iteration type: regular (R) or inner (I). Exit flag of inner solver ‘F’ has three states: subproblem converged (C), not solved (E), or current iteration is inside subproblem or tensor model not used (-). For details on the interpretation of rho and tau, see Section 11.
If Tensor-Newton model is chosen, then details of each inner iteration can be printed by setting Print Level=4, output is similar to:
---------------------------------------------------------------------------------------------------
  Iter |   error  |   optim    |  rel optim |  Delta  |   rho   |S2IF| inn it|  step | loop|LS| tau
---------------------------------------------------------------------------------------------------
      0 3.6953E+01  1.30501E+01  1.51801E+00  1.00E+02 -1.00E+00 ----
---------------------------------------------------------------------------------------------------
      0 3.6953E+01  2.00107E+01  2.32768E+00  1.00E+02 -1.00E+00 --I-
      1                                       3.91E-01 -1.35E-01 UNI-
      1 3.2210E+01  2.19541E+01  2.73529E+00  3.91E-01  3.15E-01 SNI-         9.6E-01 TR Ok    1.00
      2                                       1.95E-01 -6.51E-01 UYI-
      2 3.2210E+01  2.19541E+01  2.73529E+00  1.23E-30  1.00E+00 SYI-         9.2E-16 TR Ok    1.00
      3                                       3.08E-31 -2.00E+00 UYI-
      3 3.2210E+01  2.19541E+01  2.73529E+00  6.16E-31  1.00E+00 SYI-         6.5E-16 TR Ok    1.00
---------------------------------------------------------------------------------------------------
      1 1.9040E+01  9.07931E+00  1.47131E+00  1.00E+02  3.77E+00 STRC       4 9.6E-01 TR Ok    1.00
---------------------------------------------------------------------------------------------------
      0 1.9040E+01  1.59354E+01  2.58235E+00  1.00E+02 -1.00E+00 --I-
      1                                       2.50E+01 -4.67E+00 UNI-
      1 1.5631E+01  1.59186E+01  2.84704E+00  3.91E-01  4.09E-01 SNI-         6.4E-01 TR Ok    1.00
      2 1.5631E+01  1.59186E+01  2.84704E+00  9.86E-30  1.00E+00 SYI-         1.9E-15 TR Ok    1.00
      3 1.5631E+01  1.59186E+01  2.84704E+00  6.16E-31  1.00E+00 SYI-         4.8E-16 TR Ok    1.00
---------------------------------------------------------------------------------------------------
      2 8.5311E+00  6.26943E+00  1.51779E+00  1.00E+02  3.08E+00 STRC       9 6.4E-01 TR Ok    1.00
Note the iteration type flag ‘I’ change under the S2IF column, the output reports on 2 (R) regular iterations where each required 3 (I) inner iterations.
Additionally, if Print Level=5, each iteration produces more information that expands over several lines. This additional information can contain:
The output might look as follows:
*** Solving the trust region subproblem using More-Sorensen ***
A is symmetric positive definite
iter    nd            sigma         sigma_shift
   0    3.6778E-01    0.0000E+00    0.0000E+00
nq =   7.7000E+02
   1    1.2571E-02    6.4469E-06    6.4469E-06
We're within the trust region radius
Leaving More-Sorensen
Model evaluated successfully: m_k(d) =   2.3089E-08
*** Subproblem solution found ***
Actual reduction (in cost function) =   1.2065E-09
Predicted reduction (in model) =   4.1866E-09
rho returned =   2.8819E-01
Successful step -- Delta staying at  1.2570E-02
Summary
Once the solver finishes, a summary is produced:
 -------------------------------------------------------------------------------
 Status: converged, an optimal solution was found
         small (scaled) projected gradient norm
 -------------------------------------------------------------------------------
 Value of the objective             2.17328E-06
 Norm of projected gradient         1.51989E-08
 Norm of scaled projected gradient  7.29019E-06
 Norm of step                       4.98107E-04
 Iterations                                  80
   Inner iterations                           0
   LS iterations                              0
   PG iterations                              0
 Function evaluations                        81
 Gradient evaluations                        81
 Hessian evaluations (objhes)                 0
 Hessian evaluations (objhprd)                0
   LS function calls                          0
   LS gradient calls                          0
   PG function calls                          0
   PG gradient calls                          0
Optionally, if Stats Time=YES, the timings are printed:
 Timing
   Total time spent                        2.43 sec
 -------------------------------------------------------------------------------
Solution
If Print Solution=YES, the values of the primal variables are printed, furthermore if the problem is constrained, the dual variables are also reported, see Lagrangian Multipliers in e04kfc and the dual variables storage format described in Section 3.1 in e04svc. It might look as follows:
 Primal variables:
   idx   Lower bound       Value       Upper bound
     1   0.00000E+00    4.58516E-01    1.00000E+00
     2       -inf       3.05448E+00         inf
     3       -inf       4.65146E+00         inf
     4     Disabled         NaN          Disabled
 
 Box bounds dual variables:
   idx   Lower bound       Value       Upper bound       Value
     1   0.00000E+00    0.00000E+00    1.00000E+00    9.52218E-10
     2       -inf       4.66962E-11         inf       0.00000E+00
     3       -inf       6.55098E-11         inf       0.00000E+00
     4     Disabled         NaN          Disabled         NaN

9.2 Residual Weights

A typical use for weights in the least square fitting context is to account for uncertainity in the observed data, σi2, by setting the weights to
w nres  with  wi2= 1 σi2 , for ​ i=1,,nres .  
The idea behind this choice is to give less importance (small weight) to measurements which have large variance.
In order to use weights,
  1. 1.request to use weights by setting the optional parameter Bxnl Use Weights=YES (this will request the solver to query the handle for an array of weights), and
  2. 2.store the weights array in the handle. This is done by calling e04rxc with the command cmdstr='Residual Weights' and passing the array length and weights array, lrarr=nres and rarr=w, respectively. Weights are required for each residual and all weights must be positive.
These steps must be done after the handle is initialized (via e.g., e04rac) but before calling the solver e04ggc. The stored weights in the handle will only be accessed if Bxnl Use Weights=YES, otherwise all weights are assumed to be 1 and the handle is not queried for residual weights.
If the solver is expecting to use weights but they are not provided, or the array length is wrong or have non-positive values, then the solver will terminate with fail.code= NE_DC_MATCH.

9.3 Internal Changes

Internal changes have been made to this function as follows:
For details of all known issues which have been reported for the NAG Library please refer to the Known Issues.

10 Example

In this example we solve the Lanczos-3 Lanczos (1956) problem, a nonlinear least squares regression. The regression problem consists of nres=24 observations, (t,y), to be fitted over the nvar=6 parameter model
y= ϕ (t;x) + ε = x1 e -x2t + x3 e -x4t + x5 e -x6t + ε ,  
and the residuals for the regression are
ri (x) = yi - ϕ (ti,x) i = 1 : nres ,  
where
t = ( 0.0000,0 0.0500,0 0.1000,0 0.1500,0 0.2000,0 0.2500,0 0.3000,0 0.3500,0 0.4000,0 0.4500, ( 0.5000,0 0.5500,0 0.6000,0 0.6500,0 0.7000,0 0.7500,0 8.0000,0 8.5000,0 9.0000,0 0.9500,0 ( 1.0000,0 1.0500,0 1.1000,0 1.1500 ) ; y = ( 2.5134,0 2.0443,0 1.6684,0 1.3664,0 1.1232,0 0.9269,0 0.7679,0 0.6389,0 0.5338,0 0.4479,0 ( 0.3776,0 0.3197,0 0.2720,0 0.2325,0 0.1997,0 0.1723,0 0.1493,0 0.1301,0 0.1138,0 0.1000,0 ( 0.0883,0 0.0783,0 0.0698,0 0.0624).  
The following bounds are defined on the variables
0.0 x1 1.0, -1.0 x2, -1.0 x3, -1.0 x4, -1.0 x5 1.0, -1.0 x6 10.0.  
The initial guess is x0=(1.2,0.3,5.6,5.5,6.5,7.6).
The expected solution is x*=(0.4445,1.8734,3.0665,4.6404,-0.9982,4.6409) with residual norm close to 2.1×10−6.

10.1 Program Text

Program Text (e04ggce.c)

10.2 Program Data

None.

10.3 Program Results

Program Results (e04ggce.r)

11 Algorithmic Details

This section contains a short description of the underlying algorithms used in e04ggc, a bound-constrained nonlinear least squares (BXNL) solver that uses a model-based trust region framework adapted to exploit the least squares problem structure. It is based on a collaborative work between NAG and the STFC Rutherford Appleton Laboratory. For further details, see Gould et al. (2017) and references therein.

11.1 Trust Region Algorithm

In this section, we are interested in generic nonlinear least squares problems of the form
minimize x nvar f(x)= 12 i=1 nres [wiri(x)] 2 + σp x2p subject to lx x ux . (2)
where ri(x), i=1,,nres, are smooth nonlinear functions called redisuals, wi>0,i=1,,nres are the weights (by default they are all defined to 1, see Section 9.2 on how to change them), and the rightmost element represents the optional regularization term of parameter σ0 and power p>0. The constraint elements lx and ux are nvar-dimensional vectors defining the bounds on the variables. For the rest of this chapter, and without any loss of generality, it is assumed that weights are all set to the default value of 1 and are excluded from the formulae. e04ggc is an iterative framework for solving (2) which consists of a variety of algorithms that solve the trust region subproblem. The fundamental ideas of the framework follow.
At each point xk, the algorithm builds a model of the function at the next step, f(xk+sk), which we refer to as mk (see Section 11.2).
Once the model has been formed, the candidate for the next point is found by solving a suitable subproblem (see Section 11.3). Let PΩ (x) be the Euclidean projection operator over the feasible set, then the quantity
ρk= f (xk) - f (PΩ(xk+sk)) mk (xk) - mk (PΩ(xk+sk)) ,  
is used to assess the quality of the proposed step. If it is sufficiently large we accept the step and xk+1 is set to PΩ(xk+sk); if not, the trust region radius Δk is reduced and the resulting new trust region subproblem is solved. If the step is very successful (ρ is close to 1), Δk is increased.
Under certain circumstances, it is deemed that the projection of the current point with the trust region step will not produce a successful point and the new step sk is calculated using a convenient line search step.
This process continues until a point is found that satisfies the stopping criteria described in Section 11.5. More precisely, it can be described as:
Note: the use of the regularization term in (2) is optional and is not used by default. To enable regularization please refer to the optional parameters Bxnl Basereg Type, Bxnl Basereg Pow, Bxnl Basereg Term, and Bxnl Reg Order.

11.2 Models

A vital component of the algorithm is the choice of model employed. There are four choices available which are controlled by the optional parameter Bxnl Model.
Gauss–Newton
This option specifies to the solver that it use the Gauss–Newton model. For this case, r(xk+s) is replaced by its first-order Taylor approximation, r(xk)+r(xk)Ts=r(xk)+Jks. The model is, therefore, given by
mk (s) = mkGN (s) = 12 r(xk)+Jks 2 .  
Quasi–Newton
This option specifies to use a Newton type model. For this case, the model is taken to be the second-order Taylor approximation of the objective function f(xk+1). For this choice of model the gradient and Hessian are gk=JkTr(xk) and Hk=i=1nresri(xk)2ri(xk). The model is given by
mk (s) = mkQN (s) = f(xk) + gkTs + 12 sT Hk s .  
If the second derivatives of r(x) are not available (i.e., the optional parameter Bxnl Use Second Derivatives=No), then the method approximates the matrix Hk. If Print Level3, the flag ‘2’ in the iteration log will display (A), see Iteration log in Section 9.1.
Hybrid
This option specifies to the solver that it use the hybrid model. In practice the Gauss–Newton model tends to work well far away from the solution, whereas the Newton model performs better once it is near to the minimum (particularly if the residual is large at the solution). This option tells the solver to switch between the previous two models, picking the model that is most appropriate for the step. In particular, it starts by using mkGN and switches to mkQN when it considers it is close enough to the solution. If, in subsequent iterations, it fails to get a decrease in the function value, then the algorithm interprets this as being not sufficiently close to the solution and switches back to using the Gauss–Newton model.
Tensor–Newton
This option specifies to the solver that it use the tensor model. The model is based on a second-order Taylor approximation to the residual, ri(xk+s)(tk(s))iri(xk)+(Jk)is+12sT2ri(xk)s, where (J)i is the ith row of J. The tensor model used is
mkTN(s)=12 tk(s) 2. (3)

11.3 Subproblems

The next point xk+1 is estimated by finding a step, sk, that minimizes the model chosen in Bxnl Model, subject to a globalization strategy. e04ggc supports the use of two such strategies: trust region or regularization, these can be set using the optional parameter Bxnl Glob Method=TR or REG respectively. If Bxnl Model=Gauss-Newton, Quasi-Newton or Hybrid, then the model is quadratic and the available methods to solve the subproblem are described in the next two subsections. If the Bxnl Model=TENSOR-NEWTON, then the model is not quadratic and the methods available are described in Section 11.3.3.

11.3.1 Trust region method

The methods mentioned in this subsection are only used when Bxnl Model=Gauss-Newton, Quasi-Newton or Hybrid and Bxnl Glob Method=TR. The trust region subproblem to solve is
sk = arg min snvar mk (s)   subject to   sΔk . (4)
The next step is taken to be the solution of the previous problem and the method used to solve it is selected using the optional parameter Bxnl Nlls Method. The methods available are:

11.3.2 Regularization

The methods mentioned in this subsection are only used when Bxnl Model=Gauss-Newton, Quasi-Newton or Hybrid and Bxnl Glob Method=REG. The regularized subproblem to solve is
sk = arg min snvar mk(s) + 1 Δk 1p sp . (5)
The next step to take is the solution to the previous problem. The methods provided to solve (5) are

11.3.3 Tensor Newton subproblem

This section describes the regularized methods used to solve the non-quadratic tensor model (3) subproblem, i.e., the step subproblem when Bxnl Model=TENSOR-NEWTON. The schemes implemented find the next step by solving
minimize snvar 12 i=1 nres (tk(s)) i 2 . (6)
Note that (6) is also a sum-of-squares problem and, as such, can be solved by recursively calling e04ggc. In this context, the iterations performed by the recursive call to the solver are called inner iterations, otherwise they are called regular or outer iterations. When Print Level3, the iteration type is shown under the flag ‘I’ of the ‘iteration flags’ column while the inner iteration count is shown under the column ‘inn it’ of the Iteration log (see Section 9.1). The method used to solve (6) can be chosen by the optional parameter Bxnl Tn Method and the implemented methods are:

11.3.4 Incorporating the regularizer

The method used to incorporate the regularization specified by σ and p in problem (2) is defined using the optional parameter Bxnl Basereg Type. The implemented choices are:

11.4 Bound Constraints

e04ggc handles the bound constraints by projecting candidate points into the feasible set. The implemented framework is an adaptation of Algorithm 3.12 described in Kanzow et al. (2004), where the Levenberg–Marquardt step is replaced by a trust region (TR) step. The framework consists of three major steps. It first attempts a projected TR step and, if unsuccessful, attempts a Wolfe-type line search step along the projected TR step direction, otherwise, it defaults to a projected gradient step with an Armijo-type line search, specifically,

11.5 Stopping Criteria

The solver considers that it has found a solution and stops when at least one of the following three conditions is met within the defined absolute or relative tolerances (εabsf>0,εrelf>0,εabsg>0,εrelg>0,εstep>0) ,
r(xk) max(εabsf, εrelf r(x0) ) , (7)
d k PG r( x k ) max(εabsg, εrelg d 0 PG r( x 0 ) ) , (8)
xk-xk-1 εstep. (9)
Where dkPG is the projected gradient (see PG step in Section 11.4) and is reported in the column optim of the output while the left-hand side of (8) is reported in the column rel optim, see Iteration log in Section 9.1. If the problem is unconstrained, then the projected gradient reduces to the gradient and the convergence tests are done over the gradient norm. The stopping tolerances can be changed using the optional parameters Bxnl Stop Abs Tol Fun, Bxnl Stop Abs Tol Grd, Bxnl Stop Rel Tol Fun, Bxnl Stop Rel Tol Grd, and Bxnl Stop Step Tol. see Section 12 for details. If these parameters are set too small in relation to the complexity and scaling of the problem, the function can terminate with fail.code= NE_NO_IMPROVEMENT, NE_TOO_MANY_ITER or NE_TOO_MANY_MINOR_ITER.

11.6 A Note About Lagrangian Multipliers

It is often useful to have access to the Lagrangian multipliers (dual variables) associated with the constraints if there are any defined. In the case where only simple bounds are present, the multipliers directly relate to the values of the gradient at the solution. The multipliers of the active bounds are the absolute values of the associated elements of the gradient. The multipliers of the inactive bounds are always zero.
The multipliers based on the final gradient value can be retrieved by calling e04rxc with the command string cmdstr = Dual Variables. The format is the same as for other functions, see Section 3.1 in e04svc in e04svc.
Note that if the problem has not fully converged, the provided approximation might be quite crude.

12 Optional Parameters

Several optional parameters in e04ggc define choices in the problem specification or the algorithm logic. In order to reduce the number of formal arguments of e04ggc these optional parameters have associated default values that are appropriate for most problems. Therefore, you need only specify those optional parameters whose values are to be different from their default values.
The remainder of this section can be skipped if you wish to use the default values for all optional parameters.
The optional parameters can be changed by calling e04zmc anytime between the initialization of the handle and the call to the solver. Modification of the optional parameters during intermediate monitoring stops is not allowed. Once the solver finishes, the optional parameters can be altered again for the next solve.
The option values may be retrieved by e04znc.
The following is a list of the optional parameters available. A full description of each optional parameter is provided in Section 12.1.

12.1 Description of the Optional Parameters

For each option, we give a summary line, a description of the optional parameter and details of constraints.
The summary line contains:
All options accept the value DEFAULT to return single options to their default states.
Keywords and character values are case and white space insensitive.
Defaults
This special keyword may be used to reset all optional parameters to their default values. Any value given with this keyword will be ignored.
Bxnl Basereg PowrDefault =2.0
This parameter defines the regularization power p in (1) and for the tensor Newton subproblem (when Bxnl Tn Method=IMPLICIT). Some values are restricted depending on the type of regularization specified, see Bxnl Basereg Type for more details.
Constraint: Bxnl Basereg Pow>0.
Bxnl Basereg TermrDefault =0.01
This parameter defines the regularization term σ in (1) and for the tensor Newton subproblem (when Bxnl Tn Method=IMPLICIT).
Constraint: Bxnl Basereg Term>0.
Bxnl Basereg TypeaDefault =NONE
This parameter specifies the method used to incorporate the regularizer into (1) and optionally into the tensor Newton subproblem (when Bxnl Model=Tensor-Newton and Bxnl Tn Method=IMPLICIT).
The option Bxnl Basereg Type=EXPAND-NVAR-DOF reformulates the original problem by expanding it with nvar degrees of freedom that is subsequently solved. For the case Bxnl Basereg Type=EXPAND-1-DOF the residual vector is extended with a new term of the form σpx2p; for this method a value of p=3 is recommended.
If Bxnl Basereg Type=EXPAND-NVAR-DOF then the regularization power term p must be 2.0, that is Bxnl Basereg Pow=2.0. For further details see Section 11.3.
Constraint: Bxnl Basereg Type=NONE, EXPAND-NVAR-DOF or EXPAND-1-DOF.
Bxnl Save Covariance MatrixaDefault =NO
This parameter indicates to the solver to store the covariance matrix into the handle.
If Bxnl Save Covariance Matrix=YES then the lower triangle part of the covariance matrix is stored in packed column order (see Section 3.4.2 in the F07 Chapter Introduction) into the handle and can be retrieved via e04rxc using cmdstr=COVARIANCE MATRIX with lrarr= (nvar×(nvar+1)) /2 .
In the special case where Bxnl Save Covariance Matrix=VARIANCE, only the diagonal elements of the covariance matrix are stored in the handle and can be retrieved via e04rxc using cmdstr=VARIANCE with lrarr= nvar.
Similarly, if Bxnl Save Covariance Matrix=HESSIAN then the lower triangle part of the matrix H(x) = r(x) r(x)T = J(x)T J(x) is stored in packed column order into the handle and can be retrieved via e04rxc using cmdstr=HESSIAN MATRIX with lrarr= (nvar×(nvar+1)) /2 .
Limitations: If the number of enabled residuals is not greater than the number of enabled variables, or the pseudo-inverse of H(x) could not be calculated, then the covariance matrix (variance vector) is not stored in the handle and will not be available.
For more information on how the covariance matrix is estimated, see e04ycc.
Constraint: Bxnl Save Covariance Matrix=NO, YES, VARIANCE or HESSIAN.
Bxnl Stop Abs Tol FunrDefault =2.2ε13
This parameter specifies the relative tolerance for the error test, specifically, it sets the value of εabsf of equation (7) in Section 11.5. Setting Bxnl Stop Abs Tol Fun to a large value may cause the solver to stop prematurely with a suboptimal solution.
Constraint: Bxnl Stop Abs Tol Fun>0.
Bxnl Stop Abs Tol GrdrDefault =ε12
This parameter specifies the relative tolerance for the gradient test, specifically, it sets the value of εabsg of equation (8) in Section 11.5. Setting Bxnl Stop Abs Tol Grd to a large value may cause the solver to stop prematurely with a suboptimal solution.
Constraint: Bxnl Stop Abs Tol Grd>0.
Bxnl Stop Rel Tol FunrDefault =ε12
This parameter specifies the relative tolerance for the error test, specifically, it sets the value of εrelf of equation (7) in Section 11.5. Setting Bxnl Stop Rel Tol Fun to a large value may cause the solver to stop prematurely with a suboptimal solution.
Constraint: Bxnl Stop Rel Tol Fun>0.
Bxnl Stop Rel Tol GrdrDefault =ε12
This parameter specifies the relative tolerance for the gradient test, specifically, it sets the value of εrelg of equation (8) in Section 11.5. Setting Bxnl Stop Rel Tol Grd to a large value may cause the solver to stop prematurely with a suboptimal solution.
Constraint: Bxnl Stop Rel Tol Grd>0.
Bxnl Stop Step TolrDefault =2ε
Specifies the stopping tolerance for the step length test, specifically, it sets the value for εstep of equation (9) in Section 11.5. Setting Bxnl Stop Step Tol to a large value may cause the solver to stop prematurely with a suboptimal solution.
Under certain circumstances, e.g., when in doubt of the quality of the first- or second-order derivatives, in the event of the solver exiting with a successful step length test, it is recommended to verify that either the error or the gradient norm is acceptably small.
Constraint: Bxnl Stop Step Tol>0.
Bxnl Reg OrderaDefault =AUTO
This parameter specifies the order of the regularization p in (5) used when Bxnl Glob Method=Reg.
Some values for p are restricted depending on the method chosen in Bxnl Nlls Method, see Section 11.3.2 for more details.
Constraint: Bxnl Reg Order=AUTO, QUADRATIC or CUBIC.
Bxnl Glob MethodaDefault =TR
This parameter specifies the globalization method used to estimate the next step sk. It also determines the class of subproblem to solve. The trust region subproblem finds the step by minimizing the specified model withing a given radius. On the other hand, when Bxnl Glob Method=reg, the problem is reformulated by adding an aditional regularization term and minimized in order to find the next step sk. See Section 11.3 for more details.
Constraint: Bxnl Glob Method=TR or REG.
Bxnl Nlls MethodaDefault =GALAHAD
This parameter defines the method used to estimate the next step sk in xk+1=xk+sk. It only applies to Bxnl Model=GAUSS-NEWTON, QUASI-NEWTON or HYBRID. When the globalization technique chosen is trust region (Bxnl Glob Method=tr) the methods for Bxnl Nlls Method available are Powell's dogleg method, a generalized eigenvalue method (AINT) Adachi et al. (2015), a variant of Moré–Sorensen's method, and GALAHAD's DTRS method. Otherwise, when the globalization method chosen is via regularization (Bxnl Glob Method=reg) the methods available are comprised by a linear system solver and GALAHAD's DRQS method. See Section 11.3 for more details.
Constraint: Bxnl Nlls Method=POWELL-DOGLEG, AINT, MORE-SORENSEN, LINEAR SOLVER or GALAHAD.
Bxnl ModelaDefault =HYBRID
This parameter specifies which model is used to approximate the objective function and estimate the next point that reduces the error. This is one of the most important optional parameters and should be chosen according to the problem characteristics. The models are briefly described in Section 11.2.
Constraint: Bxnl Model=GAUSS-NEWTON, QUASI-NEWTON, HYBRID or TENSOR-NEWTON.
Bxnl Tn MethodaDefault =MIN-1-VAR
This parameter specifies how to solve the subproblem and find the next step sk for the tensor Newton model, Bxnl Model=TENSOR-NEWTON. The subproblems are solved using a range of regularization schemes. See Section 11.3.3.
Constraint: Bxnl Tn Method=IMPLICIT, MIN-1-VAR, MIN-NVAR, ADD-1-VAR or ADD-NVAR.
Bxnl Use Second DerivativesaDefault =NO
This parameter indicates whether the weighted sum of residual Hessians are available through the call-back lsqhes. If Bxnl Use Second Derivatives=NO and the specified model in Bxnl Model requires user-suppied second derivatives, then the solver will terminate with fail.code= NE_DERIV_ERRORS.
Constraint: Bxnl Use Second Derivatives=YES or NO.
Bxnl Use WeightsaDefault =NO
This parameter indicates whether to use a weighted nonlinear least square model. If Bxnl Use Weights=YES then the weights wi>0,i=1,,nres in (2) must be supplied by you via e04rxc. If weights are to be used, then all nres elements must be provided, see Section 9.2. If the solver is expecting to use weights but they are not provided or have non-positive values, then the solver will terminate with fail.code= NE_DC_MATCH.
Constraint: Bxnl Use Second Derivatives=YES or NO.
Bxnl Iteration LimitiDefault =1000
This parameter specifies the maximum amount of major iterations the solver is alloted. If this limit is reached, then the solver will terminate with fail.code= NE_TOO_MANY_ITER.
Constraint: Bxnl Iteration Limit1.
Bxnl Monitor FrequencyiDefault =0
If Bxnl Monitor Frequency>0, the user-supplied function monit will be called at the end of every ith step for monitoring purposes.
Constraint: Bxnl Monitor Frequency0.
Bxnl Print HeaderiDefault =30
This parameter defines, in number of iterations, the frequency with which to print the iteration log header.
Constraint: Bxnl Print Header1.
Infinite Bound SizerDefault =1020
This defines the ‘infinite’ bound bigbnd in the definition of the problem constraints. Any upper bound greater than or equal to bigbnd will be regarded as + (and similarly any lower bound less than or equal to -bigbnd will be regarded as -). Note that a modification of this optional parameter does not influence constraints which have already been defined; only the constraints formulated after the change will be affected.
Constraint: Infinite Bound Size1000.
Monitoring FileiDefault =−1
(See Section 3.1.1 in the Introduction to the NAG Library CL Interface for further information on NAG data types.)
If i0, the Nag_FileID number (as returned from x04acc) for the secondary (monitoring) output. If Monitoring File=−1, no secondary output is provided. The information output to this file ID is controlled by Monitoring Level.
Constraint: Monitoring File−1.
Monitoring LeveliDefault =4
This parameter sets the amount of information detail that will be printed by the solver to the secondary output. The meaning of the levels is the same as for Print Level.
Constraint: 0Monitoring Level5.
Print FileiDefault =Nag_FileID number associated with stdout
(See Section 3.1.1 in the Introduction to the NAG Library CL Interface for further information on NAG data types.)
If i0, the Nag_FileID number (as returned from x04acc, stdout as the default) for the primary output of the solver. If Print File=−1, the primary output is completely turned off independently of other settings. The information output to this unit is controlled by Print Level.
Constraint: Print File−1.
Print LeveliDefault =2
This parameter defines how detailed information should be printed by the solver to the primary and secondary output.
i Output
0 No output from the solver.
1 The Header and Summary.
2, 3, 4, 5 Additionally, the Iteration log.
Constraint: 0Print Level5.
Print OptionsaDefault =YES
If Print Options=YES, a listing of optional parameters will be printed to the primary output and is always printed to the secondary output.
Constraint: Print Options=YES or NO.
Print SolutionaDefault =NO
If Print Solution=X, the final values of the primal variables are printed on the primary and secondary outputs.
If Print Solution=YES or ALL, in addition to the primal variables, the final values of the dual variables are printed on the primary and secondary outputs.
Constraint: Print Solution=YES, NO, X or ALL.
Stats TimeaDefault =NO
This parameter turns on timing. This might be helpful for a choice of different solving approaches. It is possible to choose between CPU and wall clock time. Choice YES is equivalent to WALL CLOCK.
Constraint: Stats Time=YES, NO, CPU or WALL CLOCK.
Time LimitrDefault =106
A limit to the number of seconds that the solver can use to solve one problem. If at the end of an iteration this limit is exceeded, the solver will terminate with fail.code= NE_TIME_LIMIT.
Constraint: Time Limit>0.