Multistart

Nonlinear optimization problems are often nonconvex due to the objective function, constraint functions, or both. When this is true, there may be many points that satisfy the local optimality conditions. Default Knitro behavior is to return the first locally optimal point found. Knitro offers a simple multi-start feature that searches for a better optimal point by restarting Knitro from different initial points. The feature is enabled by setting ms_enable = 1.

Note

In many cases the user would like to obtain the global optimum to the optimization problem; that is, the local optimum with the very best objective function value. Knitro cannot guarantee that multi-start will find the global optimum. In general, the global optimum can only be found with special knowledge of the objective and constraint functions; for example, the functions may need to be bounded by other piece-wise convex functions. Knitro executes with very little information about functional form. Although no guarantee can be made, the probability of finding a better local solution improves if more start points are tried.

Multistart algorithm

The multi-start procedure generates new start points by randomly selecting components of x that satisfy lower and upper bounds on the variables. Knitro finds a local optimum from each start point using the same problem definition and user options. The final solution returned from KTR_solve() is the local optimum with the best objective function value if any local optima have been found. If no local optimum has been found, Knitro will return the best feasible solution estimate it found. If no feasible solution estimate has been found, Knitro will return the least infeasible point.

Parallel multistart

The multistart procedure can run in parallel on shared memory multi-processor machines by setting par_numthreads greater than 1. See Parallelism for more details on controlling parallel performance in Knitro.

When the multistart procedure is run in parallel, Knitro will produce the same sequence of initial points and solves that you see when running multistart sequentially (though, perhaps, not in the same order).

Therefore, as long as you run multistart to completion (ms_terminate =0) and use the deterministic option (ms_deterministic =1), you should visit the same initial points encountered when running multistart sequentially, and get the same final solution. By default ms_terminate =0 and ms_deterministic =1 so that the parallel multistart produces the same solution as the sequential multistart.

However, if ms_deterministic =0, or ms_terminate >0, there is no guarantee that the final solution reported by multistart will be the same when run in parallel compared to the solution when run sequentially, and even the parallel solution may change when run at different times.

The option par_msnumthreads can be used to set the number of threads used by the multistart procedure. For instance, if par_numthreads =16 and par_msnumthreads =8, Knitro will run 8 solves in parallel and each solve will be allocated 2 threads.

Multistart output

For multistart, you can have output from each local solve written to a file named knitro_ms_x.log where “x” is the solve number by setting the option ms_outsub=1.

Multistart options

The multi-start option is convenient for conducting a simple search for a better solution point. Search time is improved if the variable bounds are made as tight as possible, confining the search to a region where a good solution is likely to be found. The user can restrict the multi-start search region without altering bounds by using the options ms_maxbndrange and ms_startptrange. The other multi-start options are the following.

Option Meaning
ms_deterministic Control whether to use deterministic multistart
ms_enable Enable multistart
ms_maxbndrange Maximum unbounded variable range for multistart
ms_maxsolves Maximum Knitro solves for multistart
ms_maxtime_cpu Maximum CPU time for multistart, in seconds
ms_maxtime_real Maximum real time for multistart, in seconds
ms_num_to_save Feasible points to save from multistart
ms_outsub Can write each solve to a file (parallel only)
ms_savetol Tol for feasible points being equal
ms_seed Initial seed for generating random start points
ms_startptrange Maximum variable range for multistart
ms_terminate Termination condition for multistart

The number of start points tried by multi-start is specified with the option ms_maxsolves. By default, Knitro will try min(200, 10*n), where n is the number of variables in the problem. Users may override the default by setting ms_maxsolves to a specific value.

The ms_maxbndrange option applies to variables unbounded in at least one direction (i.e., the upper or lower bound, or both, is infinite) and keeps new start points within a total range equal to the value of ms_maxbndrange. The ms_startptrange option applies to all variables and keeps new start points within a total range equal to the value of ms_startptrange, overruling ms_maxbndrange if it is a tighter bound. In general, use ms_startptrange to limit the multi-start search only if the initial start point supplied by the user is known to be the center of a desired search area. Use ms_maxbndrange as a surrogate bound to limit the multi-start search when a variable is unbounded.

The ms_num_to_save option allows a specific number of distinct feasible points to be saved in a file named knitro_mspoints.log. Each point results from a Knitro solve from a different starting point, and must satisfy the absolute and relative feasibility tolerances. Different start points may return the same feasible point, and the file contains only distinct points. The option ms_savetol determines that two points are distinct if their objectives or any solution components (including Lagrange multipliers) are separated by more than the value of ms_savetol using a relative tolerance test. More specifically, two values x and y are considered distinct if:

|x-y| \geq  \max(1,|x|,|y|) * \mbox{ms\_savetol}.

The file stores points in order from best objective to worst. If objectives are the same (as defined by ms_savetol), then points are ordered from smallest feasibility error to largest. The file can be read manually, but conforms to a fixed property/value format for machine reading.

Instead of using multi-start to search for a global solution, a user may want to use multi-start as a mechanism for finding any locally optimal or feasible solution estimate of a nonconvex problem and terminate as soon as one such point is found. The ms_terminate option, provides the user more control over when to terminate the multi-start procedure.

If ms_terminate = optimal the multi-start procedure will stop as soon as the first locally optimal solution is found or after ms_maxsolves – whichever comes first. If ms_terminate = feasible the multi-start procedure will instead stop as soon as the first feasible solution estimate is found or after ms_maxsolves – whichever comes first. If ms_terminate = maxsolves, it will only terminate after ms_maxsolves.

The option ms_seed can be used to change the seed used to generate the random initial points for multistart.

Multistart callbacks

The multistart procedure provides two callback utilities for the callable library API.

int  KNITRO_API KTR_set_ms_process_callback (KTR_context_ptr       kc,
                                             KTR_callback * const  fnPtr);

int  KNITRO_API KTR_set_ms_initpt_callback (KTR_context_ptr                 kc,
                                            KTR_ms_initpt_callback * const  fnPtr);

The first callback can be used to perform some user task after each multistart solve and is set by calling KTR_set_ms_process_callback(). You can use the second callback to specify your own initial points for multistart instead of using the randomly generated Knitro initial points. This callback function can be set through the function KTR_set_ms_initpt_callback().

See the Knitro API section in the Reference Manual for details on setting these callback functions and the prototypes for these callback functions.

In the object-oriented interface, the following functions are used to set the callbacks:

void KTRProblem::setMSProcessCallback(KTRMSProcessCallback* MSProcessCallback);

void KTRProblem::setMSInitPtCallback(KTRMSInitptCallback* MSInitPtCallback);

See the Object-oriented interface reference section for details on setting these callback functions in the object-oriented interface.

AMPL example

Let us consider again our AMPL example from Section Getting started with AMPL and run it with a different set of options:

1
2
3
4
5
ampl: reset;
ampl: option solver knitroampl;
ampl: option knitro_options "ms_enable=1 ms_num_to_save=5 ms_savetol=0.01";
ampl: model testproblem.mod;
ampl: solve;

The Knitro log printed on screen changes to reflect the results of the many solver runs that were made during the multistart procedure, and the very end of this log reads:

Multistart stopping, reached ms_maxsolves limit.

MULTISTART: Best locally optimal point is returned.
EXIT: Locally optimal solution found.

Final Statistics
----------------
Final objective value               =   9.35999999745429e+02
Final feasibility error (abs / rel) =   1.44e-07 / 3.83e-10
Final optimality error  (abs / rel) =   6.48e-07 / 4.28e-08
# of iterations                     =        415
# of CG iterations                  =         90
# of function evaluations           =        545
# of gradient evaluations           =        475
# of Hessian evaluations            =        422
Total program time (secs)           =       0.02660 (     0.027 CPU time)

===============================================================================

Knitro 10.0.0: Locally optimal or satisfactory solution.
objective 935.9999997; feasibility error 1.44e-07
415 iterations; 545 function evaluations

which shows that many more functions calls were made than without multistart. A file knitro_mspoints.txt was also created, whose content reads:

// Knitro 10.0.0 Multi-start Repository for feasible points.
// Each point contains information about the problem and the point.
// Points are sorted by objective value, from best to worst.


// Next feasible point.
numVars = 3
numCons = 2
objGoal = MINIMIZE
obj =   9.3600000342420878e+02
knitroStatus = 0
localSolveNumber = 1
feasibleErrorAbsolute =   0.00e+00
feasibleErrorRelative =   0.00e+00
optimalityErrorAbsolute =   2.25e-07
optimalityErrorRelative =   1.41e-08
x[0] =   2.0511214409048425e-07
x[1] =   4.1077619358921463e-08
x[2] =   7.9999996834308824e+00
lambda[0] =  -4.5247620510168322e-08
lambda[1] =   2.2857143915699769e+00
lambda[2] =  -1.0285715141992103e+01
lambda[3] =  -3.2000001143071813e+01
lambda[4] =  -2.1985040913238130e-07

// Next feasible point.
numVars = 3
numCons = 2
objGoal = MINIMIZE
obj =   9.5100000269458542e+02
knitroStatus = 0
localSolveNumber = 2
feasibleErrorAbsolute =   0.00e+00
feasibleErrorRelative =   0.00e+00
optimalityErrorAbsolute =   3.67e-07
optimalityErrorRelative =   2.62e-08
x[0] =   6.9999996377946481e+00
x[1] =   7.4479065893720198e-08
x[2] =   2.6499084231411754e-07
lambda[0] =  -6.3891336872934633e-08
lambda[1] =   1.7500001368019027e+00
lambda[2] =  -2.1791026695882249e-07
lambda[3] =  -1.7500002055167382e+01
lambda[4] =  -5.2500010586300956e+00

In addition to the known solution with value 936 that we had already found with a single solver run, we discover another local minimum with value 951 that we had never seen before. In this case, the new solution is not as good as the first one, but sometimes running the multistart algorithm significantly improves the objective function value with respect to single-run optimization.

MATLAB example

In order to run the multistart algorithm in MATLAB, we must pass the relevant set of options to Knitro via the Knitro options file. Let us create a simple text file named knitro.opt with the following content:

ms_enable 1
ms_num_to_save 5
ms_savetol 0.01
hessopt 2

(the last line hessopt 2 is necessary to remind Knitro to use approximate second derivatives, since we are not providing the exact hessian). Then let us run our MATLAB example from Section MATLAB example again, where the call to knitromatlab has been replaced with:

knitromatlab(@obj, x0, A, b, Aeq, beq, lb, ub, @nlcon, [], options, 'knitro.opt');

and where the knitro.opt file was placed in the current directory so that MATLAB can find it. The Knitro log looks simlar to what we observed with AMPL.

C example

The C example can also be easily modified to enable multistart by adding the following lines before the call to KTR_solve():

// multistart
if (KTR_set_int_param_by_name (kc, "ms_enable", 1) != 0)
exit( -1 );
if (KTR_set_int_param_by_name (kc, "ms_num_to_save", 5) != 0)
exit( -1 );
if (KTR_set_double_param_by_name (kc, "ms_savetol", 0.01) != 0)
exit( -1 );

Again, running this example we get a Knitro log that looks similar to what we observed with AMPL.

Object-oriented example

The object-oriented example can be modified to enable multistart by adding the following lines before the call to solver.solve():

// multistart
solver.setParam("ms_enable", 1);
solver.setParam("ms_num_to_save", 5);
solver.setParam("ms_savetol", 0.01);

Again, running this example we get a Knitro log that looks similar to the trace obtained with AMPL.