Please note that the recommended version of Scilab is 2025.0.0. This page might be outdated.
See the recommended documentation of this function
optim
non-linear optimization routine
Calling Sequence
[f,xopt]=optim(costf,x0) [f [,xopt [,gradopt [,work]]]]=optim(costf [,<contr>],x0 [,algo] [,df0 [,mem]] [,work] [,<stop>] [,<params>] [,imp=iflag])
Arguments
- costf
external, i.e Scilab function list or string (
costf
is the cost function, that is, a Scilab script, a Fortran 77 routine or a C function.- x0
real vector (initial value of variable to be minimized).
- f
value of optimal cost (
f=costf(xopt)
)- xopt
best value of
x
found.- <contr>
keyword representing the following sequence of arguments:
'b',binf,bsup
withbinf
andbsup
are real vectors with same dimension asx0
.binf
andbsup
are lower and upper bounds onx
.- algo
'qn'
: quasi-Newton (this is the default solver)'gc'
: conjugate gradient'nd'
: non-differentiable.Note that the conjugate gradient solver does not accept bounds on
x
.
- df0
real scalar. Guessed decreasing of
f
at first iteration. (df0=1
is the default value).- mem :
integer, number of variables used to approximate the Hessian. Default value is 10. This feature is available for the Gradient-Conjugate algorithm "gc" without constraints and the non-smooth algorithm "nd" without constraints.
- <stop>
keyword representing the sequence of optional parameters controlling the convergence of the algorithm.
'ar',nap [,iter [,epsg [,epsf [,epsx]]]]
- "ar"
reserved keyword for stopping rule selection defined as follows:
- nap
maximum number of calls to
costf
allowed (default is 100).- iter
maximum number of iterations allowed (default is 100).
- epsg
threshold on gradient norm.
- epsf
threshold controlling decreasing of
f
- epsx
threshold controlling variation of
x
. This vector (possibly matrix) of same size asx0
can be used to scalex
.
- <params>
keyword representing the method to initialize the arguments
ti, td
passed to the objective function, provided as a C or Fortran routine. This option has no meaning when the cost function is a Scilab script. <params> can be set to only one of the following values."in"
That mode allows to allocate memory in the internal Scilab workspace so that the objective function can get arrays with the required size, but without directly allocating the memory. "in" stands for "initialization". In that mode, before the value and derivative of the objective function is to be computed, there is a dialog between the optim Scilab primitive and the objective function. In this dialog, the objective function is called two times, with particular values of the "ind" parameter. The first time, ind is set to 10 and the objective function is expected to set the nizs, nrzs and ndzs integer parameters of the "nird" common.
common /nird/ nizs,nrzs,ndzs
This allows Scilab to allocate memory inside its internal workspace. The second time the objective function is called, ind is set to 11 and the objective function is expected to set the ti, tr and tz arrays. After this initialization phase, each time it is called, the objective function is ensured that the ti, tr and tz arrays which are passed to it have the values that have been previously initialized.
"ti",valti
In this mode, valti is expected to be a Scilab vector variable containing integers. Whenever the objective function is called, the ti array it receives contains the values of the Scilab variable.
"td", valtd
In this mode, valtd is expected to be a Scilab vector variable containing double values. Whenever the objective function is called, the td array it receives contains the values of the Scilab variable.
"ti",valti,"td",valtd
This mode combines the two previous.
The
ti, td
arrays may be used so that the objective function can be computed. For example, if the objective function is a polynomial, the ti array may may be used to store the coefficients of that polynomial.Users should choose carefully between the "in" mode and the "ti" and "td" mode, depending on the fact that the arrays are Scilab variables or not. If the data is available as Scilab variables, then the "ti", valti, "td", valtd mode should be chosen. If the data is available directly from the objective function, the "in" mode should be chosen. Notice that there is no "tr" mode, since, in Scilab, all real values are of "double" type.
If neither the "in" mode, nor the "ti", "td" mode is chosen, that is, if <params> is not present as an option of the optim primitive, the user may should not assume that the ti,tr and td arrays can be used : reading or writing the arrays may generate unpredictable results.
- "imp=iflag"
named argument used to set the trace mode. The possible values for iflag are 0,1,2 and >2. Use this option with caution : most of these reports are written on the Scilab standard output.
iflag=0: nothing (except errors) is reported (this is the default),
iflag=1: initial and final reports,
iflag=2: adds a report per iteration,
iflag>2: add reports on linear search.
iflag<0: calls the cost function with ind=1 every -imp iterations.
- gradopt
gradient of
costf
atxopt
- work
working array for hot restart for quasi-Newton method. This array is automatically initialized by
optim
whenoptim
is invoked. It can be used as input parameter to speed-up the calculations.
Description
Non-linear optimization routine for programs without constraints or with bound constraints:
min costf(x) w.r.t x.
costf
is an "external" i.e a Scilab function, a
list or a string giving the name of a C or Fortran routine (see
"external"). This external must return the value f
of
the cost function at the point x
and the gradient
g
of the cost function at the point
x
.
- - Scilab function case
If
costf
is a Scilab function, the calling sequence forcostf
must be:[f,g,ind]=costf(x,ind)
Here,
costf
is a function which returnsf
, value (real number) of cost function atx
, andg
, gradient vector of cost function atx
. The variableind
is described below.- - List case
If
costf
is a list, it should be of the form:list(real_costf, arg1,...,argn)
withreal_costf
a Scilab function with calling sequence :[f,g,ind]=costf(x,ind,arg1,... argn)
. Thex
,f
,g
,ind
arguments have the same meaning that above.argi
arguments can be used to pass function parameters.- - String case
If
costf
is a character string, it refers to the name of a C or Fortran routine which must be linked to Scilab- * Fortran case
The interface of the Fortran subroutine computing the objective must be :
subroutine costf(ind,n,x,f,g,ti,tr,td)
with the following declarations:
The argument
ind
is described below.If ind = 2, 3 or 4, the inputs of the routine are :
x, ind, n, ti, tr,td
.If ind = 2, 3 or 4, the outputs of the routine are :
f
andg
.- * C case
The interface of the C function computing the objective must be :
The argument
ind
is described below.The inputs and outputs of the function are the same as in the fortran case.
The ind
input argument is a message sent from the solver to the
cost function so that the cost function knows what to compute.
If
ind=2
,costf
must providef
.If
ind=3
,costf
must provideg
.If
ind=4
,costf
must providef
andg
.If
ind=1
,costf
can print messages.
When the imp
option is set to a negative integer value, say
m
for example, the cost function is called back every -m
iterations, with the ind=1
parameter. This can be used to
print messages, fill a log file or creating an interactive plot.
See below for sample uses of this feature.
On output, ind<0
means that
f
cannot be evaluated at x
and
ind=0
interrupts the optimization.
Example: Scilab function
The following is an example with a Scilab function. Notice, for simplifications reasons, the Scilab function "cost" of the following example computes the objective function f and its derivative no matter of the value of ind. This allows to keep the example simple. In practical situations though, the computation of "f" and "g" may raise performances issues so that a direct optimization may be to use the value of "ind" to compute "f" and "g" only when needed.
// External function written in Scilab function [f, g, ind]=cost(x, ind) xref=[1;2;3]; f=0.5*norm(x-xref)^2; g=x-xref; endfunction // Simplest call x0=[1;-1;1]; [f,xopt]=optim(cost,x0) // By conjugate gradient - you can use 'qn', 'gc' or 'nd' [f,xopt,gopt]=optim(cost,x0,'gc') //Seen as non differentiable [f,xopt,gopt]=optim(cost,x0,'nd') // Upper and lower bounds on x [f,xopt,gopt]=optim(cost,'b',[-1;0;2],[0.5;1;4],x0) // Upper and lower bounds on x and setting up the algorithm to 'gc' [f,xopt,gopt]=optim(cost,'b',[-1;0;2],[0.5;1;4],x0,'gc') // Bound on the number of call to the objective function [f,xopt,gopt]=optim(cost,'b',[-1;0;2],[0.5;1;4],x0,'gc','ar',3) // Set max number of call to the objective function (3) // Set max number of iterations (100) // Set stopping threshold on the value of f (1e-6), // on the value of the norm of the gradient of the objective function (1e-6) // on the improvement on the parameters x_opt (1e-6;1e-6;1e-6) [f,xopt,gopt]=optim(cost,'b',[-1;0;2],[0.5;1;4],x0,'gc','ar',3,100,1e-6,1e-6,[1e-3;1e-3;1e-3]) // Additionnal messages are printed in the console. [f,xopt]=optim(cost,x0,imp=3)
Example: C function
The following is an example with a C function, where a C source code is written into a file, dynamically compiled and loaded into Scilab, and then used by the "optim" solver. The interface of the "rosenc" function is fixed, even if the arguments are not really used in the cost function. This is because the underlying optimization solvers must assume that the objective function has a known, constant interface. In the following example, the arrays ti and tr are not used, only the array "td" is used, as a parameter of the Rosenbrock function. Notice that the content of the arrays ti and td are the same that the content of the Scilab variable, as expected.
// External function written in C (C compiler required) // write down the C code (Rosenbrock problem) C=['#include <math.h>' 'double sq(double x)' '{ return x*x;}' 'void rosenc(int *ind, int *n, double *x, double *f, double *g, ' ' int *ti, float *tr, double *td)' '{' ' double p;' ' int i;' ' p=td[0];' ' if (*ind==2||*ind==4) {' ' *f=1.0;' ' for (i=1;i<*n;i++)' ' *f+=p*sq(x[i]-sq(x[i-1]))+sq(1.0-x[i]);' ' }' ' if (*ind==3||*ind==4) {' ' g[0]=-4.0*p*(x[1]-sq(x[0]))*x[0];' ' for (i=1;i<*n-1;i++)' ' g[i]=2.0*p*(x[i]-sq(x[i-1]))-4.0*p*(x[i+1]-sq(x[i]))*x[i]-2.0*(1.0-x[i]);' ' g[*n-1]=2.0*p*(x[*n-1]-sq(x[*n-2]))-2.0*(1.0-x[*n-1]);' ' }' '}']; cd TMPDIR; mputl(C,TMPDIR+'/rosenc.c') // compile the C code l=ilib_for_link('rosenc','rosenc.c',[],'c'); // incremental linking link(l,'rosenc','c') //solve the problem x0=[40;10;50]; p=100; [f,xo,go]=optim('rosenc',x0,'td',p)
Example: Fortran function
The following is an example with a Fortran function.
// External function written in Fortran (Fortran compiler required) // write down the Fortran code (Rosenbrock problem) F=[ ' subroutine rosenf(ind, n, x, f, g, ti, tr, td)' ' integer ind,n,ti(*)' ' double precision x(n),f,g(n),td(*)' ' real tr(*)' 'c' ' double precision y,p' ' p=td(1)' ' if (ind.eq.2.or.ind.eq.4) then' ' f=1.0d0' ' do i=2,n' ' f=f+p*(x(i)-x(i-1)**2)**2+(1.0d0-x(i))**2' ' enddo' ' endif' ' if (ind.eq.3.or.ind.eq.4) then' ' g(1)=-4.0d0*p*(x(2)-x(1)**2)*x(1)' ' if(n.gt.2) then' ' do i=2,n-1' ' g(i)=2.0d0*p*(x(i)-x(i-1)**2)-4.0d0*p*(x(i+1)-x(i)**2)*x(i)' ' & -2.0d0*(1.0d0-x(i))' ' enddo' ' endif' ' g(n)=2.0d0*p*(x(n)-x(n-1)**2)-2.0d0*(1.0d0-x(n))' ' endif' ' return' ' end']; cd TMPDIR; mputl(F,TMPDIR+'/rosenf.f') // compile the Fortran code l=ilib_for_link('rosenf','rosenf.f',[],'f'); // incremental linking link(l,'rosenf','f') //solve the problem x0=[40;10;50]; p=100; [f,xo,go]=optim('rosenf',x0,'td',p)
Example: Fortran function with initialization
The following is an example with a Fortran function in which the "in" option is used to allocate memory inside the Scilab environment. In this mode, there is a dialog between Scilab and the objective function. The goal of this dialog is to initialize the parameters of the objective function. Each part of this dialog is based on a specific value of the "ind" parameter.
At the beginning, Scilab calls the objective function, with the ind parameter equals to 10. This tells the objective function to initialize the sizes of the arrays it needs by setting the nizs, nrzs and ndzs integer parameters of the "nird" common. Then the objective function returns. At this point, Scilab creates internal variables and allocate memory for the variable izs, rzs and dzs. Scilab calls the objective function back again, this time with ind equals to 11. This tells the objective function to initialize the arrays izs, rzs and dzs. When the objective function has done so, it returns. Then Scilab enters in the real optimization mode and calls the optimization solver the user requested. Whenever the objective function is called, the izs, rzs and dzs arrays have the values that have been previously initialized.
// // Define a fortran source code and compile it (fortran compiler required) // fortransource=[' subroutine rosenf(ind,n,x,f,g,izs,rzs,dzs)' 'C -------------------------------------------' 'c Example of cost function given by a subroutine' 'c if n<=2 returns ind=0' 'c f.bonnans, oct 86' ' implicit double precision (a-h,o-z)' ' real rzs(1)' ' double precision dzs(*)' ' dimension x(n),g(n),izs(*)' ' common/nird/nizs,nrzs,ndzs' ' if (n.lt.3) then' ' ind=0' ' return' ' endif' ' if(ind.eq.10) then' ' nizs=2' ' nrzs=1' ' ndzs=2' ' return' ' endif' ' if(ind.eq.11) then' ' izs(1)=5' ' izs(2)=10' ' dzs(2)=100.0d+0' ' return' ' endif' ' if(ind.eq.2)go to 5' ' if(ind.eq.3)go to 20' ' if(ind.eq.4)go to 5' ' ind=-1' ' return' '5 f=1.0d+0' ' do 10 i=2,n' ' im1=i-1' '10 f=f + dzs(2)*(x(i)-x(im1)**2)**2 + (1.0d+0-x(i))**2' ' if(ind.eq.2)return' '20 g(1)=-4.0d+0*dzs(2)*(x(2)-x(1)**2)*x(1)' ' nm1=n-1' ' do 30 i=2,nm1' ' im1=i-1' ' ip1=i+1' ' g(i)=2.0d+0*dzs(2)*(x(i)-x(im1)**2)' '30 g(i)=g(i) -4.0d+0*dzs(2)*(x(ip1)-x(i)**2)*x(i) - ' ' & 2.0d+0*(1.0d+0-x(i))' ' g(n)=2.0d+0*dzs(2)*(x(n)-x(nm1)**2) - 2.0d+0*(1.0d+0-x(n))' ' return' ' end']; cd TMPDIR; mputl(fortransource,TMPDIR+'/rosenf.f') // compile the C code libpath=ilib_for_link('rosenf','rosenf.f',[],'f'); // incremental linking linkid=link(libpath,'rosenf','f'); x0=1.2*ones(1,5); // // Solve the problem // [f,x,g]=optim('rosenf',x0,'in');
Example: Fortran function with initialization on Windows with Intel Fortran Compiler
Under the Windows operating system with Intel Fortran Compiler, one must carefully design the fortran source code so that the dynamic link works properly. On Scilab's side, the optimization component is dynamically linked and the symbol "nird" is exported out of the optimization dll. On the cost function's side, which is also dynamically linked, the "nird" common must be imported in the cost function dll.
The following example is a re-writing of the previous example, with special attention for the Windows operating system with Intel Fortran compiler as example. In that case, we introduce additionnal compiling instructions, which allows the compiler to import the "nird" symbol.
fortransource=['subroutine rosenf(ind,n,x,f,g,izs,rzs,dzs)' 'cDEC$ IF DEFINED (FORDLL)' 'cDEC$ ATTRIBUTES DLLIMPORT:: /nird/' 'cDEC$ ENDIF' 'C -------------------------------------------' 'c Example of cost function given by a subroutine' 'c if n<=2 returns ind=0' 'c f.bonnans, oct 86' ' implicit double precision (a-h,o-z)' [etc...]
Example: Logging features
The imp flag may take negative integer values, say k. In that case, the cost function is called once every -k iterations. This allows to draw the function value or write a log file.
In the following example, we solve the Rosenbrock test case. For each iteration of the algorithm, we print the value of x, f and g.
function [f, g, ind]=cost(x, ind) xref=[1;2;3]; if ( ind == 1 | ind == 4 ) then f=0.5*norm(x-xref)^2; end if ( ind == 1 | ind == 4 ) then g=x-xref; end if ( ind == 1 ) then mprintf("===========\n") mprintf("x = %s\n", strcat(string(x)," ")) mprintf("f = %e\n", f) g=x-xref; mprintf("g = %s\n", strcat(string(g)," ")) end endfunction x0=[1;-1;1]; [f,xopt]=optim(cost,x0,imp=-1)
The previous script produces the following output.
=========== x = 1 -1 1 f = 6.500000e+000 g = 0 -3 -2 =========== x = 1 0 1.6666667 f = 2.888889e+000 g = 0 -2 -1.3333333 =========== x = 1 2 3 f = 9.860761e-031 g = 0 -4.441D-16 1.332D-15 =========== x = 1 2 3 f = 0.000000e+000 g = 0 0 0
In the following example, we solve the Rosenbrock test case. For each iteration of the algorithm, we plot the current value of x into a 2D graph containing the contours of Rosenbrock's function. This allows to see the progress of the algorithm while the algorithm is performing. We could as well write the value of x, f and g into a log file if needed.
// 1. Define rosenbrock for contouring function f=rosenbrockC(x1, x2) x = [x1 x2]; f = 100.0 *(x(2)-x(1)^2)^2 + (1-x(1))^2; endfunction x0 = [-1.2 1.0]; xopt = [1.0 1.0]; // 2. Draw the contour of Rosenbrock's function xdata = linspace(-2,2,100); ydata = linspace(-2,2,100); contour ( xdata , ydata , rosenbrockC , [1 10 100 500 1000]) plot(x0(1) , x0(2) , "b.") plot(xopt(1) , xopt(2) , "r*") // 3. Define Rosenbrock for optimization function [f, g, ind]=rosenbrock(x, ind) if ((ind == 1) | (ind == 4)) then f = 100.0 *(x(2)-x(1)^2)^2 + (1-x(1))^2; end if ((ind == 1) | (ind == 4)) then g(1) = - 400. * ( x(2) - x(1)**2 ) * x(1) -2. * ( 1. - x(1) ) g(2) = 200. * ( x(2) - x(1)**2 ) end if (ind == 1) then plot ( x(1) , x(2) , "g." ) end endfunction // 4. Plot the optimization process, during optimization [ fopt , xopt ] = optim ( rosenbrock , x0 , imp = -1)
Example: Optimizing with numerical derivatives
It is possible to optimize a problem without an explicit knowledge of the derivative of the cost function. For this purpose, we can use the numdiff or derivative function to compute a numerical derivative of the cost function.
In the following example, we use the numdiff function to solve Rosenbrock's problem.
function f=rosenbrock(x) f = 100.0 *(x(2)-x(1)^2)^2 + (1-x(1))^2; endfunction function [f, g, ind]=rosenbrockCost(x, ind) if ((ind == 1) | (ind == 4)) then f = rosenbrock ( x ); end if ((ind == 1) | (ind == 4)) then g= numdiff ( rosenbrock , x ); end endfunction x0 = [-1.2 1.0]; [ fopt , xopt ] = optim ( rosenbrockCost , x0 )
In the following example, we use the derivative function to solve Rosenbrock's problem. Given that the step computation strategy is not the same in numdiff and derivative, this might lead to improved results.
function f=rosenbrock(x) f = 100.0 *(x(2)-x(1)^2)^2 + (1-x(1))^2; endfunction function [f, g, ind]=rosenbrockCost2(x, ind) if ((ind == 1) | (ind == 4)) then f = rosenbrock ( x ); end if ((ind == 1) | (ind == 4)) then g= derivative ( rosenbrock , x.' , order = 4 ); end endfunction x0 = [-1.2 1.0]; [ fopt , xopt ] = optim ( rosenbrockCost2 , x0 )
Example: Counting function evaluations and number of iterations
The imp
option can take negative values.
If the imp
is equal to m
where
m
is a negative integer, then the cost function is
evaluated every -m
iterations, with the ind
input argument equal to 1.
The following example uses this feature to compute the number of iterations.
The global variable mydata
is used to store the
number of function evaluations as well as the number of iterations.
xref=[1;2;3]; x0=[1;-1;1]; global _MYDATA_ _MYDATA_ = tlist ( ["T_MYDATA","niter","nfevals"]) _MYDATA_.niter = 0 _MYDATA_.nfevals = 0 function [f, g, ind]=cost(x, ind) global _MYDATA_ disp(ind) if ( ind == 1 ) _MYDATA_.niter = _MYDATA_.niter + 1 end _MYDATA_.nfevals = _MYDATA_.nfevals + 1 f=0.5*norm(x-xref)^2; g=x-xref; endfunction [f,xopt]=optim(cost,x0,imp=-1) mprintf ( "Number of function evaluations:%d\n",_MYDATA_.nfevals) mprintf ( "Number of iterations:%d\n",_MYDATA_.niter)
While the previous example perfectly works, there is a risk that
the same variable _MYDATA_
is used by some internal function used by optim
.
In this case, the value may be wrong. This is why a sufficiently weird variable
name has been used.
Example : Passing extra parameters
In most practical situations, the cost function depends on extra parameters which are required to evaluate the cost function. There are several methods to achieve this goal.
In the following example, the cost function uses 4 parameters a, b, c
and
d
.
We define the cost function with additionnal input arguments,
which are declared after the index argument.
Then we pass a list as the first input argument of the optim
solver. The first element of the list is the cost function. The additionnal variables
are directly passed to the cost function.
function [f, g, ind]=costfunction(x, ind, a, b, c, d) f = a * ( x(1) - c ) ^2 + b * ( x(2) - d )^2 g(1) = 2 * a * ( x(1) - c ) g(2) = 2 * b * ( x(2) - d ) endfunction x0 = [1 1]; a = 1.0; b = 2.0; c = 3.0; d = 4.0; costf = list ( costfunction , a , b , c, d ); [ fopt , xopt ] = optim ( costf , x0 , imp = 2 )
In complex cases, the cost function may have so many parameters, that having a function which takes all arguments as inputs is not convenient. For example, consider the situation where the cost function needs 12 parameters. Then, designing a function with 14 input arguments (x, index and the 12 parameters) is difficult to manage. Instead, we can use a more complex data structure to store our data. In the following example, we use a tlist to store the 4 input arguments. This method can easily be expanded to an arbitrary number of parameters.
function [f, g, ind]=costfunction(x, ind, parameters) // Get the parameters a = parameters.a b = parameters.b c = parameters.c d = parameters.d f = a * ( x(1) - c ) ^2 + b * ( x(2) - d )^2 g(1) = 2 * a * ( x(1) - c ) g(2) = 2 * b * ( x(2) - d ) endfunction x0 = [1 1]; a = 1.0; b = 2.0; c = 3.0; d = 4.0; // Store the parameters parameters = tlist ( [ "T_MYPARAMS" "a" "b" "c" "d" ]); parameters.a = a; parameters.b = b; parameters.c = c; parameters.d = d; costf = list ( costfunction , parameters ); [ fopt , xopt ] = optim ( costf , x0 , imp = 2 )
In the following example, the parameters are defined before the optimizer is called. They are directly used in the cost function.
// The example NOT to follow function [f, g, ind]=costfunction(x, ind) f = a * ( x(1) - c ) ^2 + b * ( x(2) - d )^2 g(1) = 2 * a * ( x(1) - c ) g(2) = 2 * b * ( x(2) - d ) endfunction x0 = [1 1]; a = 1.0; b = 2.0; c = 3.0; d = 4.0; [ fopt , xopt ] = optim ( costfunction , x0 , imp = 2 )
While the previous example perfectly works, there is a risk that
the same variables are used by some internal function used by optim
.
In this case, the value of the parameters are not what is expected and the
optimization can fail or, worse, give a wrong result.
It is also difficult to manage such a function, which requires that all the parameters
are defined in the calling context.
In the following example, we define the cost function with the classical header.
Inside the function definition, we declare that the parameters a, b, c
and
d
are global variables.
Then we declare and set the global variables.
// Another example NOT to follow function [f, g, ind]=costfunction(x, ind) global a b c d f = a * ( x(1) - c ) ^2 + b * ( x(2) - d )^2 g(1) = 2 * a * ( x(1) - c ) g(2) = 2 * b * ( x(2) - d ) endfunction x0 = [1 1]; global a b c d a = 1.0; b = 2.0; c = 3.0; d = 4.0; [ fopt , xopt ] = optim ( costfunction , x0 , imp = 2 )
While the previous example perfectly works, there is a risk that
the same variables are used by some internal function used by optim
.
In this case, the value of the parameters are not what is expected and the
optimization can fail or, worse, give a wrong result.
Example : Checking that derivatives are correct
Many optimization problem can be avoided if the derivatives are computed correctly. One common reason for failure in the step-length procedure is an error in the calculation of the cost function and its gradient. Incorrect calculation of derivatives is by far the most common user error.
In the following example, we give a false implementation of Rosenbrock's
gradient. In order to check the computation of the derivatives,
we use the derivative
function. We define the
simplified
function, which delegates the computation of f
to the rosenbrock function. The simplified
function is passed
as an input argument of the derivative
function.
function [f, g, index]=rosenbrock(x, index) f = 100.0 *(x(2)-x(1)^2)^2 + (1-x(1))^2; // Exact : g(1) = - 400. * ( x(2) - x(1)**2 ) * x(1) -2. * ( 1. - x(1) ) // Wrong : g(1) = - 1200. * ( x(2) - x(1)**2 ) * x(1) -2. * ( 1. - x(1) ) g(2) = 200. * ( x(2) - x(1)**2 ) endfunction function f=simplified(x) index = 1; [ f , g , index ] = rosenbrock ( x , index ) endfunction x0 = [-1.2 1]; index = 1; [ f , g , index ] = rosenbrock ( x0 , index ); gnd = derivative ( simplified , x0.' ); mprintf ( "Exact derivative:[%s]\n" , strcat ( string(g) , " " )); mprintf ( "Numerical derivative:[%s]\n" , strcat ( string(gnd) , " " ));
The previous script produces the following output. Obviously, the difference between the two gradient is enormous, which shows that the wrong formula has been used in the gradient.
Exact derivative:[-638 -88] Numerical derivative:[-215.6 -88]
References
The following is a map from the various options to the underlying solvers, with some comments about the algorithm, when available.
- "qn" without constraints
n1qn1 : a quasi-Newton method with a Wolfe-type line search
- "qn" with bounds constraints
qnbd : a quasi-Newton method with projection
RR-0242 - A variant of a projected variable metric method for bound constrained optimization problems, Bonnans Frederic, Rapport de recherche de l'INRIA - Rocquencourt, Octobre 1983
- "gc" without constraints
n1qn3 : a conjugate gradient method with BFGS.
- "gc" with bounds constraints
gcbd : a BFGS-type method with limited memory and projection
- "nd" without constraints
n1fc1 : a bundle method
- "nd" with bounds constraints
not available
Author
The Modulopt library : J.Frederic Bonnans, Jean-Charles Gilbert, Claude Lemarechal
The interfaces to the Modulopt library : J.Frederic Bonnans
This help : Michael Baudin
<< numdiff | Optimization and Simulation | qld >> |