Re: Question about fmins!

[Michael Creel]

Hmm, I don't see any problem. I put the following in a file named test.m

Contents of test.m:
-------- cut here -------------
1;
function a = f(x)
        a = cos(x) + 0.01*(x-pi)^2;
endfunction

# Set up the controls:
ub = 20;
lb = -ub;
nt = 20;
ns = 5;
rt = 0.5;
maxevals = 1e10;
neps = 5;
functol = 1e-10;
paramtol = 1e-5;
verbosity = 1;
minarg = 1;

# Put them in a cell array:
control = {lb, ub, nt, ns, rt, maxevals, neps, functol, paramtol, verbosity, 
minarg};

# Call the minimizer (function argument also in cell array):
samin("f", {-8}, control)

-------- cut here -------------

Then I run it from octave, and I get

octave:2> test

================================================
SAMIN results

Normal convergence

Convergence tolerances: Func. tol. 1.000000e-10 Param. tol. 1.000000e-05
Obj. fn. value -1.000000

           parameter        search width
            3.141593            0.000009
================================================
ans =  3.1416
octave:3>

How did you obtain samin? If you think you might have an old version, please get 
the newest one from 
http://octave.cvs.sourceforge.net/*checkout*/octave/octave-forge/main/optim/src/samin.cc?revision=1.8
and compile it using "mkoctfile samin.cc"
Cheers, M.

rambovn wrote:
> Thank you very much for your help. I read about samin and tried the example
> that Octave provided, but it appeared error :O:O. Here is the example:
>
> function a = f(x)
>         a = cos(x) + 0.01*(x-pi)^2;
> endfunction
>
> Set up the controls:
> ub = 20;
> lb = -ub;
> nt = 20;
> ns = 5;
> rt = 0.5;
> maxevals = 1e10;
> neps = 5;
> functol = 1e-10;
> paramtol = 1e-5;
> verbosity = 2;
> minarg = 1;
>
> Put them in a cell array:
> control = {lb, ub, nt, ns, rt, maxevals,
>         neps, functol, paramtol, verbosity, minarg};
>
> Call the minimizer (function argument also in cell array):
> samin("f", {-8}, control)
>
> error: samin: 1st element of controls must be LB: a vector of lower bounds
>
> Do you know why? :-/
> Thank you for your patience.
> Regards
> Anh
>
>
> I ran across some Python bindings to the Fortran lbfgsb code by Nocedal and 
> co-authors (http://www.ece.northwestern.edu/~nocedal/lbfgsb.html), at 
> http://setiathome.ssl.berkeley.edu/~aparsons/python/aipy.lbfgsb.html
>
> As far as I know, lbfgsb is about the state-of-the-art for minimization of 
> differentiable functions of high-dimensional parameters, subject to bounds. 
> Maybe someone who has experience with Octave/Fortran would like to try to
> make 
> some bindings?
>
> Otherwise, samin in octave-forge does bounds, without using derivatives.
> It's a 
> stochastic method. In spite of that, it is pretty quick for many problems. 
> Etienne's suggestion is probably the first thing to try, though.
> M.
>
> /quote>