Thread Subject: fminsearch and normalization

afaik scaling possibility is absent in fminsearch.
try using ralg from openopt, in my examples it works better
http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objectId=13115&objectType=file
prob=ooAssign(objFun, x0, <optional params>)
prob.useScaling=1;
if no prob.ScaleFactor is provided -> x0 will be used as scale factor
r=ooRun(prob, @ralg);
x=r.xf;
fval = r.ff;

WBR, D.

On Mar 26, 6:55 pm, Peter Bone <peterb...@hotmail.com> wrote:
> I'm using fminsearch to find the optimum variable values that give a minimum in my error function. My variables currently have different ranges. For example, some are between 0 and 1 and some are between 100 and 1000. Would the fminsearch function work better if I normalize all my variables so that they're in the same range or doesn't it matter? Thanks.

What are the problems for testing it by yourself? You could already
had tried it many times.

let for example scf=[0.1 1 100 0.1] (you can just set it to your x0 or
typicalX; if zeros are present in x0 - replace them to for example
min(coords of x0 that are not equal to zero)), than try using
[x_ fval] = fminsearch(myobjfun2, x0 ./ scf);
x=x_ .* scf
------------
function r=myobjfun2(x)
global scf % or pass it as 2nd param of myobjfun2
x2=scf .* x
r= myobjfun(x)

btw openopt ralg enables all types of constraints and (sub)gradient
provided by user to objfun and/or constraints.

WBR, D.

On Mar 27, 11:34 am, Peter Bone <peterb...@hotmail.com> wrote:
> Thanks. I know that the fminsearch function doesn't scale the values automatically. My question is; will scaling them myself help the solver find the optimum solution?
> I have to use fminsearch by the way.

sorry, in
myobjfun2
there should be of course
r= myobjfun(x2)
not
r= myobjfun(x)
D.