Re: NMinimize fails to find a minimum value
- To: mathgroup at smc.vnet.net
- Subject: [mg120805] Re: NMinimize fails to find a minimum value
- From: Frank K <fkampas at gmail.com>
- Date: Thu, 11 Aug 2011 05:11:23 -0400 (EDT)
- Delivered-to: l-mathgroup@mail-archive0.wolfram.com
- References: <j1im1r$cmu$1@smc.vnet.net>
On Aug 6, 2:13 am, Hovhannes Babayan <bhovhan... at gmail.com> wrote:
> Hello,
>
> I am solving the following task. I have set of points (list of pair of
> X, Y coordinates) and list of weights for the each point. If for some
> point weight=1, that point is very important of weight=0, that point
> is not important. Weight ranges from 0 to 1.
> The task is to construct a curve, which will fit given list of points
> taking weights into account. The analytic form of function is known,
> but I have to find parameters z,u,a,b, which will give the best
> fitting curve.
> To do that, I am using NMinimize to minimize the following expression:
> Sum[ ((f[x[[i]],z,u,a,b]-y[[i]])^2)*weights[[i]], {i=
, n} ]
> where "n" is a number of points, "x" and "y" are point coordinates,
> and "f(r, z,u,a,b)" is my function, where z,u,a,b are parameters which
> should be found and "r" (x) is the only argument.
>
> NMinimize finds something, but does stupid things:
> 1) when I declare function "f" as f[r_, z_,u_,a_,b_] := z*(r^(-a/
> 110)*(1+r/u)^(b/100)); I got normal results
> 2) when I declare function "f" as f[r_, z_,u_,a_,b_] := z*(r^(-a/
> 10)*(1+r/u)^(b/10)); I got worse results
> 3) when I declare function "f" as f[r_, z_,u_,a_,b_] := z*(r^(-a)*(1+r/
> u)^(b)); I got the worst results
>
> Why I have to divide "a" to get better results?? Is it possible to
> avoid using such tricks and get NMinimize work with the 3rd case?
> Calculated values of parameters are not very small or very large,
> normal numbers....
> I'll be grateful for any remarks/suggestions/ideas. To try, just copy/
> paste the code below into Mathematica and run it.
>
> points = {
> { 4.68, 0.156},
> { 7.63, 0.100},
> {10.56, 0.074},
> {13.48, 0.061},
> {16.55, 0.048},
> {20.33, 0.029},
> {25.39, 0.023},
> {30.28, 0.018},
> {36.05, 0.016},
> {42.60, 0.016},
> {55.49, 0.028}} ;
>
> weights = {0.2523, 0.4007, 0.5801,0.6193, 0.6798, 1, 0.7957, 0.5144,
> 0.37030, 0.1079, 0.0189};
>
> f[r_, z_,u_,a_,b_] := z*(r^(-a/110)*(1+r/u)^(b/100)); (* Here z,u,a,b
> are parameters *)
>
> x = points[[All,1]];
> y = points[[All,2]];
> n = Length[points];
> bestFitVals = NMinimize[
> Sum[
> ((f[x[[i]],z,u,a,b]-y[[i]])^2)*weights[[i=
]],
> {i, n}
> ],
> {z,u,a,b}
> ]
> Show[
> Plot[
> Tooltip[f[r,z,u,a,b] /. bestFitVals[[2]]]=
,
> {r, Max[xdata], Max[ydata]}
> ],
> ListPlot[points],
> PlotRange->{{0,Automatic},{0,0.2}},
> AxesLabel->{"r", "f[r]"}
> ]
NMinimize has a default starting search range of -1 to 1 for the
variables. That may be why you get better results when you scale.
You can change the default range in the variable specification, for
example {{x, -10,10},{y,-10,10}} etc.