|
Le 04/10/2018 à 01:11, Doug Stewart a
écrit :
On Wed, Oct 3, 2018 at 6:19 PM Przemek
Klosowski < address@hidden>
wrote:
On
10/03/2018 05:15 PM, Doug Stewart wrote:
On Wed, Oct 3, 2018 at 4:25 PM
James Sherman Jr. < address@hidden>
wrote:
Hi, I'm
trying to fit some data to a function of
degree 4 or 5, but when I
plot it, it doesn't fit:
DATA:
C30 =
0.063330 0.057200 0.052870
0.047380 0.044860 0.039710
0.037070 0.036030 0.034990
t =
1.0000 1.5000 2.0000 3.0000
5.0000 7.0000 10.0000
13.0000 18.0000
FIT:
p=polyfit(t,C30,4)
0.0000027990 -0.0001193780
0.0018221783 -0.0125401220
0.0727802050
The problem is that when I plot(t,C30) and
plot(p) I get totally different
curves, am I missing somethiing???
Yes, I believe it is safe to say that
you are missing something. Look at vector
that is returned from polyfit that you
print out:
p=
0.0000027990 -0.0001193780
0.0018221783 -0.0125401220
0.0727802050
these are the coefficients of the
polynomial (see help polyfit for more
explanation). And the "curve" from
plotting that vector is just the
coefficients wrt their index. The
function you're probably looking for is
polyval (help polyval for info) so that
you can evaluate the fitted curve at the
values of x that you want.
Hope this helps,
James Sherman Jr.
Polyfit can do that
[P, S] = polyfit (X, Y, N)
S.yf should have the fitted Y values for your
x values.
True, but it does not show how bad the fit is. These
datapoints aren't likely to fit a polynomial model, and to
see it it helps to see the actual fitted polynomial with
more resolution than the 9 points provided:
C30 = [ 0.063330 0.057200 0.052870 0.047380
0.044860 0.039710 0.037070 0.036030 0.034990]
t = [ 1.0000 1.5000 2.0000 3.0000
5.0000 7.0000 10.0000 13.0000 18.0000]
plot(t,C30,'o',x=1:.1:20,polyval(polyfit(t,C30,4),x))
which shows nicely the deficiency of the fourth-order
polynomial, by creating artifacts where the data has none.
Depending on what you're actually trying to do with your data,
non-parametric regression might be a suitable alternative.
Here is a solution using kriging (a.k.a. Gaussian process
regression) :
```
pkg load stk % assuming stk installed
C30 = [ 0.06333 0.0572 0.05287 0.04738 0.04486
0.03971 0.03707 0.03603 0.03499]';
t = [ 1.0000 1.5000 2.00000 3.00000 5.00000
7.00000 10.00000 13.00000 18.00000]';
model = stk_model (@stk_materncov52_iso);
model.lognoisevariance = nan; % assume noisy data
[model.param, model.lognoisevariance] = stk_param_estim
(model, t, C30);
t_box = stk_hrect ([0; 20], {'t'});
t_pred = stk_sampling_regulargrid (100, [], t_box);
C30_pred = stk_predict (model, t, C30, t_pred);
stk_plot1d (t, C30, t_pred, [], C30_pred); legend show
```
See attached figure for the result.
@++
Julien
|
kriging.eps