Timeline for Non-algebric curve-fitting along weighted pointcloud (if possible using python)
Current License: CC BY-SA 3.0
13 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Nov 9, 2012 at 18:53 | comment | added | heltonbiker | I tested your last code against my datasets, and although there were very very promising results (lots of perfect fits), more problematic cases yelded unstable and oscillating results. By doing more research, I think this is a good use case for non-parametric regression, so I posted another question (mentioning this one, and your very useful help) here: stats.stackexchange.com/q/43249/16656 | |
| Nov 9, 2012 at 14:04 | comment | added | Josef | Standard least trimmed squares assumes we have at most 50% outliers. In my trimming to the centerline, 70% of the observations are removed. Exploiting the symmetric structure and some manual trimming might be necessary in cases like this. | |
| Nov 9, 2012 at 13:48 | vote | accept | heltonbiker | ||
| Nov 9, 2012 at 13:48 | comment | added | heltonbiker | Your last example (using my provided sample points) shows a result which represents very much what I was hoping to find. I'm gonna try it with the whole set of points, study the concept of least-trimmed-squares (which I didn't knew), and post some feedback soon after. Thanks again, you've been of much help! | |
| Nov 9, 2012 at 4:45 | history | edited | Josef | CC BY-SA 3.0 |
add comments about variance in RLM
|
| Nov 9, 2012 at 3:48 | history | edited | Josef | CC BY-SA 3.0 |
add a possible solution with trimming to center points
|
| Nov 9, 2012 at 1:20 | comment | added | Josef | About RLM: the default parameters are defined, as usual, with the normal distribution plus outliers as the reference case. I haven't seen data like yours yet, so I don't know which options for robust estimation would produce the best results. (maybe concentrated weights in the center and then flat weights from there.) | |
| Nov 9, 2012 at 1:04 | comment | added | Josef |
about np.vander : yes, it's essentially the same as x[:,None]**np.arange(degree) (degree=order+1, I think np.vander has columns reversed.) Chebychev polynomials would be better, since they are orthogonal. Did you try a low order polynomial first to see whether it works.
|
|
| Nov 8, 2012 at 19:03 | comment | added | heltonbiker |
I ran your code and the results were poor, probably because it is still not properly tuned. I run code from the statsmodels RLM examples and the appear to work for what I need to achieve. I am curious: why did you use np.vander? Does it have to do with the degree of the desired polynomial? Thanks again!
|
|
| Nov 8, 2012 at 15:36 | comment | added | heltonbiker | That seems promising. I am not at work now, but in a few hours I'm gonna make some tests and post some feedback. Thanks for now! | |
| Nov 8, 2012 at 14:51 | review | First posts | |||
| Nov 8, 2012 at 14:53 | |||||
| Nov 8, 2012 at 14:40 | history | edited | Josef | CC BY-SA 3.0 |
fix incorrect explanatory polynomial in x
|
| Nov 8, 2012 at 14:33 | history | answered | Josef | CC BY-SA 3.0 |