[
https://issues.apache.org/jira/browse/MATH-176?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12554853
]
Luc Maisonobe commented on MATH-176:
------------------------------------
I have started implementing this feature, using the provided patch as a guide.
Since the patch code is not Levenberg-Marquardt specific, I intend to add an
AbstractEstimator class which will be the base class for the two existing
estimators (Gauss-Newton and Levenberg-Marquardt).
I encountered one problem during tests. The LevenbergMarquardtTest.testTrivial
is based on a simple linear one parameter problem, solved with one measurement.
Basically, we solve 2x = 3. Since the residual after solve is exactly 0, I
would have expected that the guessed error would be 0 too. In fact get a NaN.
The reason for this is because the error guessing performs a division by dof
(which stands for degrees of freedom I suppose) which is 0 since there are
exactly as many measurements as parameters.
My statistics skills are rather poor, so I don't understand either the
rationale for the chi square and dof values.
If anybody could provide references or hints, please add comments to this issue.
> Getting errors in estimated parameters using Estimator
> ------------------------------------------------------
>
> Key: MATH-176
> URL: https://issues.apache.org/jira/browse/MATH-176
> Project: Commons Math
> Issue Type: Improvement
> Affects Versions: 1.2
> Reporter: Kazuhiro Koshino
> Priority: Minor
> Fix For: 1.2
>
> Attachments: LevenbergMarquardtEstimator.java.diff
>
>
> To validate estimated parameters by GaussNewtonEstimator and
> LevenbergMarquardtEstimator, we need errors (or covariances) in the
> parameters.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
