How to find uncertainties of estimated parameters in Levenberg- Marquardt algorithm

13 次查看（过去 30 天）

Chris Martin 2015-11-4

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/252736-how-to-find-uncertainties-of-estimated-parameters-in-levenberg-marquardt-algorithm

评论： Star Strider 2015-11-10

Hello I have used Levenberg-Marquardt algorithm in 3 parameter problem but I unable to estimate uncertainties associated with it, Please help me in this regard

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Chris Martin 2015-11-6

I need the formula

请先登录，再进行评论。

请先登录，再回答此问题。

回答（2 个）

Torsten 2015-11-4

1
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/252736-how-to-find-uncertainties-of-estimated-parameters-in-levenberg-marquardt-algorithm#answer_198399

If you have the statistics toolbox, use nlinfit together with nlparci and nlpredci.

Best wishes

Torsten.

4 个评论
显示 2更早的评论隐藏 2更早的评论

Chris Martin 2015-11-5

There is nothing in that link

Torsten 2015-11-5

Just wanted to give a link to this book:

http://www.amazon.com/Nonlinear-Regression-George-F-Seber/dp/0471471356/ref=sr_1_2?ie=UTF8&qid=1446721830&sr=8-2&keywords=Seber+wild

Best wishes

Torsten.

请先登录，再进行评论。

Star Strider 2015-11-5

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/252736-how-to-find-uncertainties-of-estimated-parameters-in-levenberg-marquardt-algorithm#answer_198610

Finding the covariance matrix of a nonlinear regression requires that you calculate the Jacobian matrix at the convergence (the last parameter estimates calculated). In the context of the Wikipedia notation, the partial derivatives are with respect to each parameter at each data point, so the partial derivative of f(1) with respect to x(1) refers to the value of the partial derivative of the function at the first value of the independent variable with respect to the first parameter. It continues row-wise down the values of the independent variable, and column-wise across the parameters.

In a nonlinear regression, some or all of the partial derivatives of the function with respect to each parameter are by definition functions of themselves or of other parameters, so expect that in your result. (This is known as the model being ‘nonlinear in the parameters’.) This is easier if you have the Symbolic Math Toolbox, since it can calculate the symbolic Jacobian for you.

After that, you can plug your evaluated Jacobian matrix into the design matrix for a linear problem to calculate the covariance matrix, as explained in Least squares, regression analysis, and statistics. You can calculate the parameter confidence intervals from the diagonals of the covariance matrix and the t-distribution.