Choosing the initial values of a parameter at beginning of procedure like estimation of parameters is crucial. AS far as I know, good starting point leads to fast converging, for example, the good starting point, the fast converging in a calculation of hessian matrix.
now, How do I choose an initial value for the parameter?
thanks
Dear Behrouz,
I suggest to you a links and attached files in topics.
-How to choose initial values for nonlinear least squares fit - Cross ...
stats.stackexchange.com/.../how-to-choose-initial-values-for-nonli...
- r - How to find good start values for nls function? - Stack Overflow
stackoverflow.com/.../how-to-find-good-start-values-for-nls-functi...
- r - Why is nls() giving me "singular gradient matrix at initial parameter ...
stats.stackexchange.com/.../why-is-nls-giving-me-singular-gradient...
- PROC NLMIXED: Computational Problems :: SAS/STAT(R) 9.2 User's ...
https://support.sas.com/.../default/statug_nlmixed_sect032.htm
- GAUSS's Maximum Likelihood Procedure - Faculty Washington
https://faculty.washington.edu/ezivot/maxlik.htm
Best regards
I would say that scaling your data is the more critical issue. If you data contains a lot of very small and/or very large numbers, it will be difficult to choose start values. In fact convergence may fail altogether. If your data are properly scaled then starting values are easy. You could then start with a vector of ones, for example.
If your model is some variation of one with a closed form solution, regression say, you could run a regression for start values.
- Badges
- Science topic
- Similar topics
- Chemistry
- Theoretical Chemistry