I have many curves that I want to fit using a convolution of some functions. These functions include Weibull distributions with 2 parameters lambda and k, as well as a function B(t) such as measured curves to fit to model = F(lambda1, k1, kambda2, k2) + B(t)

The main problem here is that even if the lambda's, k's and B are not colinear, they can be "kind of" substituted and the optimization can lead to different local minima, with a close final error, but parameters not close at all.

This is a problem because I intend to interpret the value of these parameters as natural characteristics of the objects I observe.

Our actual approach is to minimize the number of parameters, i.e. fixing some of the lambda's and k's, as we would do if there were a function linking them. However this is arbitrary + this is a sacrifice as I can not interpret this parameters value anymore.

So question : is there a method / analysis / related problem / science paper dealing with this problem of unstable optimization when parameters are not exactly perpendicular degrees of liberty ?


Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Browse other questions tagged or ask your own question.