Better estimation of gamma parameter for location shifted distributions
Closed this issue · 1 comments
The current method of estimating the gamma parameter for location shifted distributions (Weibull_3P, Gamma_3P, Exponential_2P) is not always accurate. The method is an over-simplification and works by setting the gamma parameter equal to a value that is slightly less than the lowest failure time and then adjusting the rest of the data so that it looks like a non-location shifted distribution.
This was done because reliability.Fitters uses scipy.stats to provide the initial guess into the optimization routine when trying to minimize the log-likelihood. Unfortunately, scipy.stats is really bad at giving an accurate guess for the gamma parameter and it was found that convergence rarely worked or the solution provided was very wrong.
In order to fix this, I suspect that we will need to come up with another method of providing the initial guesses for the optimizer so that we do not need to rely on the fit (of the failure data) provided by scipy.stats.
This problem is most prevalent when there is a lot of censored data or when there is not much data.
This is now resolved in version 0.4.0 and also incorporates a fitter for Lognormal_3P