But they all fail... except for verifying that my objective is correct at the converged values.
On Fri, Sep 30, 2011 at 4:51 PM, Greg Sterijevski <gsterijev...@gmail.com>wrote: > Hello All, > > I have been playing with the non linear optimizers in Commons for a few > days now. The context in which I am working is Probit/Logit/Tobit type > limited dependent variables models. These are well known problems. I found a > very well known example of a probit estimation ( > http://support.sas.com/rnd/app/examples/ets/margeff/index.html) , the data > and attempted to fit the probit model. The only technique which generated > 'close' answers to the ones in the publication was Marquardt-Levenberg. The > rest of the routines failed in a variety of ways. > > That got me to thinking about the tests. While there are tests in Commons, > maybe there should be more. So I began writing a test collar for all of the > methods (I attached it for comments). To my chagrin, the 'easy' problem from > NIST (http://www.itl.nist.gov/div898/strd/nls/data/LINKS/DATA/Misra1a.dat) > fails all of the optimizers I tried. Powell gets close on the second > starting point. It is possible I screwed up the derivatives or did something > foolish-since I am new to the commons optimizers. It is hard to verify the > derivatives numerically since they (the objective functions) are designed to > fool numerical derivation. > > What are people's thoughts? Before I continue adding tests will-nilly... > > > Thank you, > > -Greg >
