Re: SV: [Help-gsl] On conjugate gradient algorithms in multidimentional minimisation problems.

[Max Belushkin]

    Hi list,

  I've ran into a funny issue. I have a certain point, not too far from 
a minimum, which I use as a starting point for fits.

  A simple random walk routine (pick a random direction in which the 
chi squared is smaller by throwing parameters inside a small sphere 
around the current point) works ok. James Bergstra's mm_hess routine 
works fine.

  Built-in GSL conjugate gradient algorithms (Fletcher-Reeves, 
Polak-Ribiere and the vector gradient algorithm) all bail out after 
15-20 iterations claiming iterations are not making progress toward a 
solution.

  Can someone explain what's up with that? Why are the advanced 
conjugate gradient methods failing when even the simplest random walk 
does the job? Slowly, yes, but it is, at least, going on.