[SciPy-dev] GSoC weekly report
dmitrey
openopt at ukr.net
Sat Aug 25 02:24:06 EDT 2007
Hi all,
this week 2 solvers were implemented in optimizers module:
- based on cubic interpolation
- Barzalai and Borwein (monotone and non-monotone versions)
Unfortunately, as I had informed Matthieu, cubic interpolation algorithm
contained a mistake (after block "h>0?", "No" branch, there should be
added "phi1', phi2' = phi2', phi1'"), and it took a significant time for
me to find and fix the one.
but when I tried my examples it turned out that as for Barzalai and
Borwein (nonmonotone), the book I'm following doesn't provide default
values of all those initial settings (sigma, sigma1, sigma2, alpha_min,
alpha_max, pho). Ok, I know that 0 < sigma1 < sigma2 < 1 and 0<pho<1,
sigma>0, 0<eps<<1, but I don't know any suggestions about alpha_min and
alpha_max. I took alpha_min =1e-6 and alpha_max=2, but alpha_max=1e6
works much better (on the other hand, maybe my func was very badly
scaled and that's the matter). Web search "Barzalai and Borwein" also
doesn't yield me suggestions about alpha_min and alpha_max.
Now I'm working on second order Armijo step rule and second order
Wolfe-Powell step rules. These ones, as well as some other steps
assigned to me, require Cholesky factorization. I intend to use that one
from scipy/numpy (don't write by myself, despite there is a little bit
modified alg), because that one from the book will require sufficient of
time to be implemented (and that one is not vectoriezed).
Regards, D.
More information about the SciPy-Dev
mailing list