These are chat archives for elemental/chat

22nd
Oct 2016
Ryan H. Lewis
@rhl-
Oct 22 2016 16:29
@poulson: do you think that there is any possibility that my inverse hessian computation may be numerically unstable
Jack Poulson
@poulson
Oct 22 2016 18:23
I haven't had a chance to evaluate that yet. Have you compared your implementation to the relevant software literature?
Ryan H. Lewis
@rhl-
Oct 22 2016 18:23
Do you have any suggestions, I find many existing algorithms tough to follow.
Ryan H. Lewis
@rhl-
Oct 22 2016 18:24
I was thinking that righting down B_k = L_kL_k' can still be done via rank-1 updates.
but this should involve less arithmetic.
perhaps the same thing could be done with the inverse.
e..g the same thing can be done with the inverse.
Jack Poulson
@poulson
Oct 22 2016 18:26
it would be a good idea to read all of the papers from the linked page
Ryan H. Lewis
@rhl-
Oct 22 2016 18:27
Yeah, true.
i'll do that.
hopefully I can still access :)
Of these papers, it seems that this one is the most salient one for me atm: http://users.iems.northwestern.edu/~nocedal/PDFfiles/representations.pdf
i gotta run now though
Jack Poulson
@poulson
Oct 22 2016 18:31
OK
Ryan H. Lewis
@rhl-
Oct 22 2016 18:36
The implementation I have follows Nocedal & Wright, with a few minor modifications. I would really appreciate it if you wouldn't might taking A look at the line search and BFGS iteration itself and giving me your thoughts.
I'll see what that nocedal paper has to add.
it looks like it has a nice characterization of the updates
gotta run for real.
Jack Poulson
@poulson
Oct 22 2016 18:39
OK
my main thought is that it would be useful to describe all of the variables with words rather than with symbols
e.g., lowerBound and upperBound rather than c1 and c2
and to comment each step with the logic behind it