[mlpack] GSoC 2014 simulated annealing optimizer
Ryan Curtin
gth671b at mail.gatech.edu
Thu Jul 3 13:05:33 EDT 2014
On Tue, Jul 01, 2014 at 11:46:43PM -0500, Zhihao Lou wrote:
> Hi Ryan,
>
> I've noticed the SA hasn't been merged into the repo. Is there anything I
> can help to speed things up?
Hello Zhihao,
Please accept my apologies. I had meant to take care of this, but then
I became busy with other things and forgot. I have committed your
contribution to the trunk repository and it will be a part of the next
release.
I made some minor changes; a list of changes is given at
http://www.mlpack.org/trac/log/trunk/src/mlpack/core/optimizers/sa
and
http://www.mlpack.org/trac/log/trunk/src/mlpack/tests/sa_test.cpp
If any of the changes I have made are bad, please let me know -- you are
the expert on simulated annealing. I have simply tried to refactor a
few things, sometimes for speed, sometimes for consistency with other
mlpack code.
I asked you some time back to refactor the move distribution into its
own template class, but I took a closer look at your email and the code,
and I think you are right that it doesn't really make sense (or would be
difficult) to use another distribution there, since MoveControl() is
derived for the univariate Laplace distribution. So I moved the
LaplaceDistribution back into sa_impl.hpp, but also expanded on it a
little bit to make it a multivariate Laplace distribution with unit
covariance, and added a LaplaceDistribution to src/mlpack/core/dists/:
http://www.mlpack.org/trac/browser/trunk/src/mlpack/core/dists/laplace_distribution.hpp
Now I think it will be possible to use that LaplaceDistribution class as
a PDF for other machine learning algorithms; one example might be an
HMM.
Thank you so much for your contribution, and again, I am sorry it took
so long.
Thanks,
Ryan
--
Ryan Curtin | "Lots of respectable people have been hit by
ryan at ratml.org | trains." - Penny
More information about the mlpack
mailing list