[mlpack] GSoC 2014 simulated annealing optimizer

[Ryan Curtin]

On Sun, Mar 23, 2014 at 11:22:30AM -0500, Zhihao Lou wrote:
> Dear all,
> 
> I'm a PhD student in Computer Science Department at the University of
> Chicago, currently working on a scalable parallel simulated annealing
> algorithm, which I recently presented a poster at SIAM Parallel Processing
> 2014. (Anyone interested in my poster please feel free to let me know. The
> file is a few megabytes so I'm not sure if I can sent it to the mailing
> list.) I'm greatly excited when I learned you need someone to work on a
> simulated annealing optimizer in your Google Summer of Code 2014 ideas
> posting. I believe my understanding in the simulated annealing algorithm
> itself, its parallelization and implementation can add greatly to your
> project.
> 
> On the other side, my committee requires me to have a lot more problems to
> test against than the one test problem I shown in my poster. Working on
> your project will help me to learn how to structure an optimization
> algorithm to be a general purpose c++ library, which I urgently needed. I
> believe such a collaboration will be mutually beneficial.
> 
> Currently, I have a working C++ template based implementation, both serial
> and parallel (using MPI), that aims at maximum modularity, rather than
> performance. Naturally, as an on going research project, I'll have to test
> the effects of different cooling schedule or move generation independent of
> other components in the code. Besides the genetic pattern formation model
> (which is the main reason behind this project), I tested it on Rastrigin
> function http://en.wikipedia.org/wiki/Rastrigin_function and it can solve a
> 1000-dimension problem in 40 minutes, with final results around 0.0005
> given 0 as theoretical optimal. (And as I said, there're a lot to gain in
> performance once the algorithm settles.)
> 
> Although I never used or worked with your mlpack library, I'm confident I
> can learn quickly. Any help on where should I started will be deeply
> appreciated.
> 
> In addtion, there's a question I'd like to ask, though. After glancing at
> the paper presented at the BigLearning workshop, I'm feeling the
> scalability in the paper means something different than what I'm familiar
> with, i.e., parallel scalability. Can someone please elaborate a little bit
> on this, and tell me what the ideal use case of mlpack would be?
> 
> I believe I'll be the perfect fit for this GSoC14 project and looking
> forward to work with you.

Hi Zhihao,

I'm sorry to say that the deadline for GSoC proposal submission was last
Friday, and was not to be done on the mailing list anyway, but instead
the GSoC Melange website.

Scalability in the paper refers to the idea that the implemented
algorithms are more scalable (i.e., mlpack implements O(n log n)
algorithms instead of more common O(n^2) implementations).  This is
different than the usual parallel meaning of the word but it was used to
differentiate from "fast" or "efficient" (which only imply that the
implementations are well-written, not better algorithms).  In the future
it is hoped that mlpack can be made parallel, but it is difficult to do
this, especially while maintaining API cleanliness.

At the same time, it is not a requirement to be a GSoC student to
contribute to mlpack, so feel free to contribute.

Thanks,

Ryan

-- 
Ryan Curtin    | "Are you or are you not the Black Angel of Death?"
ryan at ratml.org |   - Steve