Re: [Bug-gnubg] Distributed computing

[Øystein Johansen]

Christopher Yep wrote:
> Hi all,
> 
> Has anyone considered using a distributed computing platform to improve
> gnubg?  E.g. BOINC (http://boinc.berkeley.edu/).  Some games already use
> BOINC (for examples see http://boinc.berkeley.edu/projects.php).

Yes, honestly I have considered some things, yes, but I've never dared
to do anything ;-)

> Some possible applications:
> 
> 1. Rolling out a large MET (e.g. G25 instead of G11) at 2-ply (instead
> of 0-ply).

This is really something a distributed system could do. It shouldn't be
to hard, I believe. The message system can be quite straight foward I
believe, but I have not thought this through, though.

> 2. Training for future versions of gnubg.

The training it self is really hard to distribute. The next "state" of
the neural net, which is the result of one call to backprop., is
depending on the current state. Distributing the backpropagation will
therefore involve distributing the whole neural net and then every
update must be atomic. We would need a full transaction system, I believe.

The JOONE project has something the call DTE, Distributed Training
Environment, but when I read the description of this, it looks more like
a genetic algorithm, which trains the net on several nodes, and the a
master continues with the best net from these nodes. A bit of a quasi
distributed algorithm in my opinion.

However, some thing in the training process _can_ be done distributed.
Rolling out benchmark databases, collecting position training databases,
and rolling out these positions.

> 3. Rolling out an accurate opening book (this could eventually be added
> to gnubg).  Many people are working on rolling out an opening book today
> (I'm involved in some of these efforts myself), but if gnubg were
> integrated with BOINC (or another distributed computing platform), the
> work would go a lot faster; it would be easier to coordinate and
> tabulate results.

Same as above, distributing a collection of rollouts should not be very
complicated if the distributed system algorithms are there.... This may
be the starting point to make such system.

> I'm sure that more people would participate in the project (for #1, 2, 3
> above) under BOINC.

Whatever system as long as it is simple to use.

> Has anyone looked into this (BOINC or any other distributed computing
> platform) to see how much work it would require from an initial
> programming standpoint as well as from a weekly/monthly administrative
> standpoint (to coordinate the rollout work)?

I'm aware of BOINC, but I've also looked at the Spread Toolkit. Both
looks "usable".

-Øystein

signature.asc
Description: OpenPGP digital signature