[computer-go] Pay-as-you-go cluster?

[sylvain.gelly at m4x.org]

Le Dimanche 01 Octobre 2006 19:40, Rémi Coulom a écrit :
> Chris Fant wrote:
> > Can someone please do this?  I'd love to see a good MC algo scaled out
> > to 100 CPUs in the next tourney.
>
> Well, I have access to that machine: https://www.grid5000.fr/
>
> But it is not so easy to scale a Monte-Carlo algorithm to work on a
> large cluster. SMP with fast access to shared memory is very easy.
> Finding a way to efficiently distribute UCT over a fast network looks
> like a fun research project.
>


Hello,
as we have access here to an interesting cluster, I have worked several weeks 
on a way to efficiently distribute UCT. I have not succeed yet, and I think 
it is a quite difficult problem. I think there are two main issues:
- find the good data chunk size as stated by David.
- UCT mixes the exploration frequency and the weights in the averaged value 
(as the value of a node is the average of values of the childs weighted by 
the frequency of exploration of theses childs). So if you explore more one 
node than UCT would want to explore it (for example you send it to a machine, 
and wants the result for all the possible moves), then the result will be 
highly biased. 

I have tried a lot of solutions, which seemed good ideas, but none of theses 
were so far successful. 

Of course, I would let you know if it becomes successful, and I would be very 
happy to hear if someone has found a good solution.

Sylvain