Petr Baudis: <[email protected]>:
>On Wed, May 25, 2011 at 10:41:14AM +0900, Hideki Kato wrote:
>> Mark Boon: <[email protected]>:
>> >But for MCTS it's not very efficient to simulate the same leaf many
>> >times at once. So it will probably only be worth it if you can get in
>> >a fairly large number of simulations at similar cost to one simulation
>> >done by a CPU.
>> 
>> That's not correct.  Leaf parallelism is efficient enough and can 
>> compete root parallelism.
>
>Well, I'd argue root parallelism is not efficient enough either. ;-)

Either only reduces the variance; never expand the tree.  We have to 
develop some better way. ;-)

>> Past bad results are mainly due to synchronizing all simulations
>> invoked from a leaf.
>
>Can you elaborate on an alternative to this? I thought that treating all
>simulations invoked from a leaf as a single unit was the very point of
>leaf parallelism.

Just simply avoid synchronization.  Tree-part updates the info in the 
search tree as soon as a result arrives, start descending tree, and send 
the leaf position to be simulated.
# I used broadcasting (udp/ip) but point-to-point is also possible.

For detail, see 
<http://www.geocities.jp/hideki_katoh/publications/gpw08-private.pdf>.

Hideki
-- 
Hideki Kato <mailto:[email protected]>
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to