In dimwit we simply increase the number of visits to a node before it
is added to the UCT tree, to slow down its growth. I wasn't too happy
about how selective the tree got with a long time to think, but it's
unclear if this particular hack had anything to do with that.

Álvaro.


On Mon, Jun 1, 2009 at 4:54 AM, Isaac Deutsch <[email protected]> wrote:
> Hi.
>
> I've been thinking about pondering, and the way the tree has to be built to
> support pondering. Because with pondering, the thinking time for a move can
> be very big theoretically, I would like to handle automatic pruning of the
> tree to avoid running out of memory. Right now I have a fixed size pool of
> nodes, and I simply stop the tree from growing when I see that all nodes
> are used. However I'm afraid this could hurt the performance when thinking
> times are very long.
>
> This brings me to my question: When I see that I'm running out of memory,
> which leaves/subtrees of the UCT tree should be pruned?
>
> -Prune moves with a low winrate and a low variance. This would favor nodes
> near the root, and often lots of memory would be freed this way. However,
> one has to be careful not to prune potentially good moves.
>
> -Prune leaf nodes with little visits that are old. This would have a small
> impact on the UCT search but the memory freed is very little, meaning I
> would have to do a lot of pruning.
>
> Another approach would be of course to just let the tree grow
> indefinitely and hope that it will not use too much memory, but I'm not sure
> it would work well in all situations.
>
> What do you do in your programs? Have you tested other approaches? Do you
> think hard pruning is bad in general?
>
> Regards,
> -ibd
> --
> Nur bis 31.05.: GMX FreeDSL Komplettanschluss mit DSL 6.000 Flatrate und
> Telefonanschluss nur 17,95 Euro/mtl.!* http://portal.gmx.net/de/go/dsl02
> _______________________________________________
> computer-go mailing list
> [email protected]
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/

Reply via email to