[Computer-go] "Power of Forgetting" paper

[Jean-loup Gailly]

> I have become more cautious about modifying the policy because often ideas
that are good with
> few simulations and light playouts do not scale to Fuego with long
thinking times on big hardware.

I also tried the killer-reply heuristic in Pachi, and it didn't help even
with forgetting. As Martin, I suspect
this heuristic is mostly useful for low end hardware but I didn't check
this.

Jean-loup

2011/1/11 Martin Mueller <mmueller at ualberta.ca>

> > What I perceive is that Orego, while "Mogo-like," still has a fairly
> light
> > playout policy. For example, in the paper, Orego (using the
> > killer-reply-with-forgetting heuristic) defeats Gnugo maybe 87% at 32K
> > trials on 9x9, whereas Pebbles defeats GnuGo 93% using 10K trials.
> >
> > Orego places the killer-reply heuristic as the very first rule applied. I
> > speculate that engines that employ heavier playouts will benefit from
> > placing a killer heuristic farther down the rule set.
>
> Yes, this is pretty close to what we found with Fuego.
> The method in the old paper (without forgetting) did not work for us at
> all.
> The new method with forgetting works, but only on big boards and only as a
> later rule. I think we used it as the last rule before random. I forget the
> details, but it was something like 62% wins in selfplay on 19x19 in
> relatively fast games. Arpad did the tests, maybe he can reply if he reads
> this.
>
> This is not yet in the Fuego codebase - it needs more testing to see if it
> works on 13x13 and how it behaves in tournament-like (more time, multicore)
> conditions.
>
> I have become more cautious about modifying the policy because often ideas
> that are good with few simulations and light playouts do not scale to Fuego
> with long thinking times on big hardware. I think the MoGo team has made
> similar observations.
>
>        Martin
> _______________________________________________
> Computer-go mailing list
> Computer-go at dvandva.org
> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20110112/b1f9c988/attachment.html>