[Computer-go] RAVE questions

Daniel Shawul dshawul at gmail.com
Sun Apr 10 18:47:58 PDT 2011


Ok I will try to have them both play at the same time on CGOS.
Surprisingly though the version with RAVE only score (i.e beta = 1) beats
the one with UCT only (beta = 0).
I consider 24 half plies from the current node downwards to calculate the
RAVE score, which
can sometimes get into the playout zone as well. The RAVE score has the UCT
score at that particular
node but is hugely biased by the rest of the nodes. I am surprised it beats
UCT only by 80 - 90 %.

On Sun, Apr 10, 2011 at 8:52 PM, Michael Williams <
michaelwilliams75 at gmail.com> wrote:

> It sounds like a major bug that is resulting (by luck) that the RAVE
> version beats non-RAVE.  You should compare your non-RAVE version to the
> reference bots in terms of elo.  If they are on par, then.. even more
> intereasting...
>
>
>
> On Sun, Apr 10, 2011 at 7:58 PM, Daniel Shawul <dshawul at gmail.com> wrote:
>
>> Sorry if this is a double post.
>>
>> I am getting very contrasting results using RAVE with UCT. In self play
>> games
>> the one which uses RAVE almost scores 90% against the pure UCT version.
>> But
>> when I get the RAVE version up on CGOS , it seems to score 50 or more elos
>> weaker...
>> Is this a problem with self-test ? I made the self tests from 1000 - 30000
>> playouts per move
>> still the RAVE version wins by a big margin.
>>
>>
>> _______________________________________________
>> Computer-go mailing list
>> Computer-go at dvandva.org
>> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
>>
>
>
> _______________________________________________
> 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/20110410/7a78640b/attachment.html>


More information about the Computer-go mailing list