[Computer-go] How well does pure monte carlo play?

Hiroshi Yamashita yss at bd.mbn.or.jp
Mon Feb 13 08:30:12 PST 2012


I tried pure MC + UCT + RAVE 10000 playouts on KGS.
It got 15k. (name is pureMcUct, but it uses Rave.)

Program         playouts  KGS  CGOS(19x19) CGOS(9x9)

GNU Go          classic   6k   1800        1800
AyaMC               600   6k   1741        1992
AyaBot          classic   8k   1522        1750
pureMcUctRave     10000  15k   1380        1595
pureMcUct         10000        1081        1389

AyaMC         is my MC program.
AyaBot        is my old non-MC program. 
pureMcUctRave is pure MC + UCT + RAVE.
pureMcUct     is pure MC + UCT.

I was surprised there is big difference KGS rank and CGOS 19x19 rating.
AyaBot and pureMcUctRave's Elo difference is only 142(1522-1380), but 
 KGS rank difference is 7 rank (15k-8k).

Hiroshi Yamashita

----- Original Message ----- 
From: "Matthew Woodcraft" <matthew at woodcraft.me.uk>
To: <computer-go at dvandva.org>
Sent: Sunday, January 22, 2012 9:45 AM
Subject: Re: [Computer-go] How well does pure monte carlo play?

> Hiroshi Yamashita wrote:
>> If CGOS rank is right, myCtest-10k-UCT is around 8k on KGS.
>> But I think pure MC+UCT is not so strong.
>> CGOS 19x19 population is few, so it is not reliable.
>> For example, RandomGo's rate 505 is maybe overrated.
> Ah, yes. Looking at the cross-table for myCtest-10k-UCT, it has 104/104
> against myCtest-10k, 0/144 against Gnugo level 8ish, and 3/34 against
> intermediate opponents.
> I don't think there's enough information for any kind of ranking there.
> -M-
> _______________________________________________
> Computer-go mailing list
> Computer-go at dvandva.org
> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

More information about the Computer-go mailing list