[Computer-go] a pro game which is computer-unreadable

Brian Sheppard sheppardco at aol.com
Wed Nov 30 20:05:16 PST 2011


I am not sure that it is worthwhile to force MCTS to find such sequences,
but it isn't hard to do so.

For example, if Black can play a ladder out for N moves, then the search can
include all N positions in which Black can break off the ataris as
additional nodes of search. That is, treat the N sequences of atari/escape
pairs as single "macro" moves.

In effect, instead of having a node with, say, 200 legal moves, you have 200
+ N legal "moves."

If anyone has an opportunity to test such an idea, please let me know how it
turns out.

Brian

-----Original Message-----
From: computer-go-bounces at dvandva.org
[mailto:computer-go-bounces at dvandva.org] On Behalf Of Thomas Wolf
Sent: Wednesday, November 30, 2011 10:24 PM
To: computer-go
Subject: Re: [Computer-go] a pro game which is computer-unreadable

This lader together with a hand full of others from professional games are
discussed in my book Mastering Ladders, Richmond, VA: Slate & Shell, 2008.
ISBN 1-932001-40-9 On the accompanying CD there are about 150 more
professional games with longer ladders, many intentionally lost because this
gives very heavy sente moves that may secure more territory than the ladder
loses.

Thomas

On Wed, 30 Nov 2011, Olivier Teytaud wrote:

> Hi;
> sorry for taking some of your time with non-technical long-term AI/GO 
> dreaming, but if sometimes you find Go fascinating you might like the 
> video below :-)
> 
> As many of you I guess, I've spent time trying to design some sort of 
> learning in MCTS, so that monte-carlo simulations would be "adaptive" 
> to the current situation. This idea looks like a very natural solution 
> to the problems we have for reaching human top-level.
> 
> I've met this incredible game; I'm not a Go player, but like many 
> not-so-strong players at first view the moves by black look like a big 
> mistake (misunderstood ladder):
> In fact, it's (as far as I see...) a very clever idea by black (Lee 
> Sedol, pro player), in spite of the fact that it's a failed ladder.
> 
> http://www.youtube.com/watch?v=beic62XoHnM
> 
> We tried various things for having machine learning in MCTS:
>    - Contextual Monte-Carlo for online learning simulations:
> http://hal.inria.fr/inria-00456422/
>    - poolRave (using RAVE values in simulations):
> http://hal.inria.fr/inria-00485555/
>    - Bernstein Races for offline learning patterns 
> http://hal.inria.fr/inria-00622150/
>     (a synthesis of these papers in 
> http://hal.inria.fr/inria-00544758/ ) and many of you have published 
> related stuff; but when a computer will be able to understand a 
> situation as the game above, it will be very impressive to me :-) Go 
> looks like a combination between feeling and mathematical reasoning. 
> One day the people of the Go-sect will convince me that this game has 
> something really special :-)
> 
> In particular, my feeling is that a 10kyu can not play this pro game, 
> but a 10-kyu can understand a posteriori. It's difficult the discuss 
> the possibility for a computer to understand a posteriori, but with a 
> little bit of provocation from this point of view computers are not 
> yet 10-key  :-)
> 
> Best regards,
> Olivier
> 
>




More information about the Computer-go mailing list