Re: [gnugo-devel] endgame module for GNU Go

[Martin Girard]

On Sep 6, 2004, at 21:46, Gunnar Farnebäck wrote:

> Xavier wrote:
> > > That does depend on how exactly you differentiate between the 
> > > concepts
> > > but the most likely answers are "GNU Go doesn't really have any
> > > strategy at all" and "yes".
> > Oh, I did not realize that by reading the documentation.
> > Thanks a lot. So one part of my job would be to realize a startegy,
> > before gluting it with existing tactic.
>
> How do *you* define strategy and tactics in go? It could probably be
> argued that GNU Go has the strategy to get as good score as possible,
> but I'm not sure whether that is at all helpful in this discussion.
>
> /Gunnar
>
>
> _______________________________________________
> gnugo-devel mailing list
> address@hidden
> http://lists.gnu.org/mailman/listinfo/gnugo-devel

More about strategy:

I've just played a game (see attachment) against GNU Go 3.4 (package 
released by Sente for Mac OS X, strength 10). If you replay that game, 
you shall realize two things:

1) I suck;
2) It's not too hard to slaughter GNU Go even if you suck.

GNU Go's weakness is that it greedily and shortsightedly attempts to 
achieve the best possible score as only strategy, which is worse than 
having no stragegy at all, since that makes it utterly predictible.

The strategy I used in that game (killing stones in the center while 
giving away corners) was utterly obvious, yet GNU Go nonetheless fell 
for it and made careless moves a human player of that rank would never 
have made.

What's more, even though it was utterly obvious the stones in the 
center were threatened, GNU Go couldn't react in time because no amount 
of reading could possibly detect such cases in practice (threat is too 
distant), although a quick look is enough to see that vital points 
needed to be defended.

GNU Go won't improve significantly as long as it keeps taking simple 
baits like that; reading more moves ahead won't make much difference 
for obvious reasons. In order for it to improve beyond 1 dan (given it 
ever reaches it), it would need to figure out the opponent's 
intentions, forge a plan in order to foil them while making profit, and 
use influence to attack instead of just playing safe. In short, GNU Go 
is unable to understand intangible value, figure out distant threats 
and come with abstract goals; this is where effort should be put in the 
long run.

--Martin

20040906-222523.sgf
Description: Binary data