[computer-go] MC and tactics?

[Don Dailey]

Chrilly,

Most of the MC players have a search layer before the MC starts.

But yes, the mutual stupidity principle makes the MC part of the search
much stronger.   

That's why MC by itself is not so good.  If you have a group of stones
subject to capture but can be defended with only 1 specific move,  then
it's a matter of who reacts first, and this is random.   However, that
doesn't mean MC won't find it and play correctly because the winning
move will get a far higher score.   But in general, simple MC without
enhancements is terrible at tactics.   (The big surprise to me is that
it is so good at what us chess players call "positional play" that this
is largely compensated for, at least on small boards.)

But with the better MC players this is noticed very quickly and the
defending move will get sampled far more often.

I have a GO problem that I use to test with - the program must play a
move that is self-atari, then play just the right move into the newly
created hole (to prevent the opponent from eventually making an eye.)

The program can solve it pretty quickly but not because it see's all the
way to the end of the tactic,  only because it see's enough to get to a
position that the MC evaluation can understand.
 
But in general, simple MC without enhancements is terrible at tactics.
The big surprise to me is that it is so good at what us chess players
call "positional play" that this is largely compensated for.   With some
work you can fix the tactics.  
 

- Don


On Mon, 2006-09-18 at 08:12 +0200, Chrilly wrote:
> How does a MC "see" a forced tactical sequence?
> Lets assume a simple 10 moves deep ladder. Even if the probability is 0.5 
> for playing the correct move, the final probability is 0.001.
> Does the MC player reley on the principle of mutual stupidity? That the 
> defender does not play the correct defence-moves either? In case of 
> learning/updating of the playing statistics, this would mean, that both side 
> must have the same learning rule, otherwise one side learns to win sequences 
> which are no win against correct defence and similarily to defend positions 
> that are in fact lost.
> 
> The principle of mutual stupidity is also important for database statistics. 
> E.g. the Hydra chess expert GM-Lutz analysed the wining/drawing statistics 
> of certain endgames from the ChessBase MegeBase. In most cases the 
> statistics are from a certain playing level on (about 2400 Elo) not anymore 
> related to the Elo-rating. 2 Super-GMs have about the same chances than 2 
> IMs playing each either. There can be no doubt, that the Super-GMs play the 
> endgame better, but the higher error rate of the IMs compensates each other. 
> It is also known that GMs have difficulties to mate God with the queen 
> against a rook. But they win relative easily against another human of the 
> same strength.
> 
> Chrilly
> 
> Note: A MC programm can of course augmented with a ladder-routine. But thats 
> not the point of my question. 
> 
> _______________________________________________
> computer-go mailing list
> computer-go at computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/