Mark Boon wrote:
> Question: how do MC programs perform with a long ladder on the board?
>
> My understandig of MC is limited but thinking about it, a crucial long
> ladder would automatically make the chances of any playout winning
> 50-50, regardless of the actual outcome of the ladder. If this is the
> case then:
>
> a) when winning it will avoid all ladders, even the ones working for it.
> b) when losing it will look for situations involving ladders, even
> when they're not working.
>
> Is this correct? Probably on 9x9 this is not a problem because all
> ladders will be relatively short.
>
> Mark
The best first search is pretty focused with good heavy play-outs but
the strength of MC programs is definitely not tactics and I think long
ladders are a problem - though not as much as one might think.
I think your analysis is more or less correct - if it can't read a
ladder it is likely to see it as rather ambiguous which pushes the
winning probability towards even. If it's happy with it's position
otherwise, it would probably gravitate away from them and otherwise
gravitate towards them just as you say.
In fact, this illustrates a wonderful strength of these programs. They
are very pragmatic even about what they don't do well. Any good human
player has the sense to avoid tricky situations he doesn't understand if
he is winning anyway. And a good human in a game losing situation
knows that it's time to introduce some risk if the straightforward
conservative play is leading down a dead end. In other words, the MC
programs do not always calculate risk correctly (otherwise the score
would always be 1 or 0) but they deal gracefully with it and they are
"true to themselves", a characteristic we humans lack. We prefer to
struggle with self-doubt many times which often clouds our judgment.
Naturally, if we can improve our judgment it's a win and this applies
equally to MC programs.
I myself have used this same principle in a chess tournament. I was
playing a much weaker player and ended up with a losing position due to
bad play on my part and good play on my opponents part. However, at
one point it became clear that I was going to lose if I continued with
standard play, I was not making it difficult for my opponent. I
discovered an interesting (but unsound) king side attack but it was
profound enough that it took some work for me to see that it was
unsound. It turned out that there was a clever defense that would
have given the opponent the win and the game would be over quickly if he
could find this move. Fortunately, it also turned out that the
"clearly obvious" defense was losing - but it was not easy to see this.
I think at this point I played like a monte carlo player. I estimated
my odds of winning with the swindle to be much greater than the odds of
winning with correct play (correct in the sense that it delayed the loss
as much as possible.)
So I did exactly what MC programs do, in a losing position I chose the
murky waters. In this case I won the game by playing the opponent
instead of the board. In that same tournament a friend of mine also
made a calculated choice to avoid a line of play against a weaker
opponent that was very likely to end in a draw, even though the move he
chose instead was inferior. He wanted to win, not draw and he chose
the move that he calculated would give him the best chances of "not
drawing." In other words, as he admitted later, the move he chose
gave him a greater chance of losing or winning instead of drawing.
His gamble also payed off in this case.
It's usually non-productive to play the odds in this way because we are
actually playing the opponent, not the board and it's more important to
play the board, not to mention the fact that playing the opponent
usually means purposely changing your own style - usually weakening you
game a bit. Still, in some situations the choice is not so
difficult and you can get away with playing the opponent by applying the
MC principle of estimated your winning chances.
Of course MC programs don't consider the opponent when the make these
estimates, just the position.
- Don
> _______________________________________________
> computer-go mailing list
> [email protected]
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
