On Tue, 2008-11-18 at 19:42 +0000, Oliver Lewis wrote:
>         It doesn't make any sense to me from a theoretical
>         perspective.  Do you have empirical evidence? 
> 
> 
> I agree that empirical evidence is required, but theoretically, if MC
> converges to something that is not perfect play, then as the number of
> playouts goes up, the probability of playing a different move (the
> perfect move) goes down, so the programme could get weaker.  Does that
> make sense? 

No.  Your reasoning is wrong.  It's true that there are many cases where
it won't play the best move with infinite playouts,  but there are also
many times when it will.   

Game playing ability is about errors, not playing good moves once in a
while.   A good move is really just a move that isn't an error.
Playing at a high level is like walking through a minefield - if you can
avoid the bad places, you win.  

This is real obvious in chess, not quite as obvious in Go.  A 1 ply
chess search still plays reasonable looking chess but with many tactical
errors.  Add an extra ply of search,  and you eliminate a whole class of
stupid errors, but you still have a lot of errors left.   Each
additional ply eliminates more classes of errors.   

As humans we tend to think in the other direction, we view moves in
terms of goodness as if it's on a sliding scale.   How many times have
you heard the terminology, "that was a master level move",  or "that was
a 5 dan move."    Or "that was ok, but this is better."      It makes it
sound as if each move has a rank associated with it.   

You should also consider that even weak players play the very best move
some of the time.   In chess I probably play the same moves a
"grandmaster" would play a high percentage of the time, but I am far
from a grandmaster in strength.  It's those pesky bad moves that kill
you.   

- Don



>  The effect may depend on the opponent.

Attachment: signature.asc
Description: This is a digitally signed message part

_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Reply via email to