[computer-go] New version of Crazy Stone

[Don Dailey]

You make a good point about complexity.   Even if you could design
a program that plays perfect go or chess it would not necessarily
get the best results unless it also knows how to complicated
things.     

One can imagine that if chess is a draw,  a perfect player could make
things too easy by quickly exchanging off pieces and getting down to
a simple draw.    

However, I believe that even a naively coded "perfect player" would
be way too much for even the best grandmasters.   They would rarely
if ever achieve a draw.    Of course it's far better to design such
a player to keep things complicated and make it difficult for a player
to find his way.   Once the human opponent makes his first game
theoretic blunder it's all over and the program can then play very
straightforward (as opposed to trying to trick the human opponent
into losing even faster.)

I learned this principle myself when I played club chess years ago.
I was slowly but surely losing to a much weaker player but I found
a move which was was really quite bad in the game theoretic sense 
but required a very precise response.  If my opponent had played the
correct response it would have been over very quickly for me.  But if
I had played "normally" he would have won the game without much 
resistance possible as the game was about to simplify into a slow but
easy win for him.   With my best poker face I played the "tricky" move
and my opponent immediately assumed I had blundered.  Actually, the
move was a calculated blunder because it had the appearance of a 
more trivial blunder.    It was truly an ugly swindle and I was almost
ashamed that I had stooped so low.   But it worked and after that I
realized that GO as played but us mortals is more like poker than
I had previously believed.

I had since learned that many stronger players take into consideration
the weaknesses and strengths of their opponents.   Being more of an
idealist I didn't like this at all - at first - until I realized that
it was really an important part of the game.  

By the way,  what do you consider a "blunder"?   The most basic
definition is any move than changes the game theoretic result.   If you
have a draw but make a move that converts this to a loss, it is a
blunder.    But if you have a 37 move checkmate and you play a move
that turns this into a 47 move checkmate - is it a blunder?

By the most basic definition, if you blunder into a loss you cannot
blunder again until your opponent blunders the game back into a draw
or win for you.   If you have a win you can blunder twice but no more
unless the opponent counter-blunders!   

So if you wanted to build a 32 man endgame database it would be nice
if you could do it 1 bit per position.   But you could only represent 2
states.   Which 2 states would you keep?   Would you play to win, or
NOT to lose?    Which type of database would make the better opponent
for human players if we assume the initial position is a draw?

        state 0     state 1
        --------    -----------
    1.  win         not wins
    2.  lose        not lose    

I think that you would want the lose database.   If hydra was backed up
by the loss database, it would achieve better results that if it had
access to a win database (that's my assertion.)     However,  if your
contention is that humans will inevitably play a game theoretic losing
move (whether the opponent actually realizes this or not) then a win
database might give better results although it would make it possible
for hydra to lose an occasional game.

- Don


On Sun, 2006-06-04 at 09:08 +0200, Chrilly wrote:
>  don't have a link, but I read a description many years ago.  They
> looked at the variance in score between players of the same strength,
> and found that weaker players have higher variance.  They plotted the
> variance against the strength, and extrapolated to zero variance,
> which would be perfect play, at about 4 stones above top pro.
>  
> According to "Chrilly's law" humans make in complicated positions
> within 10 moves a serious blunder. This 10 moves is not the mean, but
> the "beta-bound". The mean is about 7 moves. This are not blunders in
> the sense of god, but blunders Hydra sees. And Hydra is of course very
> far away from God. The opening repertoire but also the evaluation of
> Hydra is tuned according this law. I do not care too much, if the eval
> is correct, the position must be just complicated. Its along
> Bronsteins "Do not solve problems, create them".
> It was ever known that play of average club-players is a balance of
> blunders. Who makes the second-last blunder wins. But it was thought
> that GM play with a few exceptions - mainly in Zeitnot - almost
> perfect. The latest generation of programms has proven this wrong. GMs
> make even in "simple" endgames like Queen against Rook serious
> blunders. Although the longest mating distance is 31, they are not
> usually able to mate a perfect opponent in 50 moves. In human-human
> games the mating distance is close to the theoretical one, because
> both make the same error.
>  
> I am convinced that God can play easily with a Knight-handicap in
> chess. 
>  
> Chrilly
>  
> _______________________________________________
> computer-go mailing list
> computer-go at computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/