Re: [computer-go] KGS game-end protocol

[drd@xxxxxxx]

Robin,

This might be nice,  but it has little or no meaning for some programs.  My 
program only cares about winning and not the actual score.    

If Botnoid is completely crushing an opponent,  it will stop making good moves 
as long as it doesn't feel that will prevent it from winning.    It's the 
same if it is losing.    Even against an equal opponent,  Botnoid will tend 
to "barely win"  in most of the won games and it will get "badly beaten" in 
most of the lost games.      

Of course it's the same for Botnoid's opponent.  It doesn't matter how much 
you win or lose by when you are playing Botnoid,  you can't really draw too 
many conclusions from this.

Let me show you some real data as an example:    

Botnoid plays a 19 game match with GNU Go 3.7.4.   On average,  GNU Go wins 
48.1  points of territory per game and Botnoid wins 32.9 points of territory.

Based on the previous paragraph,  which program would you expect won that 
match?

Botnoid actually won that match,  12 games to 7.    It won by a fairly 
substantial margin, although the average territory when summed over all the 
games was significantly lower for Botnoid.


Notes:   a) Botnoid was playing at an "impractically high level"
         b) GNU Go 3.7.4  --mode gtp --score aftermath --capture-all-dead 
--chinese-rules

         c) Botnoid spends over 30 minutes per game at this level.
         d) Gnugo spends less than  1 minute per game.

    

 games    win%  score  Match Up
------  ------  -----  ------------------------------------------
    19   63.2%   32.9  Botnoid   gg_3.7
    19   36.8%   48.1  gg_3.7    Botnoid


 Rating  Win perc  Tot Gms  Ave Time  Player
-------  --------  -------  --------  ------
 1552.6    63.158       19    2122.6  Botnoid
 1447.4    36.842       19      52.9  gg_3.7



Don



On Monday 25 July 2005 11:13 am, Robin Kramer wrote:
> One point I can see.  Is that for the interest of communally improving
> the ability of play of the computer-go community it would be best to
> have the most accurate scoring, it is better to know that you lost by
> a few points than to know that you simply lost.
>
> For the purposes of guiding future strategy knowing which robots were
> the best and how much better they were is a definite advantage.
>
> While, it might take Nick some time to score the board.  It probably
> would take each programmer much more time to implement the "perfect"
> endgame protocol, and their time might actually be better spent
> improving their strategy.
>
> The solution that I see is to either implement the protocol, pay a fee
> for scoring the games given that it is not implemented, or lose(in the
> event that the protocol was not implemented or implemented
> incorrectly).
>
> -Robin
> _______________________________________________
> computer-go mailing list
> computer-go@xxxxxxxxxxxxxxxxx
> http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/