Re: computer-go: Fast scoring program?

[Don Dailey <drd@xxxxxxxxxxxxxxxxxxxxx>]

TT scoring does   indeed capture the  spirit  of  the game.   Japanese
scoring is  just a shortcut way  to evalute the  end result before the
game actually ends.   

In my opinion, you should be training your NN  to understand the WHOLE
game, not just the part humans play  (which does not include that full
end game.)  That is why I say TT captures  the true spirit of the game
and Japanese does not.

I did some experiments with temporal difference  learning in chess.  I
had  these endgame databases  that KNOW when a  game is over simply by
database lookup (despite  the fact  that checkmate  had  not yet  been
delivered.)   I also   would   end the games   when  one  side had  an
overwhelming material  advantage  and  just  score it   infinite (like
checkmate.)

Much  to my surprise,  this hurt   the  learning process  a lot.   The
program could not perform basic checkmates  because it didn't have the
basic technique to move the king towards the corner.  It didn't really
need this during learning because I made it lazy!

You are trying   to teach your  program  to  have subtle  end of  game
scoring  sense,  by using   the japanese  model  somehow, which   just
supplies the answer for it  and does not  let it make the mistakes  it
needs to learn with.

I don't think of  Japanese scoring as  really "scoring", I think of it
as an "evaluation function."  It's exactly the same as taking TT rules
and building a function that is good at guessing the final result when
the game is close to being finished.  If you think of it this way, you
will probably come to realize that it doesn't really matter what rules
you use, you should play the game to the  end, and if you really think
you have to end  early it still doesn't matter  what rules you use  as
long as you have   an evaluator  that is   accurate when the  game  is
"almost" over.

Don






   From: "Grajdeanu, Adrian" <adrian.grajdeanu@xxxxxxxxxxxxxxxxx>

   TT scoring does not capture the sprit of the game. Counting dead stones as
   points instead of prisonners and territory is bad. Playing inside own
   territory is bad. I strive towards jap scoring. But for now, TT will
   suffice. Only later in the evolution my programs need to learn subtelties
   that are not favored by TT scoring.

   > Hi Adrian,
   > 
   > But TT scoring  probably captures the spirit  of the game better  than
   > any other scoring system.  For what you are doing, I'm sure TT scoring
   > is exactly what you want and will yield better results.    
   > 
   > Don