Re: [computer-go] Minimax with random evaluations

[Don Dailey <drd@xxxxxxx>]

I see your point.

Actually my program DOES forbid plays inside it's own single point
eyes but I meant to turn that off for this test and didn't.  So I must
restart.

I don't think this will actually matter because I am counting the
moves correctly in the evaluation, it's just in the search that I
fobid moving to my own single point eyes.  The reason it won't matter
is because the evaluation should prevent it from ever wanting to move
to a single point eye anyway so with the evaluation it doesn't matter
whether I forbid this or not in the tree search.  It should still play
stronger because it will actively create eyes for itself and destroy
them for the opponent.  What a silly way to program making eyes!

Ok, after restart, I have this and it's pretty much the same so far:

  LMC(5)  vs  RM     96.7%  (30 games)
  LMC(5)  vs  RE(5)  90.2%  (51 games)



- Don



   From: "David Fotland" <fotland@xxxxxxxxxxxxxxxxx>
   Date: Mon, 24 Jan 2005 20:02:35 -0800

   Are you forbidding plays inside your own single point eyes?  If so, then
   when one player has only single
   point eyes left, then the legal move count is the territory score :)

   David

   > -----Original Message-----
   > From: computer-go-bounces@xxxxxxxxxxxxxxxxx 
   > [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx] On Behalf Of Don Dailey
   > Sent: Monday, January 24, 2005 7:59 PM
   > To: drd@xxxxxxxxxxxxxxxxx; computer-go@xxxxxxxxxxxxxxxxx
   > Cc: computer-go@xxxxxxxxxxxxxxxxx
   > Subject: Re: [computer-go] Minimax with random evaluations
   > 
   > 
   > 
   > I forgot to mention that I'm running on 5x5 boards.   
   > Everything else is
   > under the same basic conditions I posted earlier.  
   > 
   > And I might as well give the updated numbers:
   > 
   > 
   >  LMC(5) = Legal Move Count at 5 ply
   >  RE(5)  = Random Evaluation at 5 ply
   >  RM     = Play a random legal move.
   >  
   >    LMC(5)  vs  RM     97.5%  (120 games)
   >    LMC(5)  vs  RE(5)  90.5%  (126 games)
   > 
   > So as we both thought, legal-move-count is better than random 
   > evaluation which is better than random move.
   > 
   > Hmmm, I never thought of mobility as being a GO evaluation feature!
   > 
   > - Don
   > 
   > 
   > 
   > 
   > 
   > > > Ok, I will put on a legal-move-count evaluation test to see what 
   > > > happens.
   > > 
   > > And I'm running it now.  I implemented it the slow naive way, 
   > > generating a set of legal moves twice, once for each side.
   > > 
   > > I'm pretty sure this is equivalent to simply counting single point 
   > > eyes.
   > > 
   > > I don't have enough games to give the full picture but after a few 
   > > minutes the legal move count version is easily dominating both the 
   > > random evaluation and the random mover.
   > > 
   > > 
   > > LMC(5) = Legal Move Count at 5 ply
   > > RE(5)  = Random Evaluation at 5 ply
   > > RM     = Play a random legal move.
   > > 
   > >   LMC(5)  vs  RM     97.4%  (38 games)
   > >   LMC(5)  vs  RE(5)  87.1%  (31 games)
   > 
   > _______________________________________________
   > 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/