Re: [computer-go] Minimax with random evaluations

[Don Dailey <drd@xxxxxxx>]

I would be good to compare your results with mine as a code check.
The version I started also had a bug in it so I had to restart too.
I forgot srand()

Which board size would you prefer?  I started with 5x5 so that we
might get relatively deep searches but I will test any boardsize you
prefer.

On the 5x5 boards, it appears that komi of 0.5 gives the most even
black/white split.  This is remarkable since I think black should
always take the whole board with good play.

I'm starting with 1 vs 3 (level 1 is 1 ply search which is equivalant
to playing a random move.)

--------

boardsize 5, simple KO rule, passes included in search, NO stop after
2 passes (but the game ends if 2 passes in a row are actually chosen.)

Evaluation function is:   #define EVAL (8192 - (rand() & 0x3fff))
I'm not sure it matters and could just as easily have used rand()

After the game is stopped, I score using strict chinese scoring
(tromp/taylor rules).  (I don't try to resolve dead groups or
anything.)


Results of 1 vs 3 after a few minutes of running:

  Win perc  Tot Gms  Ave Time  Player 
  --------  -------  --------  ------
    54.280      514       0.0  r3
    45.720      514       0.0  r1

Black wins:      270      52.5 %
White wins:      244      47.5 %


I'll run other depths after I play 1000 games (I play 500 each color
for each side.)

I can change any of this to conform to your testing methodology so that
results can be accurately compared.


- Don






   Date: Fri, 21 Jan 2005 04:32:40 +0100
   From: Gunnar Farnebäck <gunnar@xxxxxxxxxxxxxxxxx>
   X-Virus-Scanned: by amavisd-new-20030616-p10 (Debian) at lysator.liu.se
   X-Amavis-Alert: BAD HEADER Non-encoded 8-bit data (char E4 hex) in message header 'From'
	   From: Gunnar Farneb\344ck <gunnar@lysa... ^
   X-Spam-Score: -4.9
   X-Spam-Flag: NO
   X-Scanned-By: MIMEDefang 2.42

   Don wrote:
   > From the code you included, I see you don't do this.   But isn't the game
   > technically over after 2 passes in most rulesets?

   Or 3 passes or by agreement. I consider either way reasonable for this
   experiment and I doubt it makes much difference.

   > I'm doing the same test with alpha/beta right now.  I'm using a 5x5 board
   > so I can get lots of results quickly.   I'm running with and without the
   > 2 pass rule.

   Any results yet? My code did have a bug. Although it might have been
   insignificant I have discarded the old results and started over with
   an alpha/beta implementation.

   /Gunnar

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