Re: [computer-go] Monte Carlo Go

[Don Dailey <drd@xxxxxxx>]

This is my understanding which is an educated guess based on implementing
similar ideas in one version of my program.

  1. Statistics are are gathered for each point for each player.  If
     it is hit more than once by the same player, it's not an issue
     and you count it as a single data point.

     If it is hit by both sides, count only the first player to hit
     it.  This makes sense as the data is far more relevant for the
     first player who plays to a given point.
     
  2. It doesn't make much sense to extend the list beyond the number
     of points on the board, so for example on a 9x9 board I assume
     there is an 81 element move list.  After annealing this list, you
     play on the first point possible.

There were several ambiguious points in that paper.  However, I like
Bouzy's formulation of a very similar program which is described in a
paper called, "Developments on Monte Carlo Go."  Perhaps you've
already seen that one, but he seems to have taken the good ideas from
Monte Carlo Go and discarded the useless ones.

I have also done much experimentation with this approach and have reached
many of the same conclusions that Bouzy did.   My own program has some of
it's own novelties and twists on this basic idea.   

I think this is an interesting approach but it does need a lot of work
beyond what has been written.  

- Don
     



     


   Date: Fri, 28 Jan 2005 12:12:06 -0800
   From: Peter Drake <drake@xxxxxxxxxxxxxxxxx>

   In his original paper on Monte Carlo Go, Bernd Brügmann says:

   "Each player decides beforehand in what order he wants to play his 
   moves (taking into account that pieces may be played several times onto 
   the same field after captures)."

   How is this taken into account?  Can a move be listed more than once?  
   If so, how long must the move list be?

   Later:

   "Each move is assigned the average value of all the games in which it 
   was played."

   What if it was played more than once in a game?

   Peter Drake
   Assistant Professor of Computer Science
   Lewis & Clark College
   http://www.lclark.edu/~drake/
   _______________________________________________
   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/