Re: computer-go: Learning from existing games

[Don Dailey <drd@xxxxxxx>]

Heikki,

Along  the  same  lines as  your  comment,  I  think that  even  moves
themselves could be "judged"  (very weakly and inaccurately of course)
by the final  results of the game.  I would never  recommend this as a
way  to proceed  but it's  certainly the  case that  the moves  of the
winning side are better than the moves of the loser.

So even though you could never  point to any individual move and claim
it's value  with any  reasonable certainty, you  at least have  a weak
measurement of quality of all the moves taken as a whole.  

I doubt  thinking in terms of  moves instead of positions  can get you
anywhere, but temporal difference learning is the same idea applied to
positions and has been successful in many games.

Don




   Date: Mon, 13 Jan 2003 21:32:33 +0100
   Content-Type: text/plain; charset=us-ascii
   Content-Disposition: inline
   User-Agent: Mutt/1.4i
   From: Heikki Levanto <heikki@xxxxxxxxxxxxxxxxx>
   Sender: owner-computer-go@xxxxxxxxxxxxxxxxx
   Precedence: bulk
   Reply-To: computer-go@xxxxxxxxxxxxxxxxx

   On Mon, Jan 13, 2003 at 05:15:25PM +0000, Nick Wedd wrote:
   > Bluntly, no. 
   > If there were a simple way to tell a good move from a bad move, you 
   > would just implement it in your program, you wouldn't need the training 
   > strategy.

   Careful here! This argument could be used against any and all learning
   algorithms. Yet we know of backgammon that a simple learning program can
   indeed play well, better than any hard-coded evaluation.

   Then again, learning has not turned out to be such a hreat success in go, so
   you may well have a point here...

   -H


   -- 
   Heikki Levanto  LSD - Levanto Software Development   <heikki@xxxxxxxxxxxxxxxxx>