Re: computer-go: Results on self play vs recorded games

[Don Dailey <drd@xxxxxxx>]

I do a LOT of self play.  Depending on what I'm trying to accomplish I
either play  the first n moves randomly,  or I play from  a small book
(database of starting positions.)

If you are  simply evaluating programs, you need  many position from a
book and you  may want to consider letting each  program play both the
black and white stones from each book position if you want to be fussy
about fairness.

Don




   Since the programs were deterministic (when learning was turned off), 
   repeated games wouldn't provide more information.  Is there a consensus 
   on how best to get some variety in testing?  Doing several training 
   runs would probably be good, but that's awfully time-consuming.

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