[computer-go] Noise reduction in alpha-beta search

[Chrilly]

Ingo Althoeffer has published some time ago a theoretical article about this idea. He called it "telescope" evaluation. According his theorectical findings is the error propagation not better than the usual approach.
K.Chen proposed a similar approach. Use the mean of the last and second-last evaluation. I tried this. It makes the search more stable. With the normal search we have strong odd/even effects. At odd depths Suzie evaluates the position (considerable) higher than at even depths. At odd depths Suzie has 1 move more in the variation than the opponent. This odd/even effect disappears with the K.Chen mean. Also the search depth increases, because move ordering is more stable. But the result in the autoplay-matches was worse. One gets effects like the following: 1 Ply before the horizon the opponent makes a threat. The evaluation goes down considerable, but there is a defense move and at the horizon nothing has happened.. But according the mean the programm is still in some trouble. Or if one reverses the role, Suzie would like to play the threat-move. This scheme increases therefore even the potential of the programm to "cheat". 

Chrilly

  ----- Original Message ----- 
  From: compgo123 at aol.com 
  To: computer-go at computer-go.org 
  Sent: Monday, April 09, 2007 4:48 PM
  Subject: [computer-go] Noise reduction in alpha-beta search


  I think following is a way to reduce the noise in alpha-beta search. Instead of using the evaluation values, use the cummulative evaluation values. That is the sum of the evaluation values of each node of the playing path under examination.


  Daniel Liu

------------------------------------------------------------------------------
  AOL now offers free email to everyone. Find out more about what's free from AOL at AOL.com.



------------------------------------------------------------------------------


  _______________________________________________
  computer-go mailing list
  computer-go at computer-go.org
  http://www.computer-go.org/mailman/listinfo/computer-go/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://computer-go.org/pipermail/computer-go/attachments/20070409/18783d07/attachment.htm