Author: Peter McKenzie
Date: 01:30:12 10/08/99
Go up one level in this thread
On October 08, 1999 at 03:06:56, Bruce Moreland wrote: >On October 07, 1999 at 23:45:47, Peter McKenzie wrote: > >>A partial solution to this is to be extra careful with how you recognise draws >>by repetition when near the root. Ie. only score a node as drawn if it has >>actually repeated 3 times in the tree. What I meant was: 'repeated 3 times (twice previously) in the search tree or actual game. > >This has to be too conservative. > >I think there are two better choices. > >1) Score a position as drawn if it has happened either in the game continuation >or somewhere directly above in the search tree. > >2) Score a position as drawn if it has happened twice in the game continuation >and/or once somewhere directly above in the search tree. I think this is equivalent to saying that nodes very close to the root of the search tree should be treated a little differently (because it is impossible to repeat one of those positions further up the search tree). > >The first method is the classic way of dealing with this. Basically any sort of >repetition is scored as a draw. > >The second method tries to address the problem wherein you'll play back into a >lost game if the opponent had a win, missed it, and is still either drawn or >slightly better. > >Here is an example. In a very early game my program was down two pawns. The >opponent made a reversible move that was a blunder. My program made a >reversible move that won a pawn. The opponent realized he had made a mistake >and reversed his previous move, since that was the best defense. My program >could have taken the pawn, leading to a position where it was a pawn down, but >instead it reversed its move, thinking that this was a draw, but actually it >allowed the opponent to avoid the blunder and remain two pawns up. This is the problem I was trying to describe. > >The drawback to the second method is that you get a lot of repetitions in actual >games, which can lead to enraged opponents and possibly some 50-move draws in >endings where the program would be forced to make progress otherwise. > >bruce
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.