Author: Robert Hyatt
Date: 06:33:32 09/11/99
Go up one level in this thread
On September 11, 1999 at 06:53:00, Dezhi Zhao wrote: >Posted by Robert Hyatt on September 10, 1999 at 00:19:37: > >[snip] > >>If you recall the discussion here a couple of weeks ago, I reported that I store >>absolute mate scores (EXACT scores) in the hash table, and that I adjust them >>so that they are always stored as "mate in N from the current position". This >>has always worked flawlessly for me, and still does. > >>For bounds, I once tried adjusting the bounds as well, but found quirks, and >>left them alone. Wrong answer. To fix this mate in 4 problem, I decided to >>adjust the bounds as well, but I now set any bound value that is larger than >>MATE-300, by reducing it to exactly MATE-300, but still using the "LOWER" >>flag to say that this is the lowest value this position could have. For bound >>values < -MATE+300, I set them to exactly -MATE+300 and leave the flag as is. > >Hi! > >If I understand correctly, you relax the bound mate scores to safe values, so >that these bounds will not produce cutoffs when compared with other mate scores >(esp. of EXACT type), but the bound will still generate cutoffs when compared >with non-mate scores. > >I used to adjust the bound mate scores and the exact mate scores in the same >way, and have not found any problems so far. > >Therefore, my question is: >Why the adjusted bounds should not be compared with other mate scores and thus >produce cutoffs? > >It seemed to me that the relaxed bounds would produce less cutoffs than the >adjusted ones which are tighter. However, when I tried out your relaxed bounds >on some mate positions, I found that your relaxed bounds save a lot (~ 12%)! >Thanks to Dr. Hyatt! I'm not sure why it saves a lot, but I do understand why it is better. The position from Steffen found a score (Mate in 4 but it wasn't forced) and this became a bound. Later it found a mate in 5, but since this was worse, it stored >= Mate-in-4 for that position. Later it found that same position 4 plies deeper and said >=Mate-in-4 which is now wrong... It is really >= Mate-in-6 since we are 4 plies deeper than before. This made crafty find a mate in 4 in the actual game, when was really a mate in 6. I tried it with the 'relaxed bounds' and it had no problems I could find with over 500 mate-in-N test positions... It was harder to find than to fix, naturally. :) > >Any explainations? Can we find a even better way of using mate bounds? None. I am going to experiment with adjusting the bounds (Mate) as well. However, I did this when I first added hashing to Crafty and ran into something that was a real pain. It is probably given in the comments in main.c, but I want to go back and understand why what I did was not working. Might be that there were other bugs at the time... > >Dezhi Zhao
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.