Author: Robert Hyatt
Date: 18:05:11 07/04/01
Go up one level in this thread
On July 04, 2001 at 18:02:46, Alvaro Jose Povoa Cardoso wrote: >Some time ago I saw here a discussion about adjusting mate scores. >Dr. Hyatt said we should do so only if they were exact scores. >Others have said we could also adjust upper and lower bounds mate scores. >Can some one please elaborate a 'complete' explanation on this? If you have an exact score, and it is a MATE score, then it should be adjusted (I am assuming you mean a hash table score here of course). I store "Mate in N from the current position" scores. If I get a hit on this position from ply=6 later, I simply return current score + 6, since the score is a mate in 6 more plies than the hash table says, since I am 6 plies into the tree when I get the hit. I don't see how you can adjust a "bound" mate score. IE if you know the current position is at least a mate in 5, can you adjust that to be at least a mate in 8 or whatever? When I did this a long while back it caused lots of funny search inconsistencies. > >In my checkers program if I adjust every (exact, upper or lower) mate score I >can get wrong mate positions from hash. >For example: I can get a position that says white wins in 50 moves (!!!) but in >reality it loses in 3 moves. > >On the other hand if I adjust only exact scores those errors disappear, but >another one comes in. >I get mate in 15, then mate in 14, then mate in 13, then mate in 15 again, then >mate in 14, then mate in 13... >I wonder how my program can make progress towards a win with these scores. The first error was similar to what I saw. I never see the second case however, which suggests that maybe you have a bug in your "score adjustment" code... > >Are these errors correlated? > > >As an example here is the output of a (portuguese) checkers game in wich I >adjust only exact scores: > >White: 20 ply Black: 20 ply IDS_Step = 2 > > Nº Move Score Nodes > > 1W 8-19 31978 39,680,403 > 2B 3x13 31979 11,938,167 > 3W 19-14 31980 7,539,580 > 4B 18x11 31981 88,312 > 5W 9x18 31982 20 > 6B 25x14 31983 18 > 7W 5-10 31984 90,491 > 8B 14x5 31985 16 > 9W 1x14x4 31986 113 > 10B 17-13 31987 60,257 > 11W 4-14 31988 1,281 > 12B 13-9 31989 1,485 > 13W 14-32 31990 514 > 14B 31-27 31991 300 > 15W 32-14 31992 84 > 16B 27-22 31993 84 > 17W 14-1 31994 130 > 18B 22-18 31995 59 > 19W 1-32 31996 78 > 20B 9-5 31995 238 > 21W 32x1 31996 55 > 22B 18-13 31997 6 > 23W 1-14 31998 24 > 24B 13-9 31997 54 > 25W 14-19 31998 32 > 26B 9-5 31999 2 > 27W 19x1 32000 2 > > Totals -> 59,401,805 > > >If I adjust every mate score the progress towards a win is perfect, but as I >said before I can get wrong mate positions from hash. >Can someone please help with an explanation (as complete as possible) on the >matter of adjusting mate scores? If I am at ply=5, and I store a "mate in 20 plies" score, I change that to "mate in 15 plies" before I stick it in the table. Because this position is a mate in another 15 plies... > >Thank you all, >Alvaro Cardoso
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.