Author: Robert Hyatt
Date: 20:32:42 01/28/98
Go up one level in this thread
On January 28, 1998 at 21:10:04, Don Dailey wrote: >Hi Bruce, > >My own testing on earlier chess programs indicated a small but definite >improvement by weighting moves near the root. I used depth from leaf >nodes and use d * 2^n. I didn't do it for Bob's reason though. I felt >that even if you felt the moves should be equally weighted, they were >not since on the 2nd ply you have very little representation. And >yet these are just the moves you would like to get ordered right. > >But your point is well taken, it's a rough idea and there may very >well be better algorithms. > >Bob mentions that he gets little from killers. I'm sure this is because >he already has the history heuristic. If I implement history heuristic >first, I might think killer help very little, and visa versa. > I may not have worded this very well. In Cray Blitz I *only* used killers, but I did them differently than most everyone else, in that I kept two killers per ply, but I didn't only try just those two at a ply. I also tried the killers from other odd or even plies (depending on this ply being odd or even) at this position as well. Made a big difference. In Crafty I first did killers, then did history and took killers out. Way later I re-added killers but the size of the tree did not go down, but the time saved by avoiding move generations (because a killer move is complete and can be legality checked, but history moves don't exist in such a form). So I got no tree reduction because I already used the history heuristic... but I did get a 10% speedup by avoiding move generations. But had I not used history already, killers are a huge improvement over nothing. I don't remember how much better history was than killers. I might even experiment with no history to see as it is somewhat more computationally expensive to do the selection sort than to manage killers. I'll put this on my list to test... >I remember reading an article years ago trying to show how each >additional piece of knowledge helped the program, but the order the >heuristics were presented significantly change the outcome. It turns >out the first thing you put in the chess program seems like the one >that helps the most. i.e. centralization may not seem important if >you already have mobility, but if you do not, centralization is a HUGE >improvement. > >The human mind is the same way. Studies show that we are influenced >strongly by the order information is presented to us. We tend to use >the >first piece of information as a standard to judge the next by. > >- Don > > > >On January 28, 1998 at 18:02:40, Bruce Moreland wrote: > >> >>On January 28, 1998 at 17:05:23, Robert Hyatt wrote: >> >>>from my early testing, you have to weight based on distance from the >>>root... because I trust a move backed up as best near the root rather >>>than one backed up as best near the tips. >> >>Possible topics: >> >>1) Whether or not it really matters how far the move is from the root, >>since this is just being used to get a rough idea of what moves tend to >>be good. I'm not sure it's intuitive that closer to the root is better, >>since most applications of this heuristic will occur further from the >>root, why not weight moves that tend to cause cutoffs way out there as >>higher? >> >>2) Whether the moves that tend to have high history values in one >>position also tend to have them in other positions, and if so, perhaps >>some sort of static weighting system might be more efficient. >> >>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.