Author: Robert Hyatt
Date: 10:36:28 01/27/04
Go up one level in this thread
On January 27, 2004 at 13:22:00, Ed Trice wrote: >Hello Dr. Hyatt , > >> >>Crafty needs several things. >> >>1. Better evaluation tuning. I simply have not invested the time in this to >>optimize the scoring well enough. >> > >I was wondering if you were aware that Crafty was fine-tuned using an automated >procedure that was published in: "Advances in Computer Games 10: Many Games, >Many Challenges" > >The was the first paper appearing in this book, by Gomboc, Marsland, and Buro. >It is entitled "Evaluation Function Tuning vis Ordinal Correlation." Yes... however, if you look at the results, it is not clear that this could be called a complete success. IE some terms seemed to make little difference in their testing, while others never seemed to reach a stable value... > > >>2. No forward pruning except for null-move and futility. I haven't given this >>aspect a lot of thought as I have been more interested in the parallel >>programming issues. > >The ProbCut and MultiProbCut selected search enhancements were also used to show >improvements over a vanilla version of Crafty in this book. While the gain was >not outstanding, it did demonstrated that MPC could be tweaked an applied to >chess (Othello was a successful testbed previously). I have looked at that, and I think someone even did a crafty implementation a while back and it was discussed a bit on the Crafty List. However, as always, there are lots of interactions. I liked the futility stuff better myself, from the results I saw. > > >>3. Better evaluation, period. It knows a lot about chess. And particularly >>about endings. There is still much to do here and I add things as I have time >>and see a problem. > >I see Crafty uses flat weights for the pieces and then a BADTRADE criteria to >discourage certain kinds of trades. Because the number of such trades goes >through the roof in Gothic Chess, I needed another technique. > >So, I had the piece weight change as a function of the pawn population. This >does wonders for the "rook trade" problem, where B + N appears weaker than R + >P, so active minors are usually disposed of for the R + P. Of course, not having >the rook in the late opening or very early middlegame is not nearly as bad as >being without minors that might otherwise have very good prospects. > >Since you can tweak the weight of any chess piece very granularly, you can have >the Queen gradually gain strength as pawns come off of the board, and have >Knights weaken. Bishops can usually stay the same since they are given mobilty >bonuses. I do that indirectly, as much of the scoring is tied to the number of pieces or number of pawns remaining... > >--Ed
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.