Author: Don Dailey
Date: 06:43:46 08/16/98
Go up one level in this thread
On August 16, 1998 at 08:45:30, Stuart Cracraft wrote: >On August 15, 1998 at 22:18:24, Jeff Anderson wrote: > >>Can someone perhaps give me a rundown of the piece values used by different >>chess programs? How do small changes piece values in programs affect their >>play? >>Thanks, >>Jeff > >Currently, I use these > > pawn1000 3000 3000 5000 11000 12000 This can get tricky. We have made several adjustments over time to the piece values we use because the evaluation tempers these numbers significantly. For instance, if your bishop has a certain value, and then gets nothing other than mobility bonuses, then what is your bishop value really? Really, there are a few important things to look out for, and you must keep these in mind. When will it trade a bishop for a knight? What is the value of typical pawn structures that this trade typically creates (like bishop for a knight and doubled pawns around king.) Small adjustments can be important and will influence key game critical decisions if they are wrong. What will your program trade to get a rook? Bishop and 1 pawn, bishop and 2 pawns, or does it need to be bishop and 2 pawns with some other positional advantage to consider it worthwhile? All these decisions and many more will be influenced by the values you chose. And to answer your question, if the values are well adjusted, small changes CAN make a big difference in the play, because exchanges and when to make them are often critical decisions in a chess game. Particularly the bishop for knight exchange. Less sophisticated players do not really distinguish the difference, but among grandmasters this decision can and will win and lose games. Here are the values we use in Cilkchess, but they are not necessarily "centered" which means we have inflated or deflated them to account for a preponderance of bonuses or penalties: pawn 96 knight 330 bishop 330 rook 510 queen 940 Bishop pair 24 The values for the pieces were once somewhat lower, but we realized we were giving pawns too many bonuses and had to effectively lower the base value of the pawn by raising these values. These values seem to be approximately right for us based on the evaluation terms we use, they may not work for you but will give you a good idea. I am thinking about performing a standardized normalization procedure to determine the "average" value of the pieces so that over time I am using the same approximate metric. But this is not very straightforward since some terms are not directly tied to an individual piece (like king safety.) I believe the bishop pair is an important terms and consider it strictly part of the material evaluation. I don't consider it a positional term. There are other terms like this but bishop pair is the most important. Me and Larry once considered building a program heavily organized around STATIC chess features (we never did though.) There are several ways to look at the fundamental elements of chess, but if you break them up this way: Material Pawn structure Time Space King safety ... then the static features are Material and Pawn structure, the dynamic features time, space and king safety. One way of defining these terms is to view STATIC features as the ones that tend to not change at all if there is not exchange of material. Obviously material doesn't change but neither do many of the elements of pawn structure. A doubled pawn remains doubled, a pawn island remains a pawn island, isolated pawns and others. Some pawn structure features of course can still change but most of these will tend to involve the possibility of a capture, for instance a passed pawn can never become passed without either an exchange or the offer of one. It turns out these are generally the terms that are the most reliable to evaluate (with some exceptions like passed pawns) and the dynamic features are usually lot's of guesswork on the programmers part. I heavily recommend that you view your evaluation this way, do not view material and pawn stucture as completely separate but as a sort of integrated single term if you will. An incredibly deep search will tend to sort out problems with your dynamic evaluation but a fundamental error in these static features are long lasting and your program will never see the 50 ply ahead to realize trading the bishop for the knight gives you a bad endgame. Of course I admit this is a bit of a simplification. You can still get problems with a cramped position (space) for instance that will also have to live with for a zillion ply and of course any dynamic feature. But remember it is much harder to get these right anyway. - Don
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.