Here are several conclusions based on my Komodo work:Dann Corbit wrote:Is there any scientific measurement that yields reliable pawn value estimation?
It seems that the standard methods are very good at showing the relative values of pieces (compared to one another), but not pawns.
1. It is best to define a constant value for the knight, because we have found that there is no measurable benefit to having a higher or lower endgame value than opening value for the knight, so we just keep it a constant. Of course this might not be true for all engines, it depends on how positional terms are scored. In Komodo, the bishop value is also constant, because although the bishop gains a bit relative to knights in the endgame, this is accounted for by increased mobility and some other terms. Due to the huge importance of the bishop pair it is best to define the knight as the constant piece. The rook and the pawn both rise sharply in value in the endgame, so they are not suitable to hold constant; the queen rises much less sharply but is not "level".
2. Because the pawn value rises so sharply as the pieces come off, it is not very useful to talk about its "average" value (relative to knight always). I use a ratio of 3.5 in teaching humans, but in Komodo it probably averages closer to 4 (pawns per knight).
3. The nominal value of the pawn is also severely affected by the fact that in general three pawns are part of the king's shelter and hence rather valuable. In Komodo the nominal opening value of a pawn is surprisingly low, but that's largely because the three king-protecing pawns get large bonuses. To be honest I don't fully understand why the nominal pawn value is as low as it is in Komodo; in actual games it scores positions with multiple pawns for a piece about right.
4. Regarding the recent Russian study which concluded that in the games of Carlsen and Anand a knight was worth less than 2.5 pawns, this is clearly nonsense. Perhaps they only sacrifice a piece for pawns when there is substantial positional compensation not detected by the methodology. I know that when I did my own study of master games back in 1999, I consistently found that in human practice a knight was worth on average a bit more than three pawns, which is the consensus of chess literature. I suggest that all chess engines should define a knight as something like 325 centipawns so that output will have a constant meaning for all engines.