Material tables

[brianr]

I have been wondering about using GM game databases a bit and am concerned about the results being tainted by potential tactical mistakes.

For instance, Tinker is clearly not a very strong engine, but on today's hardware when observing human IM/GM games on ICC, even Tinker often finds several tactical midgame errors (looking 30-40+ ply with all extensions in a 15-18 ply full-width search, single thread searching). Of course, vastly stonger engines like Rybka find many more "mistakes" earlier.

So, I'm wondering if it would be worthwhile to re-analyze these material combinations using some database with only computer v computer games from the last few years. Yes, they may have the typical computer endgame blind spots, but if the game is more or less already decided due to a significant material advantage, that is what we are looking for in this case anyway, I think.

CCRL and CEGT and the other top testing collections should have at least as many games as were used a decade ago (350K).

[lkaufman]

There might be (and I believe there are) differences in the values you would get from engine databases compared to human databases, but tactical mistakes are not the reason. There are far more random factors that affect databases results than the occasional blunder. Engines and humans are not equally good at all things, so some differences are to be expected. For example, the Exchange appears to be worth more in engine databases than in human ones. Same with the queen. Perhaps humans are better at coordinating the actions of the weaker pieces; this concept of piece cooperation is absent from chess programs because it's so hard to define.