TalkChess.com

Larry's Post over in tournament forum got me thinking a little bit about evaluation philosophy. It seems as if all engine still follow the typical "human" scheme of evaluating pieces as having an inherent value (1,3,3,5,9 or some slight variant there of). While this seems to be a fairly good rule of thumb for the majority of positions one can easily find positions where this actually very detrimental such as,

[d]5r2/7k/5ppp/4p1Q1/5PPP/8/6K1/8 w - - 0 1

A static evaluation of this position leads to the conclusion that white is ahead (Stockfish thinks as many as 4.5 pawns!). However obviously white is doing very poorly. However, If a "materialless" were to be tuned correctly the function would see that the queen is not doing anything for white and would evaluate its actual worth as closer to 0 thus giving black the edge. Obviously problems like this are solved by the search function but as we all know the two work hand in hand to make the engine play smart chess. Most positions don't make a queen look THAT stupid. In fact is it very difficult see how the position could have arose in an intelligent manner. Anyways, my question is if anybody has attempted to create such a function as it seems that it would be better at guiding the search away from such positions, where such a drastic evaluation discontinuity occurs.

While i can't claim to have tried anything along that vein a completely material-less evaluation seems it would likely struggle - i.e. many positions where a piece 'isn't doing anything' the piece still has value because some day it will once again see life, bad bishops are frequent culprits in this regard where they appear a liability and then a little while later the position gets cracked open. For cases like the one you listed some more general trapped-piece logic to at least somewhat correct the mistake the material eval would make might be appropriate (though not needed for this specific case perhaps it might be caught by something more useful - safe queen mobility of zero can't be a good thing!).

A significant correction for 'knight on a bad tour' might also be interesting - there was a benko gambit line in the last TCEC where both games black wound up totally hosed by a knight that became worse than useless. Something related to multimove safe mobility for the knight would probably be involved in catching it (i.e. don't just count the safe squares the knight can move to immediately but also the safe squares that can be reached from those squares in order to be able to apply a more significant correction - one retreat square isn't necessarily bad - only a few squares reachable within 3 moves for the piece, none of which constitute active outposts and that could be a good indication that knight is miserable.)

It is kind of a given that static evaluation only works in quiet positions. And the position you give as example is not quiet. It is not so much that the Queen isn't doing anything, it is that you will lose it on the next move.

Any attempt to statically evaluate the tactics that will make the position quiet will turn out horrendously inaccurate. It simply cannot compete with a Quiescence Search, the average Chess position is just too complex.

Many engines do recognize patterns for trapped pieces, however, such as the infamous Na8 or Ba7, largely discounting the trapped piece. I would not call that materialless evaluation, however. You count the material, but you have a compensating penalty for the trapping pattern. It should be seen as reversal of the 'burdon of proof'. Normally you assume that you will be able to hang on to any material you have. For a piece with no safe moves, in a spot deep in enemy territory where it cannot be easily protected, this assumption is wrong more often than not. So it works better to assume that you will lose the piece, and only count it if you can see within the horizon that it can break out of the trap (i.e. reach a position that no longer matches the pattern).

If you want to elaborate on that idea, not using a small number of very common patterns, but by a more general method, you could use future mobility. If that is very low, you might give a penalty based on proximity of enemy pieces.

Hi,

in "Elements of positional evaluation" by Dan Heisman he says that what usually is considered the material value of a piece is actually determined by an expression of mobility, the potential mobility of a piece.

The effective value of the piece in a given position is then a combination of actual and potential mobility.

I think this covers it pretty well.

Thomas...

cetormenter wrote:A static evaluation of this position leads to the conclusion that white is ahead (Stockfish thinks as many as 4.5 pawns!). However obviously white is doing very poorly. However, If a "materialless" were to be tuned correctly the function would see that the queen is not doing anything for white and would evaluate its actual worth as closer to 0 thus giving black the edge.

How would that give you a "materialless" evaluation? It seems you mainly want the material value of each piece to be evaluated as accurately as possible.