22

I've just played a game on Chess.com, and I'm surprised that the engine considers a promotion that I made a huge blunder. Like, my position was rated -62.8 (basically a won game), and after I promoted my pawn to a queen the rating dropped to -6.40. That's more than 50 points with a single move!
This is the position:

[FEN "8/8/2B5/p7/8/p4p2/P2K1kp1/8 w - - 0 1"]

1. Be4 g1=Q (1... Kg3 2. Kd3 g1=Q 3. Kc4 f2 4. Bd3 f1=Q 5. Bxf1 Qxf1) 2. Bxf3 Kxf3 3. Kc2 Qc5+ 4. Kd3 Qb5+ 5. Kd2 Qb2+ 6. Kd1 Qxa2 7. Kc1 Qb2+ 8. Kd1 a2 9. Ke1 a1=Q# 0-1

I played g1=Q, but the engine says I should have played Kg3 instead, which would have kept the rating at -62.8. I don't understand why. My pawn in f3 was threatened by the bishop, but it was protected by the king. The white king was unable to come closer, because of my king. And the pawns on the left were irrelevant.

The engine's move did lead to another pawn being promoted... Only to be captured by the bishop, which in turn was taken by the first queen. So after a few moves the situation was identical: king and queen vs king (plus the pawns on the side). And even with my "blunder", winning was easy.

Why does the engine find such a huge difference? Should I just ignore this, or am I missing something important?

3
  • 3
    I am not an expert at these chess algorithms, but it seems if you have the queen you induce a possibility of stalemate. That may be what is the idea behind the large drop in points. Aug 30 at 15:26
  • If you let the computer analyze for a long time, I'm sure that the -6 would turn into mate in 10 or more moves.
    – Issel
    Aug 31 at 20:46
  • 4
    Computers are not designed to evaluate positions with such big imbalances, but rather to be precise at more "common" positions
    – David
    Sep 1 at 11:21
30

50 points sounds huge, but there actually is no meaningful difference between -6.4 and -62 in such an endgame.*

  • 6.4 is more or less "Black has a queen for a bishop (but I haven't found a way to make progress yet)"

  • 62 on the other hand is "I have found that way to a position that is 100% won (but I haven't found a forced mate yet)"

Why does the engine rate both moves differently? That might just come down to random chance: Since Black has a winning position either way, almost all moves along the calculation tree are equal to the engine's eye and it loses its guidance so to speak which branches are best to pursue. At the same time, the tree is still relatively big (White has a lot of ways to shuffle the bishop around, etc.). So the tree gets deepened in one more or less random direction and the engine will keep expanding from there (downwards and sideways) because all moves it finds are keeping the advantage.

By chance, the trees for 1...Kg3 and 1...g1=Q start growing a bit differently. They will of course converge eventually (to the mate in 11 Allure mentions) if the engine gets enough time to calculate all of the tree. But in for example automatic game analysis, it often doesn't (20-30 seconds for all moves = not a lot of time per single move). In this case, the search for both moves just happened to be cut off at different states of incompleteness (with different temporary evaluation scores).

Which is a bit annoying if it taints your statistics for an otherwise well played game - but you can safely ignore it.

*For this reason, at least Lichess internally caps the scale at +10 and -10 so that a drop from -60 to -10 would not be labelled a "blunder". Don't know how chess.com handles this, maybe the drop from -6.4 to -10 was still enough for a "blunder" though.

22

This sounds like a bug. Using Lichess's analysis engine, it takes Stockfish 14 less than ten seconds to find a mate in 11 moves. Both 1...Kg3 and 1...g1=Q lead to mate in 11.

So there should be no difference in which move you start with, even objectively (of course as a human I'd go 1...g1=Q without thinking).

6
  • 7
    I've frequently seen chess.com mark certain moves as "blunders" that SF on Lichess thinks are perfectly fine. I'd guess the "standard evaluation" chess.com does after every game is not that good (it is done fairly quickly, so this is not very surprising). (I don't think there's a big difference in how conservative the engines are with calling something a blunder) Aug 30 at 10:06
  • @Discretelizard using the version of Stockfish at www.chess.com/analysis, I get M14 in a few seconds for both 1...g1=Q and 1...Kg3, as well. It's three moves slower than Lichess Stockfish (which I'm pretty sure has the newer version), but the two moves are still equally good.
    – Allure
    Aug 30 at 13:17
  • 3
    Right, but the analysis that chess.com does after finishing a game -- to show you how many blunders you made -- takes about one second for the entire game. (at least, it does for me, with a free account) This is the analysis that is frequently not as good as an ordinary analysis on lichess (or even on chess.com, as you note). Given that this analysis is done that quickly, the lack of quality is not surprising. Aug 30 at 14:00
  • 2
    Both Lichess and Chess.com use the web version of Stockfish which appears to be maintained by Niklas, one of the Lichess contributors. Chess.com only runs the engine at depth 18 or 19 (can't remember) per position though there isn't any reason they have to do that and is most likely to make it quicker. Lichess is also likely to have a more recent / stronger version of the engine given that its contributors are the ones working on the web versions of the engine.
    – mowwwalker
    Aug 30 at 21:02
  • 2
    Also @Discretelizard, the classification isn't something provided by the engine. The engine is just giving scores for positions (plus info related to depth and nodes searched) and programs like Lichess or Chess.com are classifying moves based on their own criteria.
    – mowwwalker
    Aug 30 at 21:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.