5

I'd like to know about computer games of some recent tournaments where two top engines diverge in evaluation of a position, in the sense that both think are winning the game (or both losing). I saw a match in the last CCC where Stockfish as white had +3 in his evaluation and Leela -3, that was the maximum I've ever seen. This is interesting because usually the two engines seem to agree about which one is winning (or at least one thinks it is winning/losing and the other thinks it is a draw) even if there's a gap between the two evaluations, however this happens the same in the divergence's situation but usually it takes some time.

Best by now I found :

https://www.chess.com/it/computer-chess-championship#event=ccc-rapid-2021-elite-round&game=508 : about +1 and -1

It seems I had an hallucination about the game between Stockfish and Leela

5
4

I finally found through a script several games where (not so big) divergence occurs!

I analized roughly 3000 games with time control 15+3 and just 10 show a notable divergence.

It's interesting and intuitively correct that this happens a lot more when black wins! Indeed in the list below there are 3 wins by white, 6 by black and 1 draw even if black wins much more rarely than white between engines.


CCC main league 2021 :

Stockfish classic - Stooflees : +1.41 vs. -2.53

Komodo - Rubi : Long divergence with pick +1.22 vs. -1.35

CCC elite round 2021 :

Stooflees - Ethereal : +0.84 vs. -1.75

Stockfish - Stooflees : +2.91 vs -0.27

SlowChess - Stoofvless : +2,76 vs. -1.04

Ethereal- Stoofvless : Long divergence with pick +1,07 vs. -0.97

CCC final 2021 Stockfish - Lc0:

Game 9 : Long divergence with pick +0.46 vs. -2.19

Game 409 : +0.37 vs -1.99

CCC 2021 Leela vs Dragon round 2 :

Game 2 : Long divergence with pick +0.34 vs. -1.04

Game 68 : Long divergence with pick +0.89 vs -1.13. This game is strange, the two engines don't agree with drawing the game, 307 moves are played!

2
  • Sounds like great work! We can agree that these positions are very important indeed. How did you download the games together with evaluations from both engines? Jul 9 at 20:17
  • Initially I thought I would have to write some HTML request but in the end I found that in the download button of one game of a certain event on chess.com you can go on the event section and download all the games of that event :-)
    – Tortar
    Jul 9 at 20:43
1

This happens, but it's rare and is getting rarer. Here's a very recent game between Stoofvlees and Komodo Dragon displaying such eval divergence. The discrepancy was only there for a few moves, too.

Edit: As I said this does not happen often and it's getting rarer. Examples of Stockfish-Leela games in the past where it happened are this game from Season 15 and this one from Season 18; both games are over a year old at this point and involve versions >100 elo weaker than today's.

5
  • this case is not so interesting because the two evalutions are always of the same sign
    – Tortar
    Jul 6 at 14:11
  • @Tortar not exactly - on move 138 Stoofvlees was +ve and Komodo Dragon was -ve.
    – Allure
    Jul 7 at 0:27
  • yes, but very tiny difference
    – Tortar
    Jul 7 at 14:24
  • @Tortar we're lucky there's even divergence in the first place. If you can find a position with a substantial eval divergence between today's Stockfish/Leela, I'm sure Jeroen (the person who makes the TCEC opening books for the suprefinal) will be interested.
    – Allure
    Jul 7 at 22:30
  • I found some interesting games between the two, but not so big gaps
    – Tortar
    Jul 8 at 19:22

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.