Free online Chess server. Play Chess now in a clean interface. No registration, no ads, no plugin required. Play Chess with the computer, friends or random opponents.
Sign in
  1. Forum
  2. General Chess Discussion
  3. Stockfish 10 problem of evaluation

It seems stockfish 10 is sur-evaluating the positions with 0.2 margin (at least) approximately, compared to stockfish-9.
For example, in a "somewhat better position" for White (carlsbad structure, early middlegame), which gives 51% for White in practice (Chessbase Database), it gives 0.71 at depth 27, whereas stockfish 9 gives the more reasonable (compared to our thinking and to the practical results) 0.50.

0.71 is crazy! This is potentially a major issue.

What do you think about this? (players >2100)


Stockfish 10 now has contempt 24 as default (contempt stands for optimism how many centipawns are your pieces better than your opponents), while Stockfish 9 has a contempt 20. So I guess if you decrease the SF 10 contempt you'll get more similar values. The most interesting feature by now is the way in which SF 10 tries to avoid repetitions and maintain the game alive compared to SF9 but usually this excessive optimism (contempt 24) makes it go for risky lines which at worse result in "inferior drawish" positions when you match it against SF9, but the frequence of getting nice positional wins is higher than the loses it gets. And since SF10 is able to hold when he gets worse, seems like it gained some rating points.

Ok, thank you, but I maintain: 0.71 fpr this position is NOT a good evaluation. We are not close to a pawn advantage, GM would say it is about 0.2-0.3 advantage.

This is a major issue. I think it can have huge impact on chess players (those who analyze with stockfish).

Some people are of the opinion that if you want an "accurate" evaluation set contempt to 0, so the engine doesn't consider that the pieces it plays are better all of the sudden. If you set contempt to 0 you'll get a healthy objective <+0.35 at depth 25. If you let contempt at 24 you'll be seeing +0.71-0.85. But the lines shown for the supposedly overestimated 0.71-0.85 seem better to me.


OK, I understand: the lines are different, more enthousiastic (so better against lower opposition, but tiny weaker against equal or stronger opposition), but the evaluation is wrong...

So, what is in stockfish 10 used in lichess now? and what was used when lichess was using stockfish 9? I mean, the contempt value.

If a position is winning (e.g. SF wins it against any opposition), then what evaluation is correct? The optimistic one or the pessimistic one?

Well, let's call the optimistic the one that seeks complications and the pessimistic the one that trades material off. Since complications are complicated it's more of a gamble sometimes it goes well other times it doesn't go so well, but neither exactly bad. And if the position is winning SF10 will go for something fancier than SF9, it wont trade and rather push for some unobvious sort of positional zugzwang, SF10 sees more subtleties than SF9 to keep pushing but sometimes wants to push too much.

For analyzing purposes, only those positions should be regareded as winning, which stockfish wins against itself.

@Sarg0n: If we have an early middlegame, well known, quiet, the engine cannot claim 0.71, this is absurd. 1 pawn advantage is near "winning", so this 0.71 value is wrong. The problem is also the threshold of "bbig advantage": in this case, we would enter the "big advantage" +/-, which is ridiculous, as the position is +/= without any doubt (once again, sstockfish arrives at 0.2-0.3 after 5-6 moves but when we enter them like Rfe1, h3, and so on).

Thank you, but you don't answer the questions:
- what was the contempt used in stockfish 9 for lichess?
- what is the contempt used in stockfish 10 for lichess?

This value influences the evaluation value of a position (about 0.2), so this question is critical.

This topic has been archived and can no longer be replied to.