Important evaluation terms in this kind of position

[xr_a_y]

In a recent TCEC game LC0cpu versus Minic, this position was reached

[d] r1b1r1k1/3n1ppp/pq2pP2/2bpP3/1p1N2P1/1P1B4/P1P1QB1P/1K1RR3 b - - 2 20

Minic is absolutely blind to what is going on here and gives a +0.9 score, although it thinks that f6xg7 may be played by white next move which is not a bad idea (but Qf3 is better ; in the game Qe4 was played by LC0, good also)

Next move

[d] r3r1k1/1b1n1ppp/pq2pP2/2bpP3/1p1N2P1/1P1BQ3/P1P2B1P/1K1RR3 b - - 4 21

Minic is seeing something (+3.8) but too late to defend.

Next move again

[d] r3rnk1/1b3pPp/pq2p3/2bpP3/1p1N2P1/1P1BQ3/P1P2B1P/1K1RR3 b - - 0 22

Minic says +7.9, game is over.

I can let Minic think for ages on the first position score is not moving, so is it a missing important evaluation feature (advanced pawn near king, ready to sac bishop, queen still in the game, ...) or horizon effect due to the required scarifies for the attack (a pawn and a bishop).

Another question might be, is the first position (r1b1r1k1/3n1ppp/pq2pP2/2bpP3/1p1N2P1/1P1B4/P1P1QB1P/1K1RR3 b - - 2 20) already too late to be defended by black. I'd like too see a top engine 100 games match from this starting position, I may do that ...

[xr_a_y]

Another question might be, is the first position (r1b1r1k1/3n1ppp/pq2pP2/2bpP3/1p1N2P1/1P1B4/P1P1QB1P/1K1RR3 b - - 2 20) already too late to be defended by black. I'd like too see a top engine 100 games match from this starting position, I may do that ...

Simple answer ...
SF11 versus SF11 30Sec+0.2

Code: Select all

     24 [Result "1-0"]
      1 [Result "1/2-1/2"]

[xr_a_y]

Another question might be, is the first position (r1b1r1k1/3n1ppp/pq2pP2/2bpP3/1p1N2P1/1P1B4/P1P1QB1P/1K1RR3 b - - 2 20) already too late to be defended by black. I'd like too see a top engine 100 games match from this starting position, I may do that ...

Simple answer ...
SF11 versus SF11 30Sec+0.2

Code: Select all

     24 [Result "1-0"]
      1 [Result "1/2-1/2"]

Only one engine was a bit silly ...
Adding Rubi, Komodo10 and DefenChess to the game ...

Code: Select all

     27 [Result "1-0"]

27 white win on 27 games ... still clear ... So even

[d] r1b1r1k1/3n1ppp/pq2pP2/2bpP3/1p1N2P1/1P1B4/P1P1QB1P/1K1RR3 b - - 2 20

is dead lost for black.

[jdart]

Pawn attacks on the King zone are tricky. Sometimes that kind of thing succeeds. Sometimes it doesn't. However, I note current Stockfish seems to like that kind of attack a lot and it has a good success rate with it.

Another factor here is that the Black Queen is on the wrong side of the Pawn chain and so cannot easily come to the King's defense.

[PK]

Can Minic detect the problems faster with null move turned off?

[xr_a_y]

Can Minic detect the problems faster with null move turned off?

No, null move doesn't seem to affect this position too much for Minic. I think this is linked to some static knowledge.

[Alayan]

King danger eval is very difficult, because when you try to get false negatives down, you'll almost certainly increase false positives too. In some positions, the defensive side is a bit late and will have to lose a lot on material or positionally to delay a mate, in others it's just in time and the attacking side has no way to build up the pressure.

KD evals in AB engines typically ignore the potential in pressure build-up and don't account well for defenders.

In your original position, white's pieces are not yet arrayed to attack so KD won't be as high as it should (though the pawn attack can create some KD)

This pattern of a center that is pawn-blocked with one side that can concentrate its attack on a weak king side is common (though I'd guess, much more in long TC games with reasonably good openings than in testing games), but isn't easy to evaluate.

From my limited experience with Stockfish and Ethereal, KD is really hard to improve even when you have example positions to think on. Tricks that would significantly improve either of the KD's eval quality or the ease of coding/tuning it would be a significant boon.