Komodo blunder

[hgm]

[pgn][Event "EXPERT-PC, Rapid 15m+15s"]
[Site "EXPERT-PC"]
[Date "2015.09.15"]
[Round "1"]
[White "Stockfish 150915 64 BMI2"]
[Black "Komodo 9.2 64-bit"]
[Result "1-0"]
[Annotator "7.93;6.86"]
[Setup "1"]
[FEN "1nnnknn1/3ppp2/2n2n2/8/8/8/3PPP2/1Q1QK1Q1 w - - 0 1"]
[Plycount "65"]
[TimeControl "900+15"]

1.Qa4 { 7.93/27 28 Both last book move } 1...Nd6 { 6.86/22 11 } 2.d4 { 8.00/31 27
(Qa8) } 2...Nde6 { 6.37/33 44 (e6) } 3.Qa8 { 8.35/33 49 } 3...Ng6 { 6.21/34 23 } 4.e4 { 8.05/
35 44 (Qh1) } 4...Ndxe4 { 4.38/27 31 } 5.d5 { 8.05/32 3 } 5...Ned8 { 4.20/27 43 } 6.dxc6 { 7.98/33 31 (Qd3) } 6...Nbxc6 { 4.38/28 18 } 7.Qd3 { 7.91/30 12 } 7...Nh6 { 4.51/33 23 (Nd6) } 8.Qh3 { 7.86/31 33 } 8...Nhg4 { 4.39/34 17 (Nhg8) } 9.f3 { 9.73/37 25 } 9...Nef2 { 4.36/35
16 (Ngf2) } 10.Qhf1 { 9.61/35 25 } 10...Kf8 { 4.35/36 23 } 11.fxg4 { 9.74/34 5 (Kd2) } 11...N2xg4 { 4.35/34 26 } 12.Qh3 { 9.39/34 139 (Kd2) } 12...Kg7 { 4.24/32 15 } 13.Qc3 { 9.39/
32 19 } 13...N4e5 { 4.18/32 20 } 14.Kd1 { 9.32/34 26 } 14...Ne6 { 4.18/36 19 } 15.Qc8 { 9.32/
35 28 (Qca3) } 15...Ncd4 { 4.19/34 14 (Ned4) } 16.Qh3 { 9.39/34 37 (Qa3) } 16...Nfg4 { 3.88/
27 13 (Neg4) } 17.Qhf1 { 8.75/33 63 (Qa3) } 17...f5 { 3.63/27 16 (d5) } 18.Qa8 { 7.97/
33 43 (Qc3) } 18...f4 { 3.57/28 22 (Nef3) } 19.Qh3 { 8.45/29 25 } 19...Ne3+ { 3.50/29 24 } 20.Kd2 { 8.55/31 3 } 20...N3c4+ { 3.60/32 28 } 21.Ke1 { 8.44/31 31 } 21...Kf6 { 3.63/33 19 } 22.Qh5 { 8.46/28 3 } 22...Nf5 { 3.71/33 28 } 23.Qa1 { 8.58/34 34 } 23...Ng3 { 3.72/33 34 } 24.Qf3 { 8.58/33 3 } 24...Nf5 { 3.76/33 25 } 25.Qe4 { 8.64/36 20 (Qfg4) } 25...Ned4 { 4.03/35 57 } 26.Qaxd4 { 8.93/41 20 (Qgxd4) } 26...Nxd4 { 4.73/32 25 } 27.Qexd4 { 10.50/40 85 } 27...d6 { 4.92/
34 35 } 28.Qf1 { 10.50/37 3 } 28...Kg5 { 6.02/37 61 (Kf7) } 29.Qdg1+ { 11.50/37 32 } 29...Kf6 { 6.32/35 17 } 30.Qfg2 { 11.51/36 4 } 30...Kg7 { 250.00/36 37 } 31.Qh3 { 11.96/40 28
(Q2h1) } 31...Ne3 { 250.00/44 136 } 32.Ke2 { 13.55/37 19 (Qg5) } 32...Kf6 { 250.00/47 78
(N3g4) } 33.Qh5 { 123.26/42 37 }
1-0[/pgn]
I did not play this game myself, but saw it posted on chess.com. Komodo manages to lose a Knight in a Pawn fork on move 5, for two Pawns. It seems to do this entirely voluntarily. Its score seems to improve from -6 to -4 by this. On move 2 it repeats that 'trick', this time by throwing away the second Knight for a single Pawn, without its score taking a dent. Its score even rises to -3.6 after that. The initial position was badly lost for Stockfish, but after throwing away two Knights Komodo of course loses without a chance.

How can it think trading 2 Knights for 3 Pawns is a good deal? It cannot think 3 connected passers are worth 2 Knights, can it? That seems a gross over-evaluation of passers, especially passers in the King's Pawn shield, in the presence of three enemy Queens.

So what is the problem here that causes this misevaluation that makes Komodo blundering the game away?

[Vinvin]

May be it's an unforeseen extreme unbalanced position ...

[hgm]

Could it think Knights are not worth anything because it already has so many of them? That doesn't seem the case, because the total score is not that far from balanced. (Stockfish has double the score.) It could not get so close if the Knights on average weren't worth at least 1/3 of a Queen.

It wouldn't be a clever thing to think anyway; against 3 Queens you would need all the Knights you can muster...

[Guenther]

Could it think Knights are not worth anything because it already has so many of them? That doesn't seem the case, because the total score is not that far from balanced. (Stockfish has double the score.) It could not get so close if the Knights on average weren't worth at least 1/3 of a Queen.

It wouldn't be a clever thing to think anyway; against 3 Queens you would need all the Knights you can muster...

I must say I don't trust such games a lot(online anonymous games).
Who knows how much Komodo was *optimized* here by some changed
params?

Guenther

[bob]

Could it think Knights are not worth anything because it already has so many of them? That doesn't seem the case, because the total score is not that far from balanced. (Stockfish has double the score.) It could not get so close if the Knights on average weren't worth at least 1/3 of a Queen.

It wouldn't be a clever thing to think anyway; against 3 Queens you would need all the Knights you can muster...

Knowing that

(1) they use a material imbalance table and

(2) only an idiot would tune such a table for this kind of crazy position;

(3) it is highly probable that this simply breaks something.

[bob]

Could it think Knights are not worth anything because it already has so many of them? That doesn't seem the case, because the total score is not that far from balanced. (Stockfish has double the score.) It could not get so close if the Knights on average weren't worth at least 1/3 of a Queen.

It wouldn't be a clever thing to think anyway; against 3 Queens you would need all the Knights you can muster...

Knowing that

(1) they use a material imbalance table and

(2) only an idiot would tune such a table for this kind of crazy position;

(3) it is highly probable that this simply breaks something.

[hgm]

I must say I don't trust such games a lot(online anonymous games).
Who knows how much Komodo was *optimized* here by some changed
params?

Well, I have no reason to suspect foul play by this person. The next game he posted was won by Komodo (from the same position with d2 removed), although Komodonear the end made s similar strange Knight sac. (But by then it was already too late to rescue Stockfish.)

[Guenther]

I must say I don't trust such games a lot(online anonymous games).
Who knows how much Komodo was *optimized* here by some changed
params?

Well, I have no reason to suspect foul play by this person. The next game he posted was won by Komodo (from the same position with d2 removed), although Komodonear the end made s similar strange Knight sac. (But by then it was already too late to rescue Stockfish.)

Well, I don't have Komodo 9.2, but I really meant *optimized* and no word
of foul play or bad intention. Good intentions are enough to produce something unwanted.

Guenther

[hgm]

(1) they use a material imbalance table and

(2) only an idiot would tune such a table for this kind of crazy position;

(3) it is highly probable that this simply breaks something.

I guess this could be an explanation: that > 2 Knights in the material index makes it wrap to some other piece combination. It would still a bit fishy that this material combination would have such a large score correction that it obliterates the difference between a Pawn and a Knight. But perhaps it overflows out of the table alltogether.

[Joerg Oster]

I must say I don't trust such games a lot(online anonymous games).
Who knows how much Komodo was *optimized* here by some changed
params?

Well, I have no reason to suspect foul play by this person. The next game he posted was won by Komodo (from the same position with d2 removed), although Komodonear the end made s similar strange Knight sac. (But by then it was already too late to rescue Stockfish.)

Maybe contempt factor set too high?