Two More Komodo 9.0 Exchange Sacrifices

[Lyudmil Tsvetkov]

Stockfish 6 is the only one that saw it besides Rxg4 at depth 33.

Yes, I was interested to check and see if stockfish would see it. And it did after several seconds. But houdini completely missed it. And what makes it interesting is that it's not easy to calculate it to a proper conclusion, but rather it is a positional gamble that worked out. Even Komodo's own evaluation didn't show a high score with Rxg4 until several moves later. So in a sense it specuated wit a sacrifice not knowing a concrete win (since it's evaluation was only slightly positive for white).

That's pretty interesting.

Sincerely,

Tim.

Was curious what Stockfish would show with a deep search of the key position:
[d]3r3k/1pr2qp1/p6p/P1Pp1p1P/BP1RpPnR/4P1P1/3Q4/6K1 w - - 41 94

Stockfish finds the exchange sac at depth 32:

Code: Select all

info depth 31 seldepth 32 multipv 1 score cp 24 nodes 44221264 nps 10501368 hashfull 29 tbhits 0 time 4211 pv a4d1 g4f6 d1b3 h8h7 d2d1 f7e6 d1e2 c7c8 h4h2 d8d7 h2h1 d7d8 h1h4 h7h8 e2g2 c8c7 g2d2 h8h7 h4h2 c7e7 d2d1 e7d7 h2h1 d7c7 h1h3 c7c8 h3h2 c8c7 h2g2 c7d7 g2d2 e6f7
info depth 32 currmove a4d1 currmovenumber 1
info depth 32 currmove a4b3 currmovenumber 2
info depth 32 seldepth 36 multipv 1 score cp 30 lowerbound nodes 86133657 nps 14418087 hashfull 52 tbhits 0 time 5974 pv h4g4
info depth 32 currmove h4g4 currmovenumber 1
info depth 32 seldepth 43 multipv 1 score cp 36 lowerbound nodes 118327094 nps 16123054 hashfull 65 tbhits 0 time 7339 pv h4g4
info depth 32 currmove h4g4 currmovenumber 1
info depth 32 seldepth 47 multipv 1 score cp 45 lowerbound nodes 178080186 nps 18232843 hashfull 88 tbhits 0 time 9767 pv h4g4
info depth 32 currmove h4g4 currmovenumber 1
info depth 32 currmove d2d3 currmovenumber 6
info depth 32 seldepth 48 multipv 1 score cp 37 upperbound nodes 244057586 nps 19861457 hashfull 109 tbhits 0 time 12288 pv h4g4 f5g4
info depth 32 currmove h4g4 currmovenumber 1
info depth 32 seldepth 48 multipv 1 score cp 48 lowerbound nodes 272668210 nps 20459834 hashfull 119 tbhits 0 time 13327 pv h4g4
info depth 32 currmove h4g4 currmovenumber 1
info depth 32 currmove g1f1 currmovenumber 4
info depth 32 currmove c5c6 currmovenumber 25
info depth 32 currmove d2c2 currmovenumber 26
info depth 32 currmove a4c6 currmovenumber 27
info depth 32 seldepth 52 multipv 1 score cp 59 nodes 330960483 nps 21265853 hashfull 136 tbhits 0 time 15563 pv h4g4 f5g4 d2h2 c7c6 h2h4 c6f6 a4d1 f7e6 d1g4 e6c6 g4d1 f6f8 d1b3 d8d7 h4g4 f8d8 g4f5 c6f6 f5h3 f6f8 g1f2 f8f6 f2g2 f6f7 h3g4 f7f6 g4d1 f6e6 g3g4 h8h7 g2g3 e6f7 d4d2

The evaluation gets as high as 68 centipawns in favor of White:

Code: Select all

info depth 47 seldepth 74 multipv 1 score cp 68 nodes 12645719298 nps 26352715 hashfull 999 tbhits 0 time 479864 pv h4g4 f5g4 d2h2 f7f6 a4d1 d8g8 g1f2 c7d7 d1g4 d7d8 h2h3 f6c6 g4d1 g8f8 d1b3 f8f7 f2g2 f7c7 h3g4 c7d7 g2f1 d7c7 f1f2 c7d7 f2g2 h8h7 g4h3 h7g8 g2f2 g8h8 b3c2 h8g8 c2d1 g8h8 h3h2 c6e6 h2h4 e6f6 h4f6 g7f6 d1g4 d7e7 d4d2 h8g8 g4f5 g8f7 d2d1 f7f8 g3g4 f8f7 f5g6 f7g7 d1d4

But then the evaluation drops, enough so that Stockfish switches away from RxN, before returning to it again:

Code: Select all

info depth 52 seldepth 92 multipv 1 score cp 37 nodes 137575161783 nps 27931587 hashfull 999 tbhits 0 time 4925433 pv h4g4 f5g4 d2h2 f7f5 h2h4 d8g8 a4d1 g7g5 h5g6 h6h5 d1b3 c7d7 b4b5 g8c8 b5b6 c8c5 b3a4 c5c1 g1g2 d7g7 a4e8 c1c2 g2f1 c2c1 f1e2 c1c5 e8f7 g7f7 g6f7 f5f7 h4g5 h8h7 g5e5 h5h4 g3h4 g4g3 e2f1 c5a5 f1g2 a5b5 g2g3 f7g6 e5g5 g6f7 g3h3 a6a5 g5e5 b5b3 e5d5 b3e3 h3g4 f7g7 g4f5 g7f8 f5e6 f8g8 e6d6 g8d8 d6c5 e3c3 c5b5 d8d5 d4d5 h7g6 h4h5 g6h6 d5d7 c3b3 b5a5 h6h5 d7e7 e4e3 e7b7
info depth 53 currmove h4g4 currmovenumber 1
info depth 53 currmove g1h1 currmovenumber 5
info depth 53 currmove d2g2 currmovenumber 7
info depth 53 seldepth 92 multipv 1 score cp 31 upperbound nodes 144727496258 nps 27919712 hashfull 999 tbhits 0 time 5183703 pv h4g4 f5g4
info depth 53 currmove h4g4 currmovenumber 1
info depth 53 currmove g1h1 currmovenumber 5
info depth 53 currmove d2g2 currmovenumber 7
info depth 53 currmove a4c6 currmovenumber 27
info depth 53 currmove d2a2 currmovenumber 29
info depth 53 currmove a4b5 currmovenumber 30
info depth 53 currmove d4e4 currmovenumber 31
info depth 53 seldepth 92 multipv 1 score cp 25 upperbound nodes 185752825284 nps 27900030 hashfull 999 tbhits 0 time 6657800 pv h4g4 f5g4
info depth 53 currmove h4g4 currmovenumber 1
info depth 53 seldepth 92 multipv 1 score cp 31 lowerbound nodes 200814086236 nps 28080910 hashfull 999 tbhits 0 time 7151267 pv h4g4
info depth 53 currmove h4g4 currmovenumber 1
info depth 53 seldepth 92 multipv 1 score cp 45 lowerbound nodes 205173625409 nps 28098805 hashfull 999 tbhits 0 time 7301863 pv h4g4
info depth 53 currmove h4g4 currmovenumber 1
info depth 53 currmove g1f1 currmovenumber 6
info depth 53 seldepth 92 multipv 1 score cp 34 upperbound nodes 213697141684 nps 28147366 hashfull 999 tbhits 0 time 7592083 pv h4g4 f5g4
info depth 53 currmove h4g4 currmovenumber 1
info depth 53 currmove h4h2 currmovenumber 6
info depth 53 currmove d2f2 currmovenumber 9
info depth 53 currmove d2h2 currmovenumber 10
info depth 53 currmove d2c3 currmovenumber 11
info depth 53 currmove d2b2 currmovenumber 12
info depth 53 currmove g1h1 currmovenumber 13
info depth 53 seldepth 92 multipv 1 score cp 20 nodes 284583972927 nps 28012643 hashfull 999 tbhits 0 time 10159126 pv a4d1 g4f6 h4h2 f7e6 d1b3 f6g4 h2h3 g4f6 h3h4 c7c8 h4h2 h8h7 d2e2 d8d7 h2h1 d7d8 d4d2 c8c7 g1g2 f6g4 d2d1 e6f7 h1h3 f7f6 d1d4 f6f7 g2g1 f7e6 b3d1 g4f6 h3h2 h7h8 e2f1 h8h7 f1h3 d8g8 h3g2 c7d7 g2b2 d7d8 d1b3 g8e8 b2d2 e8e7 d2d1 e7d7 b3a4 d7e7 d1d2 e6f7 a4b3 e7d7 d2e2 f6g4
info depth 54 currmove a4d1 currmovenumber 1
info depth 54 currmove g1f1 currmovenumber 6
info depth 54 currmove a4c6 currmovenumber 7
info depth 54 currmove b4b5 currmovenumber 29
info depth 54 currmove h4g4 currmovenumber 30
info depth 54 seldepth 92 multipv 1 score cp 27 lowerbound nodes 312037088894 nps 27924381 hashfull 999 tbhits 0 time 11174360 pv h4g4
info depth 54 currmove h4g4 currmovenumber 1
info depth 54 currmove d2e1 currmovenumber 9
info depth 54 currmove a4c2 currmovenumber 25
info depth 54 currmove c5c6 currmovenumber 26
info depth 54 currmove d2d3 currmovenumber 27
info depth 54 currmove d2c1 currmovenumber 28
info depth 54 currmove d2b2 currmovenumber 29
info depth 54 currmove d2c2 currmovenumber 30
info depth 54 currmove d4e4 currmovenumber 31
info depth 54 seldepth 92 multipv 1 score cp 25 nodes 314055955372 nps 27920505 hashfull 999 tbhits 0 time 11248219 pv h4g4 f5g4 d2h2 f7f5 a4d1 d8g8 h2h4 g7g5 h5g6 h6h5 d1b3 c7d7 b4b5 g8c8 b5b6 c8c5 b3a4 c5c1 g1g2 d7g7 a4e8 c1c2 g2f1 c2c1 f1e2 c1c5 e8f7 g7f7 g6f7 f5f7 h4g5 h8h7 g5e5 c5b5 d4d2 h7g6 d2c2 f7f6 e5b8 b5a5 b8b7 a5c5 e2d2 f6f8 b7a6 c5c2 d2c2 f8c5 c2d1 c5c6 d1d2 h5h4 f4f5 g6f5 g3h4 g4g3 a6e2 f5f6 e2h5 c6c5 h5g5 f6f7 g5g3 c5b4 d2e2 b4b2 e2f1 b2b1 f1g2 b1b6 g3f4 f7e8

White won the game, but I'm certainly not convinced there is a win here for White, with best play by Black.

Will wait for Lyudmil T. to arrive and present his assessment, but I'm not inclined to criticize Houdini's evaluation just yet.

And I will be waiting for A.R.B.

There is nothing to comment, if Kmodo played the move, it is good.

What would happen, if Houdini did not play Ng4?

In order to win, white should break in some way, and there is a single way to do so - playing g4. b5 is no good, no other reasonable breaks.

One way or another, white should prepare g4, play Bd1, improve the position of its king, queen, rook, and then try to play g4 in the best moment.

What happens after Rg4?

The 3 black connected pawns in the center disappear with the demise of f5, and when g4 pawn is captured and white pushes g4, white creates 4 connected pawns on the king side, also storming pawns, seemingly irresistible.

So the exchange sacrifice gives more than enough compensation, but black is one way or another lost. Nice b6 bind for white, weak g7 backward pawn part of the king shelter, making it inflexible, as when this pawn is moved, the king gets bare, black can do nothing but wait, while white can gradually improve the placement of its pieces to then break.

For me, Rg4 is the best move after Houdini's mistake, but even without that g4 break should also win.

In your line involving Bd1, SF does nothing but shuffle around.

One way or another, SF's score is also best for Bg4.

[TShackel]

White won the game, but I'm certainly not convinced there is a win here for White, with best play by Black.

Will wait for Lyudmil T. to arrive and present his assessment, but I'm not inclined to criticize Houdini's evaluation just yet.

That just proves this was a speculative saccrifice by komodo. As I said before komodo played it without having a huge positive score for white. But only a move or two later it shot to 60 centipawn advantage for white. And when komodo gets that high of an advantage that usually means a likely win. I say usually because sometimes there are fortresses that don't win.

Sincerely,

Tim.

[zullil]

And I will be waiting for A.R.B.

May you wait a long, long time.

There is nothing to comment, if Kmodo played the move, it is good.

For me, Rg4 is the best move after Houdini's mistake, but even without that g4 break should also win.

So you believe White wins after Rxg4. I don't doubt that Rxg4 is likely "best", but that White wins seems less clear to me.

What do you think of Stockfish's plan after 95. Qh2, namely 95...Qf5 followed by Rg8 and g5? Very different from Houdini's approach.

[Lyudmil Tsvetkov]

And I will be waiting for A.R.B.

May you wait a long, long time.

There is nothing to comment, if Kmodo played the move, it is good.

For me, Rg4 is the best move after Houdini's mistake, but even without that g4 break should also win.

So you believe White wins after Rxg4. I don't doubt that Rxg4 is likely "best", but that White wins seems less clear to me.

What do you think of Stockfish's plan after 95. Qh2, namely 95...Qf5 followed by Rg8 and g5? Very different from Houdini's approach.

Think that you might want to proceed from here (understand do a fresh search) in your main line, ending with 37cps white advantage:

[d]6rk/1pr5/p5P1/PPPp1q1p/3RpPpQ/4P1P1/8/3B2K1 b - - 0 5

I bet new search will end with 137 and not 37cps advantage.

I ran a few quick shootouts with my SF 6 on this position, and SF wavers quite a lot throughout, going from scores as low as 20cps to scores as high as 80-90cps, then repeatedly going down and repeatedly up, to finally end up with more than 200cps advantage.

The position is very tactical, very tactical positions are also very deep and in them, engines also have difficulties, albeit smaller than humans.

Simply too many lines that change quickly in nature.

For me, this is too tactical, I would do many mistakes along the way, but I am certain of my assessment that white has a won position.

[reflectionofpower]

Interestingly if you set Houdini 4 in "tactical mode" it really keeps setting itself on Rxg4. Could someone really let it look at it for a long time? I can't as my slim design in this i7 doesn't behave well doing this.

[zullil]

Think that you might want to proceed from here (understand do a fresh search) in your main line, ending with 37cps white advantage:

[d]6rk/1pr5/p5P1/PPPp1q1p/3RpPpQ/4P1P1/8/3B2K1 b - - 0 5

I bet new search will end with 137 and not 37cps advantage.

I note that you have deviated from Stockfish's line, by playing b5 without Bb3 first.

A fresh search (with clean hash table) shows nothing at all for White (at least yet):

Code: Select all

info depth 38 seldepth 61 multipv 1 score cp 0 nodes 3579262534 nps 24297155 hashfull 901 tbhits 186967 time 147312 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6
info depth 39 currmove g8c8 currmovenumber 1
info depth 39 currmove a6b5 currmovenumber 4
info depth 39 seldepth 61 multipv 1 score cp 0 nodes 4380659435 nps 24645753 hashfull 950 tbhits 234987 time 177745 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6
info depth 40 currmove g8c8 currmovenumber 1
info depth 40 currmove a6b5 currmovenumber 4
info depth 40 seldepth 61 multipv 1 score cp 0 nodes 5646362972 nps 25024987 hashfull 983 tbhits 362839 time 225629 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6
info depth 41 currmove g8c8 currmovenumber 1
info depth 41 currmove a6b5 currmovenumber 4
info depth 41 seldepth 61 multipv 1 score cp 0 nodes 7342981862 nps 25402440 hashfull 995 tbhits 555060 time 289066 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6
info depth 42 currmove g8c8 currmovenumber 1
info depth 42 currmove g8e8 currmovenumber 6
info depth 42 seldepth 61 multipv 1 score cp 0 nodes 9455142214 nps 25698209 hashfull 998 tbhits 793988 time 367930 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6
info depth 43 currmove g8c8 currmovenumber 1
info depth 43 currmove c7c5 currmovenumber 5
info depth 43 seldepth 61 multipv 1 score cp 0 nodes 12725115511 nps 26145760 hashfull 999 tbhits 1170037 time 486699 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6

Looks like Stockfish's response to the exchange sac on g4 draws for Black. Maybe Peter or someone else can investigate this.

Thanks for the feedback, in any case.

[zullil]

That just proves this was a speculative saccrifice by komodo. As I said before komodo played it without having a huge positive score for white. But only a move or two later it shot to 60 centipawn advantage for white. And when komodo gets that high of an advantage that usually means a likely win. I say usually because sometimes there are fortresses that don't win.

Sincerely,

Tim.

Maybe the evaluation rose quickly because Houdini failed to find a good counter-plan (probably because of the fast time control---is that depth 26 for 95...Qe6?). I wonder how Komodo would handle the Black pieces after RxN, given some time to search. Looks like Komodo was expecting 95...Qf5, if I'm correctly reading the comments in the original pgn.

[Lyudmil Tsvetkov]

Think that you might want to proceed from here (understand do a fresh search) in your main line, ending with 37cps white advantage:

[d]6rk/1pr5/p5P1/PPPp1q1p/3RpPpQ/4P1P1/8/3B2K1 b - - 0 5

I bet new search will end with 137 and not 37cps advantage.

I note that you have deviated from Stockfish's line, by playing b5 without Bb3 first.

A fresh search (with clean hash table) shows nothing at all for White (at least yet):

Code: Select all

info depth 38 seldepth 61 multipv 1 score cp 0 nodes 3579262534 nps 24297155 hashfull 901 tbhits 186967 time 147312 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6
info depth 39 currmove g8c8 currmovenumber 1
info depth 39 currmove a6b5 currmovenumber 4
info depth 39 seldepth 61 multipv 1 score cp 0 nodes 4380659435 nps 24645753 hashfull 950 tbhits 234987 time 177745 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6
info depth 40 currmove g8c8 currmovenumber 1
info depth 40 currmove a6b5 currmovenumber 4
info depth 40 seldepth 61 multipv 1 score cp 0 nodes 5646362972 nps 25024987 hashfull 983 tbhits 362839 time 225629 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6
info depth 41 currmove g8c8 currmovenumber 1
info depth 41 currmove a6b5 currmovenumber 4
info depth 41 seldepth 61 multipv 1 score cp 0 nodes 7342981862 nps 25402440 hashfull 995 tbhits 555060 time 289066 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6
info depth 42 currmove g8c8 currmovenumber 1
info depth 42 currmove g8e8 currmovenumber 6
info depth 42 seldepth 61 multipv 1 score cp 0 nodes 9455142214 nps 25698209 hashfull 998 tbhits 793988 time 367930 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6
info depth 43 currmove g8c8 currmovenumber 1
info depth 43 currmove c7c5 currmovenumber 5
info depth 43 seldepth 61 multipv 1 score cp 0 nodes 12725115511 nps 26145760 hashfull 999 tbhits 1170037 time 486699 pv g8c8 b5b6 c7c5 h4e7 f5g6 e7b7 g6f6 b7a6 c8d8 a6a7 h5h4 b6b7 c5b5 a5a6 h4g3 d4d2 h8g8 a7a8 f6f8 a8a7 f8f6

Looks like Stockfish's response to the exchange sac on g4 draws for Black. Maybe Peter or someone else can investigate this.

Thanks for the feedback, in any case.

SF is known to see draws everywhere.

The line is too long and tactical for me to suggest improvements along the way, but, even if the sacrifice only draws, there should have been different opportunities for white earlier on to gradually increase pressure and at some point break with g4.

No complete equality above and for most of the game, that is for certain.

[TShackel]

Maybe the evaluation rose quickly because Houdini failed to find a good counter-plan (probably because of the fast time control---is that depth 26 for 95...Qe6?). I wonder how Komodo would handle the Black pieces after RxN, given some time to search. Looks like Komodo was expecting 95...Qf5, if I'm correctly reading the comments in the original pgn.

The time control was a long one, not a short one. 90 minutes per side per game plus 20 seconds increment. Average game is 3.5 hours or so. So time being a problem is not an excuse. I think there was something to the sacrifice. It's just a matter of time before the pawns go forward in concert with the bishop and queen.

Sincerely,

Tim.

[M ANSARI]

Beautiful break from what looked like a certain draw! Love it when engines play like that ... almost like watching a cat play with a mouse!