Strange analysis by Chess Tiger 2007

[Marc MP]

I was reviewing my classsics with Chess Tiger 2007, when I came to the famous Steinitz-Von Bardeleben game from hastings 1895. I tried 16...Kf7 instead of 16...c6 and here is Chess Tiger 2007 (normal style) suggestion:

[D]r6r/pppqnkpp/5p2/8/3P4/5N2/PP2QPPP/2R1R1K1 w - - 4 1

Code: Select all

00:00:10.9	0.30	15	7564521	Qxe7+ Qxe7 Rxe7+ Kxe7 Rxc7+ Kd6 Rxg7 Rac8 h3 Rc7 Rxc7 Kxc7 Nd2 Rd8 Ne4 f5 
00:00:38.4	0.31	16	28780274	Qxe7+ Qxe7 Rxe7+ Kxe7 Rxc7+ Kd6 Rxg7 Rac8 h3 Rc7 Rxc7 Kxc7 Nd2 Re8 Kf1 Kd7 Nc4 
00:01:36.0	0.21	17	78044050	Qxe7+ Qxe7 Rxe7+ Kxe7 Rxc7+ Kd6 Rxg7 Rac8 h3 Rc7 Rxc7 Kxc7 b3 Rd8 a3 
00:06:03.0	0.21	18	311099907	Qxe7+ Qxe7 Rxe7+ Kxe7 Rxc7+ Kd6 Rxg7 Rac8 h3 Rc7 Rxc7 Kxc7 b3 Kd7 a3 

Rather strange don't you think? The exchange sacrifice coupled with the queen exchange seems unecessary, if not losing, but the eval is + 0.21.

Now if I go through Chess Tiger's main line 16...Kf7 17. Qxe7+ Qxe7 18. Rxe7+ Kxe7, which is totally forced (after 17. Qxe7+ of course), then Chess Tiger changes his mind:

[D]r6r/ppp1k1pp/5p2/8/3P4/5N2/PP3PPP/2R3K1 w - - 0 1

Code: Select all

00:00:10.0	-0.58	19	7797691	Rxc7+ Kd6 Rxg7 Rhc8 h3 Rc1+ Kh2 Rc7 Rxc7 Kxc7 b3 Kd6 g4 Kd5 Kg3 
00:00:20.9	-0.66	20	16884667	Rxc7+ Kd6 Rxg7 Rhc8 h4 Rc1+ Kh2 b6 Kg3 Rc2 h5 a5 Rxh7 Rg8+ Kh3 Rxb2 h6 Rxa2 Rg7 
00:00:57.0	-0.76	21	45611268	Rxc7+ Kd6 Rxg7 Rhc8 g4 Rg8 Rxb7 Rxg4+ Kf1 Rf4 Rb3 a5 Ke2 Re8+ Kd2 a4 Re3 Rb8 Kc3 Kd5 h4 
00:01:46.7	-0.70	22	87286219	Rxc7+ Kd6 Rxg7 Rhc8 g3 Rc7 Rxc7 Kxc7 Kf1 Re8 Nd2 Kd6 f4 Kd5 Nb3 Re3 Nc5 b6 

What can cause this kind of blindness? The moves are totally forced after 17. Qxe7??...

The pgn before someone ask:

[Event "Hastings"]
[Site "Hastings"]
[Date "1895.??.??"]
[Round "10"]
[White "Steinitz, William"]
[Black "Von Bardeleben, Curt"]
[Result "1-0"]
[ECO "C54"]
[EventDate "1895.08.05"]

1. e4 e5 2. Nf3 Nc6 3. Bc4 Bc5 4. c3 Nf6 5. d4 exd4 6. cxd4 Bb4+ 7. Nc3 d5 8.
exd5 Nxd5 9. O-O Be6 10. Bg5 Be7 11. Bxd5 Bxd5 12. Nxd5 Qxd5 13. Bxe7 Nxe7 14.
Re1 f6 15. Qe2 Qd7 16. Rac1 c6 17. d5 cxd5 18. Nd4 Kf7 19. Ne6 Rhc8 20. Qg4 g6
21. Ng5+ Ke8 22. Rxe7+ Kf8 23. Rf7+ Kg8 24. Rg7+ Kh8 25. Rxh7+ 1-0

[Tony]

What do you think is strange about it ?

Tiger gives a score for an 18 ply search, you play out 5 ply, then let it search 20 ply more, and it gives a different score.

Tony

[Marc MP]

What do you think is strange about it ?

Tiger gives a score for an 18 ply search, you play out 5 ply, then let it search 20 ply more, and it gives a different score.

Tony

I find it strange because the moves are forced after 17. Qxe7+?? losing (maybe, or at best giving away the advantage). I tought programs search futher than their ply search count because they have extensions. I tought the 5 plies beginning with 17. Qxe7+ are, from a human point of view, rather forced. These moves are checks and/or recaptures. I tought they were handlded in the node counts.

Maybe another way to see the problem is to ask: Do you know any other top engine that suggest to sac the exchange here? (I haven't check, but I would be surprised if you find one).

[Vempele]

Now if I go through Chess Tiger's main line 16...Kf7 17. Qxe7+ Qxe7 18. Rxe7+ Kxe7, which is totally forced (after 17. Qxe7+ of course), then Chess Tiger changes his mind:

Just because it's forced does't mean any of it is extended. For example, CT might cancel the check extensions after the queens are exchanged, since their main purpose is to see if there are tactics (either caused by the checks or delayed by them).

In the latter position, it finally sees (after much searching) that black's rook utterly dominates white's knight and 2 pawns, as is almost always the case in these endgames when there are pawns on both flanks.

I guess the original positive score was caused by the passed pawn and the fact that (in the middlegame) the exchange is worth slightly less than 2 pawns. CT probably doesn't know the latter doesn't apply to KRkn endgames and is therefore reliant on the search to find out before it's too late.

[Edit]

Maybe another way to see the problem is to ask: Do you know any other top engine that suggest to sac the exchange here? (I haven't check, but I would be surprised if you find one).

Toga doesn't play Qxe7, but its evaluation of the line is about the same as CT's. It's even worse at detecting black is winning. This is not surprising; I've seen it lose numerous games against Rybka where it happily exchanged everything but the rook and knight, then lost the pawns one by one. Its passed pawn scoring needs knowledge about special cases (it thought it was making them stronger by removing material).

[Marc MP]

Now if I go through Chess Tiger's main line 16...Kf7 17. Qxe7+ Qxe7 18. Rxe7+ Kxe7, which is totally forced (after 17. Qxe7+ of course), then Chess Tiger changes his mind:

Just because it's forced does't mean any of it is extended. For example, CT might cancel the check extensions after the queens are exchanged, since their main purpose is to see if there are tactics (either caused by the checks or delayed by them).

In the latter position, it finally sees (after much searching) that black's rook utterly dominates white's knight and 2 pawns, as is almost always the case in these endgames when there are pawns on both flanks.

I guess the original positive score was caused by the passed pawn and the fact that (in the middlegame) the exchange is worth slightly less than 2 pawns. CT probably doesn't know the latter doesn't apply to KRkn endgames and is therefore reliant on the search to find out before it's too late.

[Edit]

Maybe another way to see the problem is to ask: Do you know any other top engine that suggest to sac the exchange here? (I haven't check, but I would be surprised if you find one).

Toga doesn't play Qxe7, but its evaluation of the line is about the same as CT's. It's even worse at detecting black is winning. This is not surprising; I've seen it lose numerous games against Rybka where it happily exchanged everything but the rook and knight, then lost the pawns one by one. Its passed pawn scoring needs knowledge about special cases (it thought it was making them stronger by removing material).

Thank you for your informative answer. I was wrong to assume extensions even if the moves are forced. When I force Chess Tiger to analyse only 17. Qxe7+, it goes quicker (of course) down the line and see in reasonable time that the sacrifice is no good:

Code: Select all

00:00:05.4	0.12	21	4170571	Qxe7+ Qxe7 Rxe7+ Kxe7 Rxc7+ Kd6 Rxg7...
00:00:10.5	0.11	22	8542060	Qxe7+ Qxe7 Rxe7+ Kxe7 Rxc7+ Kd6 Rxg7 Rac8 ...
00:00:24.3	0.03	23	20356886	Qxe7+ Qxe7 Rxe7+ Kxe7 Rxc7+ Kd6 Rxg7 ...
00:00:47.4	-0.01	24	39704702	Qxe7+ Qxe7 Rxe7+ Kxe7 Rxc7+ Kd6 Rxg7 Rac8...
00:01:41.6	-0.05	25	86224262	Qxe7+ Qxe7 Rxe7+ Kxe7 Rxc7+ Kd6 Rxg7 Rac8 g3 ...

[rafowell]

Maybe another way to see the problem is to ask: Do you know any other top engine that suggest to sac the exchange here? (I haven't check, but I would be surprised if you find one).

Neither HIARCS 11.1 nor Shredder 9.11 like Qxd7 even remotely.

At depth 19, MultiPV = 3, Shredder 9.11 thinks:

+0.79 1. Nd2 Rhe8 2. Qe4 Nc6 3. Qxh7 Qxd4 4. Qh5
+0.62 1. Qe4 Nd5 2. Qh4 Rae8
+0.56 1. Qc2 c6 2.Qb1

At depth 19, MultiPV = 3, HIARCS 11.1 thinks:

+0.21 1. Ne5 fxe5 2. dxe5 Qe6 3. Qf3 Kg6
+0.09 1. Nd2 Nd5 2. Qd3 Rae8 3. Ne4 g6
+0.09 1. Qd3 c6 2. Nd2 Rad8 3. Qb3 Nd5

[Tony]

What do you think is strange about it ?

Tiger gives a score for an 18 ply search, you play out 5 ply, then let it search 20 ply more, and it gives a different score.

Tony

I find it strange because the moves are forced after 17. Qxe7+?? losing (maybe, or at best giving away the advantage). I tought programs search futher than their ply search count because they have extensions. I tought the 5 plies beginning with 17. Qxe7+ are, from a human point of view, rather forced. These moves are checks and/or recaptures. I tought they were handlded in the node counts.

You're basic thought is correct. But in general, only checking moves are extended by a whole ply. But maybe not al three. Recapture extensions are very expensive so people give them small extensions (or not). Hardly anyone use singular extensions ( 1 move clearly better than the rest).

For humans, mostly these 3 make a "totally forced line". ( where 2 and 3 are related)

Tony

[Uri Blass]

I believe that extensions is not the only problem here and chess tiger is simply not reliable tool for analysis because it is a preprocessor that means that it can evaluate the same position in different way dependent on the root position.

Uri

[Marc MP]

I believe that extensions is not the only problem here and chess tiger is simply not reliable tool for analysis because it is a preprocessor that means that it can evaluate the same position in different way dependent on the root position.

Uri

Interesting observation. I never realized it before. Just to check it I asked Chess Tiger to analyze 1. Nd2 Rhe8 and then I force 2. Qe4:

Code: Select all

; 00:00:00.0  2.   Score: 0,08   depth: 12 Qe4 Nf5 Qxb7 Rab8 Qxa7 Rxe1+ Rxe1 Rxb2 Nc4 Rc2 Ne3 
; 00:00:01.1  2.   Score: -0,36   depth: 13 Qe4 c6 Qd3 Nf5 Qc4+ Kf8 Qc5+ Ne7 Ne4 Qg4 Nd6 
; 00:00:02.3  2.   Score: -0,28   depth: 14 Qe4 c6 Qd3 Kg8 Qc4+ Kf8 Ne4 Nd5 Qc5+ Kg8 Nd6 Rxe1+ Rxe1 Rd8 Ne4 
; 00:00:04.6  2.   Score: -0,22   depth: 15 Qe4 c6 Qd3 Kg8 Ne4 Qd5 a3 Rac8 Nc5 b6 Ne6 Nf5 Qc4 
; 00:00:09.3  2.   Score: -0,16   depth: 16 Qe4 c6 Qxh7 Ng6 Rxe8 Rxe8 Qh5 Qxd4 Qd1 Nf4 Nc4 Nd3 Qh5+ Ke7 Qe2+ Kf7 
; 00:00:16.3  2.   Score: -0,02   depth: 17 Qe4 c6 Qxh7 Ng6 Qh5 Qxd4 Red1 Re5 Qf3 Qxb2 Rb1 Qxa2 Rxb7+ Kf8 
; 00:00:30.3  2.   Score: -0,02   depth: 16 Qe4 c6 Qxh7 Ng6 Qh5 Qxd4 Red1 Re5 Qf3 Qxb2 Rb1 Qxa2 Rxb7+ Kf8 

I closed the gui, reopened and asked to analyze 1. Qe4 Rhe8 and then 2. Nd2:

Code: Select all

; 00:00:00.1  2.   Score: 0,12   depth: 12 Nd2 Nf5 Qxb7 Rab8 Qxa7 Rxe1+ Rxe1 Rxb2 Nc4 Rc2 Ne3 
; 00:00:00.1  2.   Score: 0,12   depth: 13 Nd2 Nf5 Qxb7 Rab8 Qxa7 Rxe1+ Rxe1 Rxb2 Nc4 Rc2 Ne3 
; 00:00:00.1  2.   Score: 0,02   depth: 14 Nd2 Nf5 Qxb7 Rab8 Qxa7 Rxe1+ Rxe1 Rxb2 Nf3 Rc2 a3 Qd5 a4 
; 00:00:01.3  2.   Score: -0,34   depth: 15 Nd2 c6 Qd3 Nf5 Qc4+ Kf8 Qc5+ Ne7 Ne4 Qg4 Qc4 
; 00:00:04.7  2.   Score: -0,26   depth: 16 Nd2 c6 Qxh7 Ng6 Qh5 Qxd4 Qa5 Rxe1+ Rxe1 Qxb2 Nc4 b6 Nxb2 bxa5 
; 00:00:06.0  2.   Score: -0,18   depth: 17 Nd2 c6 Qxh7 Ng6 Qh5 Qxd4 Rxe8 Rxe8 Qd1 Nf4 Rc4 Qe5 Re4 
; 00:00:18.7  2.   Score: 0,16   depth: 18 Nd2 c6 Qxh7 Ng6 Qh5 Qxd4 Red1 Rad8 Nc4 Qf4 g3 Qc7 Rxd8 Rxd8 h3 
; 00:00:30.4  2.   Score: 0,16   depth: 17 Nd2 c6 Qxh7 Ng6 Qh5 Qxd4 Red1 Rad8 Nc4 Qf4 g3 Qc7 Rxd8 Rxd8 h3 

[Kaj Soderberg]

I believe that extensions is not the only problem here and chess tiger is simply not reliable tool for analysis because it is a preprocessor that means that it can evaluate the same position in different way dependent on the root position.

Uri

I have also noticed that CT2007 is not at it's best when changing from middle game to endgame. Just after the queen has been exchanged, say 3 forced plies later, CT comes with an altogether different evaluation of the situation and that already at a very low ply level after the exchange, so it is not only so that the evaluation changes because new and much deeper ply depths have been reached. BTW This also goes for the predecessors of CT2007, so there is structurally something not completely right in this otherwise fine animal.

Best regards,

Kaj