An endgame where humans are superior?

[Robert Flesher]

I recently saw this postion come up between Rybka 3.0 and Deep Shredder 11. I was very surprised that Shredder lost the game. When I knew I would draw this endgame without any trouble. The correct plan is get the rook to E4 and then do nothing but shuffle the king around the fortress. Rybka 3.0 is cluesless is this postion. Does any engine fully understand this endgame plan?


Analysis by Rybka 3 :

68.Qc5
+- (1.67) Depth: 2 00:00:00
68.Qc5
+- (1.66) Depth: 3 00:00:00
68.Qc5
+- (1.56) Depth: 4 00:00:00
68.Qb4
+- (1.62) Depth: 4 00:00:00
68.Qb4 Ke5
+- (1.63) Depth: 5 00:00:00
68.Qb4 Ke5 69.Qe1+
+- (1.63) Depth: 6 00:00:00 8kN
68.Qb4 Ke5 69.Qe1+ Kd5
+- (1.65) Depth: 7 00:00:00 19kN
68.Kc4 Re6 69.Qb8 g5 70.Kd5
+- (1.69) Depth: 7 00:00:00 19kN
68.Kc4 Re6 69.Qb2+ Kg5 70.Qg2+ Kf6
+- (1.75) Depth: 8 00:00:00 23kN
68.Kc4 g5 69.Qb2+ Kg6
+- (1.69) Depth: 9 00:00:00 37kN
68.Kc4 g5 69.Qb2+ Kg6 70.Qe5 Ra6 71.Qe8+ Kf6 72.Qc8
+- (1.67) Depth: 10 00:00:00 82kN
68.Qb8 Re6 69.Qd8+ Ke5 70.Kc4 Rc6+ 71.Kb5 Rd6 72.Qe7+
+- (1.77) Depth: 10 00:00:00 103kN
68.Qb8 Re6 69.Qd8+ Re7 70.Kc4 Kf7 71.Qd5+ Kf6 72.Qd4+
+- (1.81) Depth: 11 00:00:00 129kN
68.Qb8 Re6 69.Qd8+ Re7 70.Kc4 Kf7 71.Qd5+ Kf6 72.Qd4+ Ke6
+- (1.81) Depth: 12 00:00:00 140kN
68.Qb8 Re6 69.Qd8+ Re7 70.Kc4 Kf7 71.Qd5+ Kf6 72.Qd6+ Kf7 73.Qh2
+- (1.81) Depth: 13 00:00:00 169kN
68.Qb8 Re6 69.Qd8+ Re7 70.Kc4 Kf7 71.Qd2 Ke6 72.Qd5+ Kf6 73.Qd6+ Kf7 74.Kd4 Re2
+- (1.82) Depth: 14 00:00:00 260kN
68.Qb8 Re6 69.Qd8+ Re7 70.Kc4 Kf7 71.Qd2 Ke6 72.Qd5+ Kf6 73.Qd6+
+- (1.82) Depth: 15 00:00:01 295kN
68.Qb8 Re6 69.Qd8+ Re7 70.Kc4 Kf7 71.Qd2 Ke6 72.Qd5+ Kf6 73.Qd6+
+- (1.82) Depth: 16 00:00:01 422kN
68.Qb8 Re6 69.Qd8+ Re7 70.Kc4 Kf7 71.Qd2 Ke6 72.Qd5+ Kf6 73.Qd6+
+- (1.82) Depth: 17 00:00:02 636kN
68.Qb8 Re6 69.Qd8+ Re7 70.Kc4 Kf7 71.Qd2 Ke6 72.Qd5+ Kf6 73.Qd6+
+- (1.82) Depth: 18 00:00:02 857kN
68.Qb8 Re6 69.Qd8+ Re7 70.Kc4 Kf7 71.Qd2 Ke6 72.Qd5+ Kf6 73.Qd6+
+- (1.82) Depth: 19 00:00:03 1152kN
68.Qb8 Re6 69.Qd8+ Kf7 70.Kc4 Re4+ 71.Kd5 Re7 72.Qc8 Re4 73.Qb7+ Re7 74.Qb3 Kf6 75.Qc3+ Kg5
+- (1.82) Depth: 20 00:00:04 1597kN
68.Qb8 Re6 69.Qd8+ Kf7 70.Kc4 Re4+ 71.Kd5 Re7 72.Qc8 Re4 73.Qb7+ Re7 74.Qb3 Kf6 75.Qc3+ Kg5
+- (1.82) Depth: 21 00:00:06 2123kN
68.Qb8 Re6 69.Qd8+ Kf7 70.Kc4 Re4+ 71.Kd5 Re7 72.Qc8 Re4 73.Qb7+ Re7 74.Qb3 Kf6 75.Qc3+ Kg5
+- (1.82) Depth: 22 00:00:08 2910kN
68.Qb8 Re6 69.Qd8+ Kf7 70.Kc4 Re4+ 71.Kd5 Re7 72.Qc8 Re4 73.Qb7+ Kf6
+- (1.82) Depth: 23 00:00:12 4372kN
68.Qb8 Re6 69.Qd8+ Kf7 70.Kc4 Re4+ 71.Kd5 Re7 72.Qc8 Re4 73.Qb7+ Kf6
+- (1.82) Depth: 24 00:00:16 5997kN

(, Microsoft 27.08.2009)

[d] 8/8/3r1kp1/1Q3p2/8/1K6/8/8 w - - 0 68

[Robert Flesher]

I guess at least one engine does!



Analysis by Zappa Mexico X64:

68.Qe8 Re6 69.Qd8+ Kg7 70.Kc3 Re4 71.Qg5 Rg4 72.Qd2 Re4 73.Kd3 Kf6 74.Qh6 Kf7 75.Kc3 Kf6 76.Kb3 Re2 77.Kc3 Re4
² (0.28) Depth: 21 00:00:00
68.Qe8 Re6 69.Qd8+ Kg7 70.Kc3 Re4 71.Qg5 Rg4 72.Qd2 Re4 73.Kd3 Kf6 74.Qh6 Kf7 75.Kc3 Kf6 76.Kb3 Re2 77.Kc3 Re4
² (0.28) Depth: 22/42 00:00:20 151178kN

(, Microsoft 27.08.2009)

[Uri Blass]

I recently saw this postion come up between Rybka 3.0 and Deep Shredder 11. I was very surprised that Shredder lost the game. When I knew I would draw this endgame without any trouble. The correct plan is get the rook to E4 and then do nothing but shuffle the king around the fortress. Rybka 3.0 is cluesless is this postion. Does any engine fully understand this endgame plan?


Analysis by Rybka 3 :
68.Qb8 Re6 69.Qd8+ Kf7 70.Kc4 Re4+ 71.Kd5 Re7 72.Qc8 Re4 73.Qb7+ Kf6
+- (1.82) Depth: 24 00:00:16 5997kN

(, Microsoft 27.08.2009)

[d] 8/8/3r1kp1/1Q3p2/8/1K6/8/8 w - - 0 68

I do not see that rybka3 is clueless in this position.
She clearly put the rook at e4 in her pv.

A positive score for white is also justified because white has practical chances to win and the fact that rybka won against shredder proves it.

If rybka evaluates it as 0.00 then it may allow shredder a simple repetition and not win against shredder and people are going to complain that rybka's rating is smaller.

Uri

[Robert Flesher]

I recently saw this postion come up between Rybka 3.0 and Deep Shredder 11. I was very surprised that Shredder lost the game. When I knew I would draw this endgame without any trouble. The correct plan is get the rook to E4 and then do nothing but shuffle the king around the fortress. Rybka 3.0 is cluesless is this postion. Does any engine fully understand this endgame plan?


Analysis by Rybka 3 :
68.Qb8 Re6 69.Qd8+ Kf7 70.Kc4 Re4+ 71.Kd5 Re7 72.Qc8 Re4 73.Qb7+ Kf6
+- (1.82) Depth: 24 00:00:16 5997kN

(, Microsoft 27.08.2009)

[d] 8/8/3r1kp1/1Q3p2/8/1K6/8/8 w - - 0 68

I do not see that rybka3 is clueless in this position.
She clearly put the rook at e4 in her pv.

A positive score for white is also justified because white has practical chances to win and the fact that rybka won against shredder proves it.

If rybka evaluates it as 0.00 then it may allow shredder a simple repetition and not win against shredder and people are going to complain that rybka's rating is smaller.

Uri

To be more clear, Rybka does not understand this is a draw. The fact that it sees chances proves that it is somewhat clueless. If I understand the search correctly, it believes the opponent should be making the best move and the variation should show a draw. It does not! I do not use tablebases in my engine vs engine matches. Perhaps with tablebases Rybka could see the light. It position like this that allow humans to still be the odd rare draw. Zappa sees the idea and shows the game is pretty even, so this must be a search or knowledge issue.

[Robert Flesher]

I recently saw this postion come up between Rybka 3.0 and Deep Shredder 11. I was very surprised that Shredder lost the game. When I knew I would draw this endgame without any trouble. The correct plan is get the rook to E4 and then do nothing but shuffle the king around the fortress. Rybka 3.0 is cluesless is this postion. Does any engine fully understand this endgame plan?


Analysis by Rybka 3 :
68.Qb8 Re6 69.Qd8+ Kf7 70.Kc4 Re4+ 71.Kd5 Re7 72.Qc8 Re4 73.Qb7+ Kf6
+- (1.82) Depth: 24 00:00:16 5997kN

(, Microsoft 27.08.2009)

[d] 8/8/3r1kp1/1Q3p2/8/1K6/8/8 w - - 0 68

I do not see that rybka3 is clueless in this position.
She clearly put the rook at e4 in her pv.

A positive score for white is also justified because white has practical chances to win and the fact that rybka won against shredder proves it.

If rybka evaluates it as 0.00 then it may allow shredder a simple repetition and not win against shredder and people are going to complain that rybka's rating is smaller.

Uri

The fact Rybka won does not prove there are practical chances, it only proves Shredder was also clueless in this position. I can tell you this, if I was playing black, Rybka would have zero chances to win. It is a simple draw. This is the point I was trying to address. A draw should be seen as a draw, then positions like this could be avoided. IF the engine sees a +1.8 advantage , then it is wrong and will end up with more 1/2-1/2 games.

[Terry McCracken]

I recently saw this postion come up between Rybka 3.0 and Deep Shredder 11. I was very surprised that Shredder lost the game. When I knew I would draw this endgame without any trouble. The correct plan is get the rook to E4 and then do nothing but shuffle the king around the fortress. Rybka 3.0 is cluesless is this postion. Does any engine fully understand this endgame plan?


Analysis by Rybka 3 :
68.Qb8 Re6 69.Qd8+ Kf7 70.Kc4 Re4+ 71.Kd5 Re7 72.Qc8 Re4 73.Qb7+ Kf6
+- (1.82) Depth: 24 00:00:16 5997kN

(, Microsoft 27.08.2009)

[d] 8/8/3r1kp1/1Q3p2/8/1K6/8/8 w - - 0 68

I do not see that rybka3 is clueless in this position.
She clearly put the rook at e4 in her pv.

A positive score for white is also justified because white has practical chances to win and the fact that rybka won against shredder proves it.

If rybka evaluates it as 0.00 then it may allow shredder a simple repetition and not win against shredder and people are going to complain that rybka's rating is smaller.

Uri

The fact Rybka won does not prove there are practical chances, it only proves Shredder was also clueless in this position. I can tell you this, if I was playing black, Rybka would have zero chances to win. It is a simple draw. This is the point I was trying to address. A draw should be seen as a draw, then positions like this could be avoided. IF the engine sees a +1.8 advantage , then it is wrong and will end up with more 1/2-1/2 games.

It's a draw but there's a lot of ways to go wrong. I'm not sure Rybka is missing this, it might be but if it sees the key moves that draw then it may be set to be very aggressive in case of error.

But until it sees an error I too think the eval should read 0.00.

[Edmund]

It is not important to present a long pv that shows to the operator that the engine undertands the position, but rather return a 0 score.

Reporting a 1.82 Score at depth 24 shows that it is quite likely that Rybka would trade pieces to end up in such a position which could be a horrible plunder.

[Terry McCracken]

It is not important to present a long pv that shows to the operator that the engine undertands the position, but rather return a 0 score.

Reporting a 1.82 Score at depth 24 shows that it is quite likely that Rybka would trade pieces to end up in such a position which could be a horrible plunder.

There's still a lot of work to be done in endings, blocked positions etc. The programs although very very strong are far from perfection and will be for decades to come.

[yanquis1972]

it seems to be however, that zappa & occasionally shredder are more accurate in their evals in these positions. it will be especially interesting to see how shredder 12 evaluates compared to 11 (which imo is pretty inaccurate) -- supposedly its strength is at or superior to naum 4, which means it could make zappa & naum somewhat irrelevant as analysis engines (both, & probably esp zappa despite its lower elo, are easily superior to shredder 11 right now imo).

its a true shame that zappa hasnt been in development for a year or two. fantastically accurate analysis engine.