Tigran Petrosian’s engine-like wisdom!

January 29, 2021 Matthew Sadler 7 comments

In his notes to his Black game against Mikhail Tal in the 1958 USSR Championship, later World Champion Tigran Petrosian made a profound observation that has always stayed in my mind.

“Experienced players know that, in a cramped position, the main trouble often happens to be the poorer activity of the rooks. E.g., White advances his kingside pawns supported by rooks on f1 and g2; the black rooks are limited to their first and second ranks, waiting until the game will be opened”

The position after Tal’s 24.Rxa4

Petrosian played 24…Rbd8 25.Qf3 Rd6

The position after Petrosian’s 25…Rd6

26.Nb3 Nd7 27.Raa1 Rg6

The position after Petrosian’s 27…Rg6.

28.Rf1 Bd6 29.h4 Qd8 30.h5 Rf6 31.Qg4 Rf4

The position after Petrosian’s trademark exchange sacrifice 31…Rf4!

This is how he explained the fantastic defensive rook manoeuvre in the game: he was simply trying to reduce that inherent disadvantage of being a defender.

I thought of this while trying to make sense of game 80 of the TCEC SuperFinal between Leela Zero and Stockfish.

The position after Black’s 22…Qc4

This position is taken from a crucial moment in the Season 20 SuperFinal between Leela Zero and Stockfish. Stockfish had taken a 3–point lead after 79 games with a convincing win on the White side of this “slow” French Winawer (1.e4 e6 2.d4 d5 3.Nc3 Bb4 4.e5 b6) and now it was Leela’s turn to try and make something of the White pieces.

In principle, Stockfish had done reasonably with Black, flattening out White’s dynamic advantage with exchanges and avoiding major weaknesses. However, White’s space advantage (the e5–pawn), better mobility (the white rooks can use the open 4th rank to move from kingside to queenside) and Black’s awkward king on g8 (blocking in Black’s rook on h8) add up to a stable White advantage. TCEC Leela’s assessed its advantage as 0.53 which for Leela (which always has conservative evaluations) is almost a clear advantage. This figure did not change for many subsequent moves. Stockfish’s evaluation was much more interesting, varying between 1.1 and 1.3. Stockfish’s evaluations in the opening phase are often over-optimistic in my view, but you should take note in the middlegame phase!


A far from obvious starting move, supporting Rh4 to drive away the black queen from the fourth rank.


The black queen maintains its contact with the d3–square in order to prevent Qd3+ after …Kh7. Black’s  first priority is to get its king off the first rank and activate its king’s rook on h8.

24.Rh4 Kh7

The position after Black’s 24…Kh7

The first step has been accomplished; the next step will be to move the king’s rook and Black’s development will be normalised. However, there is a funny part to this. Black must activate its rook on h8 to stand any chance of defending its position: otherwise, Black is essentially a rook down. However, once the rook has moved from h8, you can argue that Black’s defence of its kingside has been weakened and that a White kingside pawn assault with g4–g5 stands a better chance of succeeding. We have seen many times after all, Stockfish turning a normal castled structure into this defensive structure with …Kh7, …Rh8 and …Kg8! In that way, once the g-file is opened, Black can defend g7 with …Rh7 and leave its king less exposed on g8; after …Kh7, Black must defend g7 with …Rg8, leaving its king somewhat in the open air on h7!

How should White proceed? From move 21, Leela developed a consistent view of the position, did not change its ideas or its basic evaluation and predicted the course of the game pretty accurately. However, while doing so, Stockfish’s evaluation of its chances improved considerably! From 1.11 on move 21 to 0.16 on move 35. It isn’t 100% obvious at first glance however what caused this. Let’s play first of all through Leela’s attempt and see how Stockfish defended.

25.Kb1 Rhf8 26.Rhd4 b5 27.g4 Rc4 28.f4

White’s position after 28.f4

Firstly, Leela plays the pawn to f4, ensuring that g5 …hxg5 can be met by fxg5.

28…Rxd4 29.Rxd4 Qb6 30.g5 a6

The position after 30…a6

Stockfish sets out its stall, claiming it can hold this setup for eternity!

31.Ka2 Qc7 32.Qd3+ Kh8 33.Qe3 Rg8 34.Qf2 Kh7 35.Qf3 Kh8 36.Qf2 Kh7 37.Qf3 Kh8 38.Qe2 Rb8 39.Kb1 Rg8 40.Ka2 Qb6 41.Qf2 Qc7 42.Qf3 Rf8 43.Qe3 Rg8 44.Rb4 Rb8 45.Qe2 Ra8 46.Qe1 Rg8 47.Qe3 Rb8 48.Qf2 Rc8 49.Qe1 Rg8 50.Rd4 Qa7 51.Qg3 Qc7 52.Rb4 Qb6 53.Rd4 Qc7 54.Rb4 Qb6 55.Kb1 Qc5 56.Rd4 Qc7 57.Ka2 Qc6 58.Qf2 Qc7 59.Qe1 Qa7 60.Qg1 Qc5 61.Qd1 Qc7 62.Qf1 Qc6 63.Rb4 Qc5 64.Rd4 Qc6 65.Qe2 Qc7 66.Rb4 Qb6 67.Rd4 Qc7 68.Rb4 Qb6 69.Qg4 Qc7 70.g6

The position after 70.g6 (finally!)

Finally Leela moves to the next phase of its plan, setting up an advanced pawn on the opponent’s third rank (a favourite AlphaZero ploy too!) thus pinning the black king to the back rank.

70…Rf8 71.a4 Qd7 72.Qe2 fxg6 73.hxg6 Qe8 74.Qg4 Kg8

The position after Black’s 74…Kg8

And this is the defensive structure that Stockfish will resist with! Black’s rook on f8 prevents White from breaking with f4–f5 while keeping a white major piece tied to the defence of the f4–pawn, the Black queen can hit the g6–pawn from time to time to distract White from attacking plans on the queenside and Black’s queen will remain active ready to check the white king as White tries to make progress.

75.Ka3 Qd7 76.Qg1 bxa4 77.Rxa4 Qe8 78.Qg3 Qb5 79.Qg4 Qc5+ 80.Ka2 Qc6 81.Rd4 a5

Black would probably be able to sit tight and hold, but Stockfish has a more active plan in mind.

82.Qg2 Qb5 83.Kb1 a4 84.Ka1 Qb3 85.Rb4 Qd1+ 86.Ka2 a3

Black uses its rook’s pawn to open up the White’s king position!

87.bxa3 Qd3 88.Kb2 h5

The position after Black’s march of the rook’s pawn with 88…h5!

And now the other rook’s pawn – which became passed almost unnoticed after White played g6! – starts to roll! It’s a little risky as the black king becomes more exposed (I once lost a rapid game – 5 pawns up if I remember correctly – to English grandmaster Michael Adams by doing exactly this!) but it also gives White an extra headache in addition to its open king.

89.a4 h4 90.Qg4 Qd2+ 91.Kb3 Qd3 92.a5 Ra8 93.Qxe6+ Kh8

It’s become very messy again, but Black is fine: the open white king gives Black many perpetual opportunities while the h-pawn is pretty fast!

94.Qb6 Rc8 95.Qf2 Qxc3+ 96.Ka2 Qxb4 97.Qxh4+ Kg8 98.Qh7+ Kf8 99.Qh8+ Ke7 100.Qxc8 Qxa5+ 101.Kb3 Qb6+ 102.Ka4 Qd4+ 103.Kb3 Qb6+ 104.Kc3 Qe3+ 105.Kb2 Qxf4 106.Qc7+ Ke6 107.Qxg7 Qb4+ 108.Kc2 Qa4+ 109.Kc3 Qa3+ 110.Kd4 Qb4+ 111.Kd3 Qc4+ 112.Kd2 ½–½

Essentially the game featured a standoff: Leela chose a strong, typical structure and Stockfish said: “I can hold this!” So, what’s the conclusion? Is this structure just nothing for White? Well firstly, I should make the caveat that Stockfish makes it look easy to sit and wait for however many moves without making a tactical mistake, but it’s a different matter for an over-the-board human game! This would be a difficult defensive task.

To investigate the objective truth however, it’s interesting to look at Stockfish’s evaluation, and at Stockfish’s preferred line. Firstly, when did Stockfish’s evaluation drop exactly? It was when White played 26.Rhd4 and Black played 26…b5.

Position after Stockfish’s “magic” 26…b5!

Stockfish’s evaluation went to just 0.32 after achieving that! “Now you see an advantage, now you don’t!” Magic!

A fellow TCEC fan on my Twitter feed made the following very appropriate comment:

“Puzzling over why 26…b5 moved SF’s eval from +1.1 to +0.3. Seemingly b5 changed nothing. The pawn structure didn’t change. So what did Fish see that it hadn’t seen before b5?! My only theory is now if there is ever f5 and ef, black queen on a6 is now an active defender???”

It’s intriguing as well that Stockfish’s preferred move 1 move earlier was not 25.Kb1 but 25.Rb4

The position after Stockfish’s preferred 25.Rb4

preventing 25…b5 due to 26.Qd3+ forking king and b-pawn. Let’s dive a bit deeper and look at Stockfish’s main line after 25.Rb4

25.Rb4 Rhf8 26.g4 Qb7 27.Kb1 Rg8 28.Ka2 Rc6 29.f4 a5 30.Rbd4 b5 31.g5

The position after 31.g5 (Stockfish’s analysis)


31…Rc4 loses here to 32.g6+ fxg6 33.hxg6+ Kh8 (33…Kxg6 34.Qd3+ Kf7 35.f5) 34.f5 exf5 (34…Rxd4 35.Qxd4 exf5 36.Qxd5) 35.Rxd5 with a powerful White position.

32.Qd3+ Kh8 33.Qd2 Kh7 34.Re1 and the line continues with a +1 advantage for White.

It doesn’t look that different does it? It’s kind of a spot the difference puzzle! The big difference is that all the rooks are on the board. As Petrosian said, the biggest problem in defending is often the passivity of the rooks. From that point of view, exchanging a pair of rooks – as Stockfish managed in the game against Leela – is a significant defensive achievement for Black!

So, does 26…b5 help Black to exchange rooks? Well, let’s see what happens if Black tries to do without it.

The position from Leela-Stockfish after 25.Kb1 Rhf8 26.Rhd4.

a) 26…Rc4 27.Qd3+ Kh8 28.Rxc4 forces the opening of the d-file with a commanding position for White. 28…dxc4 29.Qd7

The position after 29.Qd7 (variation 26…Rc4 instead of 26…b5)

b) 26…Kh8 avoids the check, but also weakens Black’s defence of the h6–pawn and also abandons Black’s blockade of the white h-pawn. 27.g4 Rc4 28.g5 hxg5 29.Qxg5 Rxd4 30.Rxd4 Kh7 (30…Qb7 31.h6 g6 32.Qf6+ shows the terrible downside of 26…Kh8!) 31.Qe7 Kg8 32.h6

The position after 32.h6 (analysis of variation 26…Kh8)

with a winning position for White.

26…b5 supports …Rc4 and enables Black to recapture on c4 with the b-pawn, avoiding the opening of the d-file! It also supports …Rc4 without needing to move the king from h7 so the g6-pawn keeps its protection, and the king stays on its optimal spot. The white queen can of course chase away the black king with Qd3+… but then the white queen is no longer on the c1-h6 diagonal and no longer attacking h6!

I’m certainly not claiming that White is winning after 25.Rb4, but it’s a beautiful prophylactic move, stopping a major Black defensive idea. As you can see from Stockfish’s main line, it costs Black a lot more time to organise the necessary …b5, which inevitably gives White extra chances.

Interestingly, in the engine games I ran on my weaker hardware and less cutting edge engine versions (not a perfect test in any way, but somewhat indicative), “mad hacker” neural net Stoofvlees II a16 went for 25.Rb4 and managed to defeat Stockfish!

Also, when I ran “my Leela” with the lovely Nibbler tool for about 23.7M nodes (this took a really very long time on the weak laptop graphics card I used for that test!), it switched its first choice at move 25 from 25.Kb1 to 25.Rb4.  

Nibbler output of Leela analysing White’s 25th move alternatives

That’s strange in a way since the reading on “TCEC Leela” on move 25 was 71.3M nodes: it had searched much more than “my Leela” and still went for a weaker move.

The TCEC output after move 25

I’m somewhat out of my expertise level but I believe this may have been an artefact of how well Leela had predicted the moves that Stockfish would make! Already on move 21, Leela had predicted the game continuation up to move 28 (move 31 if you ignore a small transposition) When that happens – as far as I understand – Leela does not throw away its previous analysis but keeps all the relevant parts and then builds on that further.

By move 24, Leela had built up a node count of 58.2M nodes (that’s a big search tree!) Between move 24 (24.Rh4) and 25 (25.Kb1), Leela’s node count increased from 58.2M to 71.3M. This increase may not have been enough to promote the stronger idea 25.Rb4 to its main line (I’m absolutely convinced that Leela saw it and dedicated some time to it). Remember, Leela always plays the move it has analysed the most, which is not necessarily the best move it finds in the position!

If I’ve got that wrong, then happy to be corrected of course!

Anyway, getting back to the chess, it seems that Stockfish assessed the general contours of the position better than Leela, understanding Black’s need to exchange rooks to reduce its dynamic disadvantage in a way that would have made Petrosian very proud!. Even after that it was still a difficult defence of course, but difficult defence is something Stockfish does very easily!

7 Comments on “Tigran Petrosian’s engine-like wisdom!

  1. Hi Matthew!
    Great article! As usual, I may add!
    It is always fascinating to read your latest insights on computer chess, and the new and valuable insights that it brings to us human players! And the way you made the connection with Petrosian’s defensive skills and concepts makes it all the more interesting!
    Incidently, I would like to learn much more of Petrosians concepts and skills and therefore I am looking for good game collections of him, with good analysis. I do possess the My Great Predecessors series, where part 3 is partly dedicated to Petrosian, but I guess there is much more valuable material to be found on this genius. My first sources in those situations are of course your excellent bookreviews in NIC Magazine! I have been reading all your columns in NIC so far, since I have been a subscriber since the very first issue! I saw your positive comments on the first volume of the new Petrosian game collection by Tibor Karyoli. Would that be a good choice, you think? Ore are there other book titles that you could maybe suggest?
    It would be great if you could driop me a line or two!! Thanks in advance and thank you for sharing all your valuable insights with us throughout the years!
    Kind Regards, Job de Lange
    (ps. 1: I am a 2200 player; I played a few times for the Dutch senior and youth teams in my younger days; I am almost 68 years old)
    (ps. 2 : It would be a great idea to collect all your book reviews in some way; then people like me don’t have to scroll through dozens of NIC Magazines to look for your reviews!)

    1. Hi Job, Thanks for your comment! Yes the Karolyi book is pretty good – there were some very good games from his earlier period that I wasn’t aware of! The other book I reviewed on the King’s Indian was also good, but most interesting if you’re a King’s Indian fan / 1.d4 player. A collection of my book reviews sounds fun, but I think I’ll have to become much more important before that happens 🙂 Glad you like them! I love receiving new review books through the post – every day feels like Christmas! Best Wishes, Matthew

  2. I really enjoyed your comments about Petrosian . i have several books about him and have learned a lot by studying his games. I managed to beat him in a simul in 1976. he was playing over 50 games and I always felt I received indirect assistance from the ten masters who were also playing. He lost only 5 games over all. About a year and a half later. I was present at another simul he gave . His opposition was not as strong and he won 54 games and drew one.I was at a table selling chess books and sets. After the exhibition he came over and glanced at a few of the books . He then looked at me for a few seconds and said to me “I lost Gruenfeld to you a year ago ” and smiled. I was impressed by his memory. I have known Roman Dzindzichashvili for many years . At a lecture he gave he told a remarkable story about Petrosian’s intuition. It concerned a game Dzindzichashvili played against jansa at the Tiflis 1965 tournament where he finished in second place after Gurgenidze. When he and Jansa were going over the game they were looking at a position where on the 25th move Dzindzichashvili played kf1. Petrosian wandered by and casually suggested that white should instead have played 25 Rf1 followed be f4 and a king side attack. This was surprising it involved a piece sacrifice in an ending. Larry Kaufman was present and remarked that the best engines then could not find this idea although they realized it’s strength if forced to play it. Several years later a more advanced engine did find Petrosian’s idea. Dzindzichashvilli’s actual move was good enough to win but he was enormously with Petrosian’s intuition which enabled him see in a flash what he himself did not even consider at the time.

  3. Hi, I created an engine that is probably the first artificial intelligence chess in the world, as it was developed between 2000 and 2001, 16 years before the development of AlphaZero. That machine runs on any Windows computer, but unfortunately it was made to use only one thread and does not have the modern techniques to achieve very deep analysis.
    Just as AlphaZero itself developed a style of direct attacks on the opponent’s king (a style typical of players like Alekhine, Tal and Kasparov), I worked directly to develop an engine that would play as close as possible to Capablanca. For example, playing against itself (5 minutes per move using a Ryzen Zen1+ 3200g processor) in the Leela vs Stockfish position it performs the following moves:

    23. g3 {+0.93/18} Qg4 {+1.07/18} 24. Rhh1 {+1.07/19} Kh7 {+1.07/19}
    25. Rd4 {+1.04/19} Qg5 {+1.06/20} 26. f4 {+1.15/20} Qe7 {+1.15/19}
    27. Qd3+ {+1.17/19} g6 28. Kc2 {+1.17/18} Kg7 {+1.17/17}
    29. g4 {+1.17/18} Qc7 {+1.17/17} 30. Rh3 {+1.17/16} Rce8 {+1.33/17}
    31. Kb1 {+1.25/17} Rd8 {+1.31/18} 32. Qg3 {+1.31/17} Rdf8 {+1.30/17}
    33. f5 {+2.84/17} g5 {+3.10/18} 34. fxe6 {+3.86/19} fxe6 {+3.86/19}
    35. Qd3 {+4.08/20} Qf7 {+3.87/20} 36. Rf3 {+3.98/21} Qe8 {+3.97/21}
    37. Rxf8 {+3.99/21} Rxf8 {+4.14/21} 38. c4 {+4.15/21} Qf7 {+4.15/21}
    39. cxd5 {+4.21/21} exd5 {+4.21/21} 40. Rxd5 {+4.21/20} Qf1+ {+4.27/20}
    41. Qxf1 {+4.34/21} Rxf1+ {+4.36/21} 42. Kc2 Rf4 {+4.26/20}
    43. Rd7+ {+4.26/20} Kg8 {+4.20/21} 44. e6 {+4.20/20} Re4 {+4.41/20}
    45. Rxa7 {+4.74/20} Rxe6 {+4.82/20} 46. Rb7 {+5.22/21} Re2+ {+5.28/21}
    47. Kc3 {+5.68/22} Re3+ {+5.91/21} 48. Kd2 {+6.02/22} Re6 {+6.02/21}
    49. b4 {+6.77/21} Rf6 {+6.80/21} 50. Ke3 {+6.89/21} Kf8 {+7.21/21}
    51. b5 {+7.19/21} Rf4 {+7.56/21} 52. Rxb6 {+7.57/21} Rxg4 {+7.57/20}
    53. Kd3 {+7.97/20} Rg3+ {+8.72/19} 54. Kc4 {+8.95/20} Rh3 {+9.17/19}
    55. Rxh6

    Due to the shallowness of the analysis, he often makes (very human) tactical mistakes. This can be partially solved by playing the same position multiple times, as he learns to re-evaluate the position according to the assessments learned in previous games.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.