The Computer Plays The Exchange Sacrifice

The year is 2013. Those of us who remember playing against chess computer programs probably fondly remember the days when chess engines were relatively weak and did not “understand” chess. But 16 years ago, when Deep Blue defeated Kasparov in a match, it became clear that however one wanted to define “understand”, chess played by a computer could be of reasonable quality. And since then, of course, chess engines running on a personal computer have become far stronger than humans. Furthermore, they happily choose to do things such as sacrifice a Pawn or an exchange for compensation.

Although I regularly use chess engines for analysis, I have up till now not looked much at computer vs. computer games, but by accident, I saw a tweet by Dan Heisman in which he noted that “A classically instructive exchange sac by Komodo in Gull-Komodo has left Komodo now on the edge of winning.” I was intrigued and found the score of the game, which was played in the latest TCEC tournament. Gull is an engine rated 3063, while Komodo is rated 3086, so it is particularly interesting to examine the kind of game in which there are no immediately losing blunders. I think there are lessons that can be observed by and used by us humans. My aim is to cast a human interpretation on what happened in this game.


Komodo as Black chose to play a passive but solid variation of the Bogo-Indian Defense, giving White the expected small advantage.


There were many opportunities for White to choose a thematic attempt to pressure Black, such as by forcing the hanging pawns structure and also forcing Black to play c4 giving White control of the d4 square. I think many of us humans would have played this way in an attempt to make progress.

An exchange sacrifice easy to see

But Gull chose to go for more by seemingly first preventing c4 with b3 before aiming at destabilizing Black’s center. Gull’s idea was revealed when it played Ne5, banking on a pin of Black’s Knight on d7 because of the Rook on c8.

But that backfired because Komodo simply took the Knight on d5, sacrificing the exchange for a Pawn. This is easily seen to be completely sound; most strong human players would do this almost reflexively after checking a few concrete lines. Without this resource, Black would have been in trouble. With this resource, Black completely equalizes.

It is an interesting question why one strong chess engine sees something correctly that another does not. In any case, this shows that even strong chess engines do not always see the subtleties in positions.

Weaknesses in White’s position

It turned out that White’s King side, lacking piece defenders, began suffering as Black provoked weaknesses, first as White played h3, and then when White played f3. What you have to remember is that these were not just patzer moves. Gull played them because they are justified given the danger to White’s King. And despite the “weakening” moves, White can still hold. But it is instructive how important it is, even in a seemingly even position, to make some kind of progress through provoking weaknesses.

Draw offer rejected?

Then there was shuffling around of pieces, but then Gull made the decision to avoid a repetition by playing aggressively on the Queen side with a4. From a long-term thinking human standpoint, this was very risky, permanently weakening the b3 Pawn as well as setting up the a4 Pawn as a target for Black’s Bishop, and creating a hole at b4. Still, because a strong engine chose it, it is not yet a losing move.

A nice Knight maneuver

Then there was much shuffling around yet again of pieces. But then Black came up with the fine Knight maneuver back to e8 and then to c7. As though by magic, after some more moves, Black had improved its position yet more, and White’s pieces were looking very strangely placed and passive. If any side had winning chances, it was clearly Black.

More weakening by White

White then made the decision to advance on the King side with h4 and h5 to gain space and drive Black’s Rook away from g6. Again, objectively this is not yet losing, but it is risky. It was especially risky to land on h5, a light square, because Black has a light-squared Bishop.


Finally, Black redeployed the Knight to b4 (highlighting a drawback of White’s early a4 advance weakening the b4 square), and White traded it off. With the reduced forces, I’ll call the resulting phase of the game an endgame.

Provoking another weakness

There was much shuffling of pieces again. Black managed to redeploy the Bishop to attack White’s h5 Pawn, forcing the loosening defensive g4. Then Black swung to the other wing to advance with b5. White defended, but Black then advanced b4.


After more maneuvering, Black was ready to play c4, and then c3.

Final error

Even then, at the 61st move of the game, White could have held the position for a bit, by preventing Black from invading on the King side. I have not done a full analysis, but it is possible that White is lost anyway, given Black’s perpetual threats of advancing to c2 and/or opening up the King side with f5.

But for sure, White’s allowing the Queen to come to h2 resulted in Black tying up White’s Rook on g3 and then striking with Bh7 and c2 followed by penetration at c3, game over effectively. It took another 50 moves for Black to gradually win more material, force and exchange of Queens, and play in a simplified endgame of Bishop and three connected passed Pawns versus Rook.

Summary of the game

The big-picture summary of the game is that White allowed a sound exchange sacrifice that resulted in a position that clearly only gave Black winning chances. White refused implicit draw offers (of pieces shuffling around) and made weakening Pawn moves that Black took advantage of with Knight and Bishop attacking the weaknesses, until White’s overextension finally resulted in a precarious position. One slip by White, allowing an invasion on the King side that immobilized one Rook, led immediately to a lost position.

One lesson that can be learned is: pay attention to Pawn moves that create long-term static weaknesses. In this game, Black did quite a bit of maneuvering to provoke and attack these weaknesses. Another lesson is, playing for a win when the draw is obvious can be dangerous. It’s strangely enjoyable seeing how even chess engines can be punished for refusing draws.

The full annotated game

Franklin Chen

This entry was posted in Annotated Games, Endgames, Franklin Chen on by .

About Franklin Chen

Franklin Chen is a United States Chess Federation National Master. Outside his work as a software developer, he also teaches chess and is a member of the Pittsburgh Chess Club in Pennsylvania, USA. He began playing in chess tournaments at age 10 when his father started playing in them himself but retired after five years, taking two decades off until returning to chess as an adult at age 35 in order to continue improving where he left off. He won his first adult chess tournaments including the 2006 PA State Game/29 and Action Chess Championships, and finally achieved the US National Master title at age 45. He is dedicated to the process of continual improvement, and is fascinated by the practical psychology and philosophy of human competition and personal self-mastery. Franklin has a blog about software development, The Conscientious Programmer and a personal blog where he writes about everything else, including his recent journey as an adult improver in playing music.