BREAKING: Knight Slays Dragon
My odds match with the chess engine Komodo is over, finishing in a 5-1 victory to me. It was a fun and very interesting match, which also generated a surprising amount of interest from the online chess community (whose numbers have swelled in recent weeks for obvious reasons). The viewing experience was heightened enormously by the two talented and entertaining commentators, GM Evgenij Miroshnichenko and IM Andras Toth, who streamed the match live. The Day 1 and Day 2 videos are still online.
https://www.facebook.com/dsmerdon/posts/10157436106988667
Though the final score was very convincing, it didn’t seem like things were going to go this way after the first game. I blundered horribly right in the opening, giving back the piece almost immediately. Once we were on equal material terms, Komodo completely crushed me, recording the first (and so far only) official victory of an engine over a grandmaster in a knight-odds rapid game.
This was definitely not the way I had wanted to start the match. Three-quarters of pre-match survey respondents from the chess community predicted a resounding victory to the engine, and during the 10 minute break before the second game, I began to fear that they were right. But then I reminded myself that I hadn’t played a serious chess game since 2017, so inevitably there was going to be some rust to be shaken off. I told myself that this gave me a rare advantage over the engine: while it would play at the same strength in each game, I – in theory anyway – should get better with every game. I also recalled that, as a human, I have the ability to adapt to my opponent as I gain experience, a flexibility that chess computers don’t (yet) possess.
But mainly, I told myself that I had an extra knight.
Things immediately got better in the second game. I tried a new strategy, swapping the queens off as early as I could at the cost of a pawn. I didn’t play the queenless middlegame particularly well and soon Komodo had drummed up some counterplay with running pawns on the kingside. But then came a key moment: 27…Rg8!, a move I spent almost all my remaining time on. This gave up a second pawn in order to exchange a pair of rooks and weaken Komodo‘s structure, leading to a simpler technical position that was easier to play under time pressure.
With scores level but human confidence a little higher, the third and final game of Day 1 was surprisingly easy. Komodo made a poor opening choice and I quickly got a slightly better position without even counting the extra knight. I spent almost a third of my total time on 18…Bxf3!, giving up my beautiful bishop in order to force a trade of queens and another technically winning endgame. This game confirmed that my best strategy was not necessarily to play the objectively strongest moves, but to instead try to simplify the position at all costs.
Overnight, GM Larry Kaufman (the man behind Komodo) made a few small changes to the engine’s software. The biggest of these was to tell it to avoid the French Defence, which Larry and I both agree makes the human’s job much easier. This meant that Day 2 revealed the rest of my pre-match opening strategy, as we played three Stonewall Dutch games in a row. I chose this opening in my preparations after I listed the three types of positions I wanted to avoid against the computer:
- Positions where Komodo has a large space advantage
- Positions that become open and tactical straight from the opening
- Positions in which bishops count more than knights (and in particular, positions where Komodo gets two bishops versus my bishop and pair of knights)
In the Stonewall Dutch, Black immediately erects a pawn blockade in the centre that he tries to maintain into the middlegame and often even the endgame. The board remains locked up for quite some time, which limits the range of tactical possibilities. Black’s king is quite well protected and he very rarely gets checkmated from an early attack. Secondly, getting squashed is not an issue; former World Champion Vladimir Kramnik famously said that the Stonewall is one of the few openings in which Black actually gets a space advantage in the main lines. Finally, the blocked nature means that knights count for much more than they normally do relative to the bishops. In fact, one of White’s main strategies is to obtain a ‘good knight versus bad bishop’ endgame, where Black’s light-squared bishop is made to suffer due to being restricted by its own pawns. What I came to realise is that I didn’t mind these sorts of endgames at all in our match, because my bad bishop would be an extra bishop!
The strength of this opening in odds chess was instantly apparent in the first game of the second day. Black’s blockade was maintained almost until the very end, and White could not avoid successive sets of exchanges that amplified Black’s material advantage.
After this game, I was much more relaxed. It was clear that the Stonewall was an even more effective weapon than I had expected, and given that Komodo was avoiding my French, there didn’t seem to be any real way for the engine to put pressure on me, other than if I made another ghastly blunder. When the Stonewall appeared on the board in the fifth game, I probably relaxed a little too much. On move 26 I decided to sacrifice a pawn in order to swap off all of the major pieces, which had worked well for me in the past games. However, I underestimated how technically difficult the resulting minor-piece endgame would be to win. I eventually found the winning plan starting with 84…Kd3!, walking my king all the way around White’s first rank to attack Komodo‘s pawns from behind. Checkmate finally came on the 103rd move.
The final game was something of an anticlimax. Another Dutch appeared on the board, this time soon turning into a ‘Double Stonewall’ (where White’s pawns on f4, e3 and d4 mirror Black’s on f5, e6 and d5). This setup was once championed by the great Anatoly Karpov, but in odds chess it just meant that all of the tactics were sucked out of the position, and with them any chance Komodo had of using its strengths. It was a slow and rather dull grind to checkmate that was finally delivered on move 85.
And with that, “humanity is saved” (a quote from some chess.com member). Larry and his colleague are already thinking about how to even things up for a future match against a grandmaster, such as making the odds a knight for a pawn, letting Komodo start with two moves at the start, not allowing the human to castle, or randomising the starting squares for the pieces (“Chess960”) so that the human can’t prepare specific openings like the French and Stonewall. These proposals are being discussed on Talk Chess, an interesting forum for computer chess enthusiasts and which features some very bright minds. I couldn’t help but feel a little proud when one user commented that I had “laid an opening template for others to follow” for future engine odds matches – though I was less enthusiastic to read that they had all been cheering for the computer!
I can understand the interest in finding out where the line is drawn that equalises the battle field for engines versus grandmasters. But I am not especially interested in these proposals or in future matches. For me, the most interesting academic question is working out why most of the chess community predicted a computer victory before the match. This includes not only the comments to my Facebook post (see above), but also more than a hundred predictions on FIDE Director Emil Sutovsky’s Facebook post (see below; you may have to click and auto-translate).
https://www.facebook.com/emil.sutovsky/posts/10157700261874681
It’s interesting that very few people predicted a close match. I think part of the difficulty in this prediction task is trying to weigh up two overwhelming but opposing advantages that are difficult to compare: A 900-point Elo advantage (corresponding to a 99.5 expected score out of 100) versus a winning material advantage (more than 300 centipawns). It’s hard to think of a comparable example. One of my friends is a Brazilian jujitsu champion, and in his gym he sometimes spars against weaker belts with his eyes closed. Would the Brisbane Roar beat Barcelona if the Spanish club’s players could only kick with their left feet? Could the 500th ranked tennis player bean Djokovic if the World #1 was forced to use a wooden racquet?
I imagine that in trying to predict what would happen, in each of these cases a normal person tends to focus on just one of the dimensions – the objective strength difference or the handicap – and then makes an overwhelming prediction based on that. Possibly the choice of the dimension has to do with whatever’s most salient to the individual; for example, older chess players tended to predict that I would win, focussing on the material handicap, while players who work a lot with chess computers (e.g. correspondence players) tended to side with Komodo. But even here, there were a lot of exceptions, so I’m not sure how much we can read into this. What is clear is that there were very few close predictions, where presumably the person had made a serious effort at comparing the two wildly different advantages and concluded there was a degree of balance.
Who knows. I’m looking forward to reading Larry’s article in the ICGA journal about the match and hearing what he thinks. But from my perspective, I’m just glad I didn’t add my name to the illustrious list of grandmasters who have been smashed by computers over the years, and happy to slink back into retirement.
I agree with Indrek, this was a great read. To be honest with you I first heard of this match after it had already happened, and therefore I didn’t get to make a prediction. I was, however, very surprised to hear that you won – well done! The reason I would have expected Komodo to win was that I would not have thought the Komodo team would have organised the match under those conditions unless they were fairly confident they could win. At least that’s the impression I’ve gotten based on how these matches have gone in the past (or at least how I think they’ve gone in the past, I can’t actually name any such matches and might be just falling prey to some narrative that ‘resistance is futile’). While as an outsider it is very difficult to predict the relative significance of the handicaps of playing strength and material, I also would have thought that the Komodo team would have the resources to make a pretty good guess.
Kudos for proving me wrong! And congrats on the new book!
Like many others, I also predicted a big win for the engine. The factor I underestimated was how effective your anti-engine strategy would be (swapping pieces, simplifying, avoiding open positions, etc). My initial (obviously wrong!) thinking was that the Elo difference was vastly more important than the material. e.g. if I was playing a knight up against someone 800 Elo higher than me, I am pretty certain I’d lose 99 times out of 100. But against an engine, your clever anti-engine strategy proved to be remarkably effective. Well played, and thanks for your analysis.
Congratulations. Great recap!