Man vs Machine in the Time of COVID: Komodo vs Smerdon

UPDATE: My match with Komodo will be played on chess.com on the evenings of April 10 and 11 at 8 p.m. New York time, which is Saturday and Sunday morning in Australia (check your local day/time here). You can watch live commentary by GM Evgenij Miroshnichenko and IM Andras Toth on Evgenij’s stream at www.twitch.tv/mironiusvulgaris. Further details here.

 

In 1997, the chess world was rocked by the defeat of the reigning world champion Gary Kasparov at the hands of IBM’s Deep Blue. This sparked a wave of interest in human-computer chess matches over the next decade, featuring other esteemed masters such as Anand, Topalov, Karjakin, Adams, Kasimdzhanov, Khalifman and Ponomariov. The humans never stood a chance. The last human-computer match on even terms was the World Champion Vladimir Kramnik’s defeat by Deep Fritz in 2006. The reasoning behind why Kramnik would agree to play a match in which he was almost surely going to be rushed is perhaps best explained by the 500,000 euros appearance fee, though it will forever be remembered for the so-called “blunder of the century” (Susan Polgar) when the World Champion infamously allowed checkmate in one move.

Some thought this would be the end of human-computer matches. McGill University computer science professor Monty Newborn is quoted as saying after the Kramnik defeat, “I don’t know what one could get out of it at this point. The science is done.” But the advances in computer chess over the last decade have been astronomical. This has particularly been true in the realm of neural network engines, such as the highly publicised victory by Google Deep Mind’s AlphaZero over the (at the time) strongest commercial chess engine Stockfish and subsequent academic papers in Science.

Still, chess players and programmers alike maintain a morbid fascination of pitting humans against these ever-evolving machines. The only way to make a contest out of things is to handicap the computer in some way, usually by removing one or more of its pieces from the starting position. It has been both fascinating and humbling to watch the offered odds growing steadily larger over time as grandmaster after grandmaster has been bested from an initial position that is without a doubt objectively winning for the human.

Next week, I’m the chosen sacrificial GM to explore the frontier of these contests. I was approached by chess computing legend Larry Kaufman to take part in “the first non-blitz match in history in which a GM will take knight odds”, with the results to possibly feature in a special edition of the ICGA journal. I will play six rapid games (15 minutes each plus an extra 10 seconds per move) with the black pieces against Komodo, the 2019 World Computer Chess Champion and currently with an estimated Elo rating of 3421. (Mine is currently 2508. Though calculating expected scores for such a large Elo difference is tough, to give you a sense of scale, the traditional Elo formula estimates that Komodo would score 99.5 out of 100 games against me under normal conditions.)

Still, Komodo may be Komodo, but a knight is a knight (to paraphrase Mikhail Tal). A rapid game is nowhere near as long as a classical game, but neither is it the tactical lottery of a blitz match, so in theory I should be able to avoid outrageous blunders. On the very interesting forum TalkChess, the experts of the computer chess world have given their predictions. These range from a comfortable computer victory to a narrow 3.5-2.5 triumph for Komodo to a comprehensive 4-2 victory by the human (though this pundit referred to me as a “young, active grandmaster”, of which both adjectives are incorrect!).

The closest empirical benchmarks to my match are the informal odds games between the elite grandmasters Hikaru Nakamura and Maxime Vachier-Lagrave on chess.com. These were also played at a time control of 15 minutes each against Komodo, albeit a slightly weaker version. The computer started with initial odds of a knight-for-a-pawn behind. Nakamura’s game ended when he repeated moves to force a draw; Vachier-Lagrave’s game was more disastrous:

Given that I have a 0% score against both these super-GMs, this doesn’t fill me with confidence. But I do have one potentially large advantage over these elite grandmasters. Both Nakamura and Vachier-Lagrave have developed their styles such that they naturally steer towards complicated positions, reasoning (correctly) that they are the better calculators no matter who their opponents might be. But this approach is folly against a machine. Conversely, calculation has never been my comparative advantage in chess. Recognising this while growing up through the junior ranks, I became more of a psychological player, modifying my play to try to exploit my opponent’s stylistic weaknesses and avoid his or her strengths. My ability to adapt my game depending on my opponent is a skill that not many grandmasters employ, but one that I hope will come in handy next weekend.

It may seem that psychological play shouldn’t matter much against a non-human, but I don’t think that’s true at all. When playing hardcore tacticians, I generally have aimed to steer the game towards simpler positions where intuition and patterns count for more than brute calculations, and this will be my approach in the match. Of course, it’s impossible to avoid all tactics, but here I also plan on using an element of bluff from time to time against the machine. A correct chess player should calculate all of his opponent’s tactical possibilities before embarking on a critical continuation, which can use up valuable time on the clock. I can’t afford to do this, but I have the advantage that Komodo will always avoid playing objectively weak moves, even if it’s clear that I haven’t calculated their consequences. So, rather than having to calculate variations to their conclusion, I only need to ‘feel’ that a possible continuation is winning for me in order to stop my calculations and move on. If my intuition is right, Komodo won’t play that way anyway. (Of course, if my intuition is wrong, then this backfires – but in that case, Komodo would win anyway!)

I am not sure how much of an advantage this sort of anti-computer strategy gives me (if any) over the elite grandmaster who have come before me. But on paper at least, I should know a thing or two by now about how to avoid being swindled from a winning position. While I have no idea how the match will turn out, it’s nice to play a small part in the history of computer chess, even if if comes at the expense of my pride.

The games will be played on chess.com on the evenings of April 10 and 11 at 8 p.m. New York time (1 a.m. on April 11/12 in London and a leisurely 10 a.m. on April 11/12 in Sydney). I will probably try to find a way to stream my webcam live, though I’ll be too focussed on the games to comment as well (so if anyone wants to volunteer to start a commentary stream, be my guest). You can watch by going to https://www.chess.com/live, clicking on the “events” tab (the binoculars symbol) and searching for smurfo, my name on chess.com. Viva Humanity!