Chess: Best Human vs. Best Computer: The Chess Muse Stands Up
Recently, world chess champion Vladimir Kramnik was playing what many consider to be the best chess program ever, Deep Fritz 10. Four games were drawn (i.e. ended in a tie). One game was blundered away on a one-move checkmate; Kramnik lost. But the last game, the Fritz program won outright. What I find interesting, is that this last game (game 6) contained moves that the other grandmasters in the room could not understand, and which some of them even disparaged. Here is a quote from George Dvorsky's excellent article:
Fritz made moves during the tournament that left other grandmasters scratching their heads wondering how and why it did what it did. In many respects, the internal machinations of the computer is beyond human comprehension.I don't think this statement about "comprehension" is quite right, for the moves are computed by an algorithm whose goal was to quantifiy a position so as to determine which move is best (as determined by the program's metrics), when given certain input paramaters. The metrics were designed by a human, and the algorithm was implimented by a human.
Instead, it's probably better to claim that the computer's play is beyond the "intuition" of even the best grandmasters under human-standard tournmament conditions. Stephen Moss, writing for The Guardian, seemingly agrees: "Deep Fritz played some brilliantly counterintuitive moves in game six, and made some manoeuvres that were so ugly, they were beautiful."
I'd argue they were beautiful in the sense that a very efficient math proof is beautiful -- the objective is achieved with a minimum of wasteful steps. However, the observed pattern of efficiency in achieving a win (or superior position) exhibited by computers does not have the kind of aesthetic appeal that has served grandmasters so well up to today. One can image, as computers continue to double in power every couple of years, the games might become what looks like an entangled mess, but which actually amounts to supremely efficient goal achievement.
But Moss goes on to draw a conclusion with which I cannot agree:
Computers are now so much better than humans [...] that man-v-machine matches have become pointless. Sponsors love them; manufacturers of chess programs love them; but chess players recognise them for what they are - sideshows, games of academic interest only. Deep Fritz is reckoned to calculate 10 million positions a second; the human brain simply can't calculate that fast or that far ahead.[...] Machines, programmed with books of opening theory and tables to allow them to calculate perfect endgames, can play chess better than humans. Checkmate. Now let humans get back to playing other humans in a thrilling struggle between fallible, organic entities.Monty Newborn, speaking at the 2nd World Computer Chess Championship in Toronto way back in 1977, made a controversial statement: "In the past Grandmasters came to our computer tournaments to laugh. Today they come to watch. Soon they will come to learn." One big lesson Grandmasters can learn is how their default intuitions are operating on the board, since their own moves and positional preferences are also being quantified under the computer's model. Thus, a grandmaster can learn certain weakness in his own psychology of board evaluation. A well-known chess teacher, Alexander Roshall, who is also the editor of a Russian chess magaine, claimed that just such a weakness was exposed for Kramnik in game two, where an odd tactical position arose that grandmasters generally just don't encounter:
[T]he mating pattern that occurred during the game, with the white queen protected by a knight on f8 [...] is extremely rare in chess. It is not one of the patterns that chess grandmasters automatically have in their repertoire. This was confirmed by a GM commentator in Bonn, who after Kramnik's move did not notice that it was a blunder and started discussing White's options – but not the mate in one. Alexander Roshal assured us that, had the white knight somehow moved to [a more commonly played square nearby the King] Kramnik would have seen the mate in micro-seconds, [as if the square] would have had a big red light blinking on it, Roshal said, because this kind of mate (or mating threat) occurs quite often in chess, and the mating pattern would be firmly anchored in his mind. With the knight in an unusual position the square remained dark and Kramnik simply did not see the danger.This kind of analysis shows, contrary to Moss, that human-machine matches are not pointless, and here are three reasons why: First, finding such blindspots in professional player strategy is an important reason to continue having human-computer tournaments. Second, having the computer analyze which players had (or have) the strongest attacks, deepest tactical intuitions, best endgame understanding, etc., would help justify which players in the history of the game were true stand-outs in their area of speciality. (This is a constant argument among followers of the game.) Third, there are always new search structures that can be tweaked, tested and evaluated (see Google video: The History of Computer Chess: An AI Perspective), and these might yeild identifiable "styles" of chess play, styles which might accurately map the styles of any past human player of note.
I like the picture which opens today's blog entry, as it shows a young woman, which I'll dub 'The Chess Muse', standing beside an empty chair. With the tournament victory of Deep Fritz 10 over the world champion, The Chess Muse has stood up. No doubt she plans to walk to the other side of the table, where there is a player who needs no chair.
 "The Future of Chess" George P. Dvorsky's Blog (Accessed 12/9/2006)
 Stephen Moss, "Man v machine (and guess who won)" The Guardian Unlimited (Accessed 12/8/2006)
"How could Kramnik overlook the mate?" Chessbase news (Accessed 12/09/06)
see also: "Kramnik Biography" Vladimir Kramnik World Champion Official Web Site