How (not) to play chess against computers

Part IV: How (not) to play chess against computers…

In Part III we saw some evidence that humans tend to wear themselves out very quickly in a head-to-head match against a tireless computer. To combat this, a top grandmaster would be well-advised to stick to very short matches against computers. Even a six-game match is quite long enough to cause serious problems for the strongest and healthiest humans, by the time Games Five and Six roll around.

Are there other ways that human players should adapt, in their ongoing struggles to gain an edge against computers? Last time we focused on physical strength and chess expertise. This time, I would like to investigate how computers and/or humans can modify their own playing style, in an effort to gain an advantage against the other.

(3) Playing style against specific opposition

It appears that computers can improve their "style" against grandmasters in general. A fairly recent innovation is the specific tuning of software to play more successfully against humans, even if the actual chess moves are not objectively stronger. Chess Tiger, one of the top commercial programs, has a "sibling" named Gambit Tiger which was specifically configured to be more successful against human opponents. And X3D Fritz has supposedly been tuned specifically to defeat human players, and even more specifically, to defeat Garry Kasparov.

There is certainly a lot that a programmer can do to "tune" a program in this way. Analyze all the games between humans and computers, and look for the types of position where computers do unusually well. Computers don't do as well when all 16 pawns are on the board? Just put in a "contempt factor" so that your computer will devalue positions with all 16 pawns, and aim for an early opening of lines.

Similarly you could analyze all known games of a particular opponent. Garry Kasparov plays 50 points stronger than usual when he has both bishops still on the board? Give a special 0.25 pawn bonus to the computer whenever it is looking at a position where one of his bishops has been exchanged.

I just made those numbers up, by the way; I don't really know which types of position are particularly suited to a computer, or to Garry Kasparov. It sounds like a fun study, but right now I can't really tell you whether this "tuning" is worth a lot or a little. What I can tell you, is that I have done some analysis about openings. I have looked into which openings have historically been problematic for computers against humans, and which openings have worked better for computers.

If I told you that the King's Indian Defense, Sicilian Defense, French Defense, and Ruy Lopez were four of the openings that showed the most significant bias toward one side or the other, in human-computer games over the past five years, you probably wouldn't be too surprised. After all, computers are supposed to play extra-well in tactical positions, and extra-weak "positionally". And the King's Indian and Sicilian are "known" to be tactical openings, whereas the French Defense and Ruy Lopez are viewed as being more "positional".

However, what if I told you that computers do poorly against humans in the King's Indian Defense and Sicilian Defense, and unusually well in the French and Ruy Lopez?

Maybe there's not enough games to establish the true pattern. I'm not going to claim that the conventional wisdom is "wrong". I will say, however, that the empirical evidence does not remotely confirm any claim that computers are stronger in the Sicilian and King's Indian, or weaker in the French and Ruy Lopez. It is time for a better distinction than "tactical" vs. "positional".

Earlier on, I pointed out that one way humans could "improve" against computers is to remove something negative from their play. We saw that scheduling shorter matches is one way to remove (or at least) limit the fatigue factor. There is another important factor, a stylistic factor, which may also be hurting humans, but which is much harder to correct for. That is the irresistible urge to play "in the grand style", to sacrifice material.

Stefan Fischl has to receive some of the credit for noticing this. We jointly discovered, in a series of online postings to the ChessNinja message boards in August, that there is a definite relationship between a player's sacrificial style against humans, and their success against computers.

Human players who frequently give their human opponent a pawn, or even a piece, and still win, will perform worse against computers than their ratings would suggest. Players who rarely win by sacrificing material, on the other hand, will do considerably better against computers than their aggressive counterparts.

For instance, Yasser Seirawan, John T.H. van der Wiel, and Joel Benjamin tend to sacrifice material and then win, against similarly rated humans (within 100 Elo points), about 7%-9% of the time, although the grandmaster average is around 10% or 11%. They have all done considerably better against computers than their rating would suggest. On the other hand, Garry Kasparov, Judit Polgar, and Alexander Shabalov have a "sacrifice percentage" that is twice as high, or even more. All three players are quite famous in the chess world for their bold, attacking style. And all three of those players have historically played at a level at least 200 rating points lower than you would expect, against computers!

Here is a chart showing this effect for 30 different contemporary players. The players on the right side have a lot of sacrificial wins; the players on the left don't sacrifice as much. The players near the top do better than expected against computers; the players near the bottom do worse than expected.

Surely Garry Kasparov doesn't need me to tell him that it is dangerous to sacrifice material against a computer. I think he probably already knows that! Although the Kasparov of six or eight years ago may have tried to use the same approach against computers that he did against humans, I don’t think we will see the older and wiser Garry Kasparov playing so far below his true strength, this time around.

The main point I am trying to make is that some of the top grandmasters can still "improve" further against computers, by suppressing their own natural urge to sacrifice and win spectacularly. If grandmasters are trying to improve by changing their style of play, and computers are trying to improve by "tuning" specific anti-human parameters, it may well turn out that grandmasters are able to win that particular race. For one thing, it is quite easy for a human to sit in the safety and privacy of their own living room and experiment with playing styles against a strong computer.

I don't know what the future will hold for grandmasters and chess computers, but I believe there is still hope for the grandmasters, if they continue to adapt. Grandmasters are constantly improving at chess, and learning more about how to change their own playing style against computers, and that may be enough to maintain the balance, or even pull ahead again.

In Part V, the final article in the series, Jeff Sonas will give us his statistical predictions for the match between Garry Kasparov and X3D Fritz.