chess: Deep Blue all at sea

Where do Deep Blue's programmers go after the defeat of their monster by Kasparov? The post-mortem on the games reveals clear causes for its demise, but there are great problems to be solved if a later version is to be inoculated against the fatal illness. Here is a three-point plan for computer domination of chess:

1. Don't make any drastic changes: Deep Blue did, after all, win one game against Kasparov. Its tactical control of the opening game was impressive. Nothing must be done that will interfere with its main advantage over humans: its ability to calculate long variations with perfect accuracy.

2. On the other hand, some drastic changes are needed to improve its long-term judgement: In the second, fifth and sixth games, Deep Blue accepted positional disadvantages that any human would have avoided instinctively.

Look at the diagram position (above right) after Kasparov's 28.a4 in the final game. By playing 28...Ra8? 29.a5 a6?? 30.b6 Bb8?? the machine condemned its rook and bishop to a living death. One cannot play like that without knowing for certain that Black will eventually free the trapped pieces, perhaps by playing f6 and e5, but even a human can see that such a dream is overly optimistic.

There is a trade-off between calculation and judgement in any computer chess program. The more sophisticated its judgement, the fewer positions it has time to analyse. Yet if it makes judgemental mistakes of this magnitude, there must be something deeply wrong with the manner in which it assesses positions.

3. Cut down the pollution of human intelligence: In the third game, Deep Blue adopted a worse move in the opening than it had played in the first game. It transpired that the machine's grandmaster consultants had been "improving" its openings book between games. Also, the machine lost the fifth game after its operators had turned down an offer of a draw. Deep Blue needs to be left plugged in for a few months to re-write its own openings book. If it is really clever, it might even devise a repertoire for itself, tailored to defeat humans. And it should be allowed to make its own decisions on draw offers. No wonder its play in the last game- and-a-half looked so sulky. Had it accepted the draw in game five, Kasparov would not have realised how badly it was capable of playing, and the illusion of artificial chess intelligence might have been sustained.

We can be reasonably sure that improvements in hardware will eventually be enough to overcome all these problems, and with the speed of computer processing doubling every 18 months, we may not have very long to wait. For now, however, programming a machine to play chess has been made to look a more difficult task than playing chess itself.