Chess / Deep Blue seizure

Deep Blue, the grandmaster-crusher from IBM, gave a most impressive account of itself last week, but predictions of the downfall of humanity may be premature. For Deep Blue has its Achilles chip, too.

Its chief programmer, Feng-Hsiung Hsu, admitted: "There are certain kinds of positions where the [human] grandmaster has an advantage [over Deep Blue] and we don't know how to program such positions yet." He was unwilling to give more than the vaguest hint of the type of position he meant, but two particular positions may provide a clue.

In the diagram above, any thinking human playing White will realise that he cannot lose if he does nothing. The black rooks and bishop can never emerge from behind their wall of pawns. But if White plays 1.bxa5, Black will eventually organise a breakthrough with ...b4 and win. Most computers, however, will play 1.bxa5 on the grounds that it's better to be one rook behind than two.

Some machines have been told about totally blocked pawn structures, but even those are all at deep blue sea if we make a slight modification.

Replace the pawn on b4 with a white bishop. Any human who has seen the earlier position will not fall for the trap of 1.Bxa5? which loses (if Black finds 1...b4+). Keep moving the king, however, and a draw is secure.

There are two separate problems for the programmer here. The first is the concept of a fortress - a defensive wall that can never be breached. Machine thought has problems with the concept of eternity. Still more difficult, however, is the idea of generalisation. A human who has seen the first position has little difficulty with the second. Machines so far have no way to learn from example, and add to their conceptual vocabulary. Solve that problem, and machines can take over.