A remarkable event took place in London towards the end of last year, when the AlphaZero computer program took on one of the leading commercial programs, Stockfish, in a 100-game match. Astonishingly AlphaZero won by the overwhelming score of 28 wins, no losses, with the remainder of the games being drawn. AlphaZero is the brainchild of Demis Hassabis and his team at Deep Mind. I had the pleasure of playing against Demis in a simultaneous display when he was just eight years old. It was absolutely evident that he possessed an extraordinary intellect, and he has gone on to invent revolutionary processes for creating new types of AI. He was awarded CBE in the New Year’s Honours list. In my opinion, he is nearly up there with Newton and Darwin in the pantheon of British geniuses.
The key to the Hassabis software system is not brute force calculation but an almost infinite capacity to learn by experimentation, based on a vast number of games it has played against itself. In the extracts that follow, AlphaZero produces moves of an extremely paradoxical nature while also overturning orthodox evaluations of material strength. The first extract, in which White retreats the queen to possibly the worst square on the board, is a revelation.
AlphaZero-Stockfish: London 2017 (See diagram 1)
26 Qh1 This astonishing move is played with the idea of exchanging the light-squared bishops. 26 ... Kg7 27 Be4 Bg6 28 Bxg6 hxg6 29 Qh3 Bf6 30 Kg2 Qxa2 31 Rh1 Qg8 32 c4 Another extraordinary move. Despite being a piece and two pawns down, Stockfish plays with a remarkable lack of urgency. 32 ... Re8 33 Bd4 Bxd4 34 Rxd4 Rd8 35 Rxd8 Qxd8 36 Qe6 The key move, enabling White to win material and achieve a won endgame. 36 ... Nd7 37 Rd1 Nc5 38 Rxd8 Nxe6 39 Rxa8 Kf6 40 cxb5 cxb5 41 Kf3 and AlphaZero won in 56 moves.
AlphaZero-Stockfish: London 2017 (See diagram 2)
Another situation where AlphaZero seems unconcerned about having a serious material deficit, in this case two pawns. 33 f4 Qd6 34 Bc1 Nd4 35 Re7 f5 36 Bxf5 Nxf5 37 Qxf5 Rf8 38 Rxd7 Rxf5 39 Rxd6 Rf7 40 g4 Kg8 41 g5. White has regained the material and stands much better in this endgame thanks to the active rooks and mobile kingside pawns. AlphaZero won in 70 moves
I write this from the Gibraltar Masters, now in progress. This week’s puzzle is from a previous event on the Rock.