The architecture that beat humans at the notoriously CPU-impervious game Go, AlphaGo by Google Deep Mind, was converted to allow the machine to tackle other “closed-rules” games. Successively, the program was given the rules of chess, and a huge battery of Google’s GPUs to train itself on the game. Within four hours, the alien emerged. And it is indeed a new class of player.
The Alpha Zero neural network uses reinforcement learning to teach itself things from scratch. It does not rely on previous knowledge – which in the case of chess is surprising, as the mass of knowledge on the game accumulated in centuries of experimentation is hard to shrug off. Combined with a powerful search algorithm, the neural network is at present unbeatable. This was demonstrated in a 100-game match against the strongest chess program around, Stockfish 8.
What impressed me when I saw a few games from that match, which was concluded with 25 wins and 75 draws, no losses from Alpha Zero, is that the machine can display an evolved treatment of openings, is keen to sacrifice material for positional gains, and has no prejudices.
I await patiently for the time when somebody will publish the 100 games and comment them – I am sure there’s a treasure of things to learn there.
Read full, original post: Alpha Zero Teaches Itself Chess 4 Hours, Then Beats Dad