Machines-learning games

Computers have been winning against meatbag chess players ever since Deep Blue beat Kasparov in the 90s. But at the end of 2017 we saw headlines about Google’s deep learning engine AlphaZero ‘teaching itself to play chess’ in four hours, and then beating Stockfish, the world’s best chess engine (see www.chess.com/news/ view/google-s-alphazero-destroys-stockfish-in-100-game-match).

This training took place on some of the most advanced hardware in the world, and probably equates to many years on commodity hardware. So we shouldn’t read too much into that four hours. What’s interesting is the style of play demonstrated by AlphaZero. In the ten games made public ( https://deepmind.com/research/

alphago/alphazero-resources) it seemed to show no concern for sacrificing material, if there was a possibility of a trade-off, and in some cases played moves that appeared to make no sense. This hearkens back to the creative, wildcard style of Mikhail Tal. Boffins have also got AIs to learn video games too, and the machines have surprised everyone. For example, in a paper ( https://arxiv.org/

abs/1802.08842) published earlier this year a team from Freiburg showed what happens when you teach them to play the 1982 Atari classic Q*bert. The agent (using Evolutionary Strategy techniques) discovered a bug in the game that it exploited to achieve a score no human could every hope for.