A computer has been created in London that can regularly beat a professional champ at the ancient board game of Go, scientific journal Nature has revealed.
The AlphaGo computer program has been developed by Google’d London-based DeepMind division – in its most successful work yet on Artificial Intelligence development.
AlphaGo defeated the European Go champion Fan Hui in October 2015 and will soon face its ultimate challenge playing against Lee Sedol, the world champion Go player over the past decade. Human and machine will battle in a 5-round tournament game on 9-15 March in Seoul, South Korea.
Go is an ancient Chinese strategy game in which two players try to surround the most territory on a grid board of 19×19 intersecting lines. Despite its easy rules, it is difficult to get a winner when players have comparable experience. The possible number of movements is enormous, making it difficult to plan far ahead or preempt one’s opponent’s moves, so becoming such a thrilling jigsaw for AI experts.
Apparently everything comes down to “deep neural networks and tree search”, David Silver and Aja Huang say in their article in Nature. The mathematical algorithm driving AlphaGo, developed by Google’s London-based company DeepMind, does not only make it able to learn a limited number of moves, but makes it capable of learning on its own directly from raw experience or data – in other words, by playing against itself.
The article explains in detail this new approach “that combines Monte Carlo tree search with deep neural networks that have been trained by supervised learning, from human expert games, and by reinforcement learning from games of self-play.” This type of computer algorithm was dubbed ‘Monte Carlo’ as a codename for top-secret research into the hydrogen bomb in the 1950s.
Robots have beaten humans at board games before. The race started with Noughts and Crosses in 1952, then backgammon in 1992, draughts in 1994, and chess in 1997 – the latter considered as an AI milestone. However, in the game of Go, humans remained unbeatable… until now.
As simple as Go may seem, while for chess the number of possible moves for an average position is 20, it is up to 200 for Go. So AlphaGo’s victory, forecasted by researchers to take at least ten more years, could be a big leap for Artificial Intelligence development.
Nature magazine asked Fan Hui, current European Go champion, how it felt to be beaten by a machine: “the problem is humans sometimes make very big mistakes. Sometimes we are tired, sometimes we so want to win the game, we have this pressure. The programme is not like this. It’s very strong and stable, it seems like a wall. For me this is a big difference.”