AlphaGo Zero

Game of Go. © iStock.

Google’s DeepMind lab has developed AlphaGo Zero, an artificial intelligence programme that is exceptional because it was able to teach itself to play Go. Google says the programme independently acquired “superhuman abilities”, becoming virtually unbeatable in just a few days. AlphaGo Zero started out not knowing a single strategy for playing Go, a centuries-old Chinese game in which two players use black and white stones to control as much territory as possible. It took AlphaGo Zero just under four days to beat its predecessor, AlphaGo, which in turn beat the best human player in the world, South Korean Lee Sedol, just last year. “The most important idea in AlphaGo Zero is that it learns completely tabula rasa — that means it starts from a blank slate and figures out for itself, only from self-play, without any human knowledge, any human data, without any human examples or features or intervention from humans”, said lead AlphaGo researcher David Silver in a Nature interview. Which isn’t to say that DeepMind has developed an AI capable of doing anything. To train itself, AlphaGo Zero has to be limited to a problem where clear rules limit its actions. Within that framework, it can discover strategies that would never have been thought up by humans, which is when it is at it’s most impressive.

⇨ Ars Technica, “New neural network teaches itself Go, spanks the pros.”