Go…Go

Google’s DeepMind subsidiary which focuses on artificial intelligence has unveiled a new and improved version of its Go-playing software, AlphaGo Zero, according to a report from The Verge.

The story claims that the new version of the program is a “significantly better player than the version that beat the game’s world champion earlier this year.”

But more notably, it asserts the new version is entirely self-taught.

The Verge explains the original AlphaGo demonstrated superhuman Go-playing ability, but needed assistance from humans to win, building on a dataset of more than 100,000 Go games as a starting point.

The new version has only been programmed with the basic rules of Go, while everything else it learned from scratch.

A paper published in Nature outlines how Zero developed its Go skills by competing against itself. Starting with random moves on the board, every time it won, Zero updated its own system and played itself over and over, millions of time, “learning” as it went.

The Verge reports it took AlphaGo Zero only three days of self-play to become strong enough to defeat the version of itself that beat 18-time world champion Lee Se-dol, winning 100 days to nil.

Lead programmer David Silver had this to say at a recent press conference about the advance:

By not using human data — by not using human expertise in any fashion — we’ve actually removed the constraints of human knowledge. It’s therefore able to create knowledge itself from first principles, from a blank slate…This enables it to be much more powerful than previous versions.

 

Elon Musk had better speed up development of that Mars program.