Alberta AI ready to retire after sweeping the competition at go
AlphaGo, created by University of Alberta alumni, won all three matches against China's top player, Ke Jie, at a go summit in Wuzhen, China, this week.
|Report an Error|
Share via Email
EDMONTON — A University of Alberta computing science professor says a clean sweep by a computer at a man-versus-machine competition in the ancient game of go is bittersweet.
Ryan Hayward says it means the retirement of AlphaGo, the name of the machine that was developed by a pair of University of Alberta alumni.
AlphaGo won all three matches against China's top player, Ke Jie, at a go summit in Wuzhen, China, this week.
Hayward says Deep Blue, the IBM supercomputer that bested chess grandmaster Garry Kasparov in 1997, was retired afterwards and it has also been announced that AlphaGo will switch off too.
AlphaGo was developed by the team at Google DeepMind led by University of Alberta PhD graduate David Silver and former postdoctorate fellow Aja Huang.
Go is considered more difficult than chess for machines to master because the near-infinite number of possible positions requires intuition and flexibility.
"If you took all the atoms in the universe and you counted them and then you multiplied that times itself. There's more continuations of a go game than there is of that number," Hayward said in Edmonton.
"Three years ago if you'd asked me, I would say I would not see in my lifetime a computer program beat a human at the game of go."
Go players take turns putting white or black stones on a rectangular grid with 361 intersections, trying to capture territory and each other's pieces by surrounding them.
AlphaGo previously defeated European and South Korean champions.
Hayward explained that the heart of the computer is algorithms that recognize patterns in a go game. The computer can then access a memory bank of other games with similar situations and figure out which moves in those situations produced the best result.
Hayward said the go community was hopeful AlphaGo could continue to run game scenarios, analysing the advantages and disadvantages of various moves and positions. That won't happen, he said, but DeepMind will publish 50 full-strength AlphaGo-versus-AlphaGo games.
He said Google has already used the algorithm to reduce energy consumption in their server farms by 20 per cent and the algorithms may also have applications for medical diagnoses.
The DeepMind team will produce a paper where they reveal the improvements they made to their algorithms, which Hayward said he's looking forward to.
He said commercial versions of go programs based on some of DeepMind's earlier improvements are already on the market.
"Probably within a few years we will see some commercially available go program that is close to the strength of AlphaGo now," Hayward said.