Demis Hassabis, DeepMind’s CEO and founder, today announced the historic news that its machine, AlphaGo, won its third game in a row versus 18-time Go world champion, Lee Sedol. With the third win under its belt, that’s the five-game match now sealed. A machine has officially beaten the world’s best player at a game which is widely considered to be very difficult to teach a machine.
#AlphaGo won game 3 and the match! Historic moment. In complete awe of Lee Sedol’s incredible genius, and proud of the amazing AlphaGo team!
— Demis Hassabis (@demishassabis) March 12, 2016
That takes AlphaGo up 3-0 versus Sedol, who initially expressed surprise at losing the first game, but was left “speechless” following the second. This victory is the second landmark win for AlphaGo, following its first 5-0 win against another human player in the latter months of 2015.
Unlike a game of chess, machines can’t be taught to play Go by algorithms alone. AlphaGo needed to learn 30 million or so moves, as well as learning to create its own strategies, like a human player would. Like a person playing the game, AlphaGo developed its own strategies by playing internal games countless times and developed its own methods. What’s interesting about watching AlphaGo play is that it doesn’t think like a human, and so doesn’t play like a classically-taught Go player would.
What this indicates is far more important than a single match-win versus a world champion. It shows that machines can combine their ability to be taught algorithms and how to crunch incredibly challenging numbers, but also, they can be taught to learn to a level beyond what a person can. Watching DeepMind’s work over the coming years is going to be intriguing, to say the least.