Humans vs machines. Sarah Connor vs the Terminator. Neo vs Agent Smith. The battle between the human brain and artificial intelligence is more than a plot device; it’s an academic endeavor. Back in 1997, AI was making the news as IBM’s Deep Blue computer beat chess champion Gary Kasparov. Since then AI has come on leaps and bounds but has always fallen short of rising to its biggest gaming challenge: to beat a human at Go.
DeepMind, now owned by Google, took on the challenge when they created an AI called AlphaGo. Employees working on the project are experts in deep neural networks, reinforced learning, and neuroscience. The DeepMind algorithms are inspired by the human brain and can actually learn from experience. In October, DeepMind’s AlphaGo made history by beating a professional human for the first time in the form of European champion Fan Hui. This was a huge victory and major milestone for Google DeepMind, but could AlphaGo beat the very best? This morning AlphaGo played the first of 5 matches against Lee Se-dol, current world champion and professional for 21 years. It’s now official: AI can beat the world’s best human Go player.
#AlphaGo WINS!!!! We landed it on the moon. So proud of the team!! Respect to the amazing Lee Sedol too
— Demis Hassabis (@demishassabis) March 9, 2016
Playing the game
Go is played on a 19×19 board and involves capturing pieces by surrounding them. All the pieces are the same and move the same way. It’s a simple game in terms of rules but exceptionally complex in terms of strategy. With such a large board and so many pieces, AI can’t memorise every single possible scenario and figure out the best move with brute force methods. Instead, AlphaGo has to learn how to win at Go. There just isn’t enough time or processing power to imagine every possible game and that’s the case for us humans too.
Chess is also a complex game but the best computers can think ahead and analyse every possible move and reliably calculate the value of each move and decide which is most likely to go well. It’s long been thought that computers could never beat professionals at Go because we just don’t have the technology to use that same approach. We either need future computers that can calculate the dizzying number of moves or we need a computer with a human brain.
AlphaGo plays Go, messes up, and learns from the experience. It watches Go games and is asked to predict the winner. Over time, it sees where it goes wrong and starts to learn how humans play the game. It doesn’t become smarter with some improved code; it becomes smarter from literally learning for itself. There are too many possibilities to really calculate the value of a move like in chess, but AlphaGo can get a feel for who is winning by looking at the board. It thinks about its previous matches and forms strategies from experience.
Rising to the challenge
In press conferences Lee had described being wary but confident that he would win outright. He told reporters that he would probably win 5-0 or 4-1 so his goal was to not lose a single match. The first match started at 4am and Lee is already on the back foot as AlphaGo won. Lee will need to play his best in the coming days because he’ll want to make sure the AI doesn’t win any more matches, and because the winner walks away with a cool $1 million.
This morning’s result doesn’t change the game of Go, just as Deep Blue’s victory didn’t change chess. The importance isn’t for the game but for the technology. How smart is an AI really if it justs runs through every possible move and chooses the action that is statistically most likely to be good? That approach won’t work for Go and DeepMind has succeeded by getting a computer to learn like a human. Where that technology goes from here is really interesting. Healthcare? Climate modelling? What problems can DeepMind solve if it can learn better than any other computer but can do it all better than us?
The matches continue at 4am on the 10th, 12th, 13th, and 14th and they are all live streamed on the DeepMind YouTube channel. You can watch the first match here with commentary:
Main image © Wikicommons/Donar Reiskoffer