What would you think if you taught your child the rules of chess at
breakfast and found that by lunchtime she had beaten the world champion?
Awe? Parental pride? A bit of fear perhaps?
That's essentially what just happened at Google's DeepMind subsidiary in London. They created an ultra-powerful game-playing computer system called AlphaZero that implements a n eural network capable of deep learning through reinforcement.
Unlike other chess-playing programs--which have outperformed humans since IBM's Deep Blue beat the human world champion Gary Kasparov
in 1997--AlphaZero was not pre-programmed with any specialized
knowledge or expertise about chess. It was simply given the rules of the
game and allowed to learn by playing against itself.
Four hours later AlphaZero crushed the World Computer Champion, Stockfish, with 28 wins and zero losses in a 100-game tournament (the remaining games were ties).
The AlphaZero algorithm developed by Google and DeepMind took just four hours of playing against itself to synthesise the chess knowledge of one and a half millennium and reach a level where it not only surpassed humans but crushed the reigning World Computer Champion Stockfish 28 wins to 0 in a 100-game match. All the brilliant stratagems and refinements that human programmers used to build chess engines have been outdone, and like Go players we can only marvel at a wholly new approach to the game.
Other chess experts describe AlphaZero's play as "divine," or "from another galaxy."
As if one superhuman feat were not enough, the AlphaZero team used the same artificial-intelligence (AI) system to tackle the games of Go and the Japanese chess game, Shogi. It took AlphaZero just two hours of play against itself to surge past Elmo, the Shogi Computer World Champion, and all of 8 hours to surpass AlphaGo (another DeepMind program), which itself dethroned the human Go champion, Ke Jie, earlier this year.
So, to summarize, in less than a day, starting as a blank slate and knowing nothing more than the rules of the games, simply by playing against itself, AlphaZero reached a superhuman level of play in three abstract games that have challenged humans for millennia. Not a bad day's work.
And just in case you're thinking that AlphaZero reached these superhuman levels simply by calculating faster than any other computer, that's far from the case. It is blindingly fast compared to humans--for example searching 80 thousand chess positions per second, but it is tortise-slow compared to other chess-playing systems. Stockfish, which AlphaZero completely dominated, searches 70 million positions every second. The system's creators explain, "AlphaZero compensates for the lower number of evaluations by using its deep neural network to focus much more selectively on the most promising variations -- arguably a more 'human-like approach to search."
In recent years some very smart people including Bill Gates, Elon Musk and Stephen Hawking have warned about the threat posed by out-of-control artificial intelligence. While the superhuman learning and game-playing of AlphaZero seem benign, and ubiquitous, mostly invisible AI applications help us every day in a huge variety of areas, there are red flags raised by the potential for robo-cops and soldiers, vital infrastructure managed by AI, increasingly autonomous robots capable of replacing most workers, and the potential for super-intelligent AI creations that may not have the interests of humans at heart. AlphaZero, for example, could equally well learn to "play" at a superhuman level at politics, finance or war.
Gates and others emphasize that we need to figure this out before such intelligences emerge because once they do, like AlphaZero, they could leave us in the dust within a few hours.
So if your child became the world champion chess player after four hours of play, wouldn't you be scared? I would.
-----
You can read the scientific paper describing AlphaZero's accomplishments here.
4
3
2
