Remember Google’s program that learned to beat champion Go players? It’s been comprehensively beaten by its successor, which learned the game from scratch.
AlphaGo Zero, which worked without any game history data, beat the original AlphaGo in a 100 match series by a margin of 100 games to zero.
Go has always been a challenge for computers as it involves a huge range of possible moves at any stage. While computers can process data quickly, they aren’t necessarily that good at weighing up multiple options at once, which is where artificial intelligence comes in.
The original AlphaGo program learned by assessing data from thousands of real-life games played by humans. This let it figure out patterns and learn techniques such as looking for the first move that met a particular threshold of likely success rather than going through more options to find the “perfect” move.
AlphaGo Zero didn’t get this game data. Instead it was programmed with the rules of the game and then left to play against itself and figure it out from there. Rather than conclude it was a strange game where the only way to win was not to play, it took just three days to get good enough to unleash a whooping on AlphaGo.
It seems the “secret” to AlphaGo Zero’s success might in fact be its lack of humanity. Because it’s physically impossible to consider every possible move, expert human players develop great instincts for having a general idea what type of move will be good. By learning from human game data, AlphaGo was restricted by trying to turn these instincts into logical rules, in effect meaning it played as if it were an expert human player with immense brain capacity.
By starting without any human-like preconceptions, AlphaGo Zero was able to develop strategies more suited to its capabilities. It still needs to be tested against human players, but one expert who analyzed the inter-computer games says it used techniques he had never previously seen.
Google’s hope is that such an approach might work in other areas of artificial intelligence, with computers that develop techniques and procedures that make best use of a computer’s capacity rather than trying to refine the way human brains approach tasks.