Google's DeepMind-powered AlphaGo artificial intelligence has officially beaten human champion Lee Se-dol at the notoriously complex game of Go, but failed to repeat its clean sweep against Fan Hui.
Go has long been a target of artificial intelligence research, thanks to its gigantic problem space. Where chess has a possible 10-to-the-power-of-120 games, Go boasts 10-to-the-power-of-761 - despite having just three rules - making it far too complex for even the most powerful computers to attack in a brute-force manner. Earlier this year, though, advertising giant Google boasted of a breakthrough
: a DeepMind-powered artificial intelligence dubbed AlphaGo, which proved its mettle by besting three-time European Go champion Fan Hui five games to zero.
To prove its victory was no fluke, Google set up a match against champion Lee Se-dol with a $1 million prize pot. Prior to the first match Lee was confident of victory, predicting he would either sweep the board 5-0 or potentially lose a single game to the AI as he got used to its playing style. That prediction, however, proved inaccurate: AlphaGo won the first match
, rattling Lee, and won the next two to take the overall win.
Unlike its battles with Fan Hui, AlphaGo's victory wasn't a clean one: after losing three matches in a row, Lee was able to secure a victory last night as he spotted some weaknesses in the AI's play style. Although not enough to make a difference to the overall outcome of the competition, Lee's victory has proven useful for the team behind AlphaGo who now know where its limitations lie and can develop means of improving its play still further.
The final match, played now for pride rather than victory by Lee, is scheduled for Tuesday.