Saturday, May 7, 2016

AlphaGo's Conquest Of A GO Master: Sending Shockwaves Around the GO Gamer World

Image result for lee sedol
The thrashing of Go Master Lee Sedol was so complete, 4 games to one, the South Korean could scarcely believe it himself.  But it was true and there was no way escaping the fact. The "Deep Mind" machine had bested the human in convincing ways. This, despite the fact that many observers (the games were streamed online to millions) believed AlphaGO had made one or more mistakes. But in the end all the critics were forced to concede that perhaps the machine had used strategies its human masters had overlooked.

The two games that stood out for their oddities were games 4 and 5, given it appeared the machine had made fatal errors. In game 4, a 'tesuji' (brilliant, clever move) by Lee seemed to throw the machine off its game, since the move had not been foreseen. Indeed, its next dozen moves were genuine errors. This indicated at least some weaknesses in the deep learning method used to direct AlphaGO's plays. And coincidentally, this was the only game Lee won.

Game five was a different story in that it showed how difficult it was to consistently exploit the machine's mistakes. Most human observers reckoned the machine had made a serious error early on but not even the GO master Lee was able to capitalize on it - with the AlphaGO clawing its way back into contention and a final win - to the evident despair of Lee.

All of this is of interest since as I had posted about earlier, the game of GO had hitherto been a redoubt of human superiority (as when compared with the ignoble chess defeats of humans such as Gary Kasparov by the likes of IBM's Deep Blue).  There are many reasons including that GO is fiendishly more complex given it has more potential moves than the atoms in the universe,

"Deep learning" has been the difference, whereby the programs-computers are allowed to extract patterns from masses of data including assimilating the moves from human to human games as well as games played with other computers and with itself. The combination has apparently proven too much for any modern GO master to surmount.


And Lee Sedol doesn't appear to be too eager to go up against his machine opponent anytime soon.

See also:

http://venturebeat.com/2016/03/12/go-board-game-champion-lee-sedol-apologizes-for-losing-to-googles-ai/


And:

http://www.wired.com/2016/03/final-game-alphago-lee-sedol-big-deal-humanity/



No comments: