Reinforcement Learning

Deepmind's amazing victories against Go world champion Lee Se-dol have led many to acclaim a new era in AI. As the FT said, the achievement was in a different class to previous machine/human clashes (like IBM Deep Blue/Chess and Watson/Jeopardy):

'Unlike chess, Go permits too many possible moves for a computer to calculate. As a result, the only approach a machine can take is to use pattern recognition to “understand” how a game is developing, then devise a strategy, and adapt it on the fly.'

And as Kottke pointed out what was particularly remarkable was the way in which AlphaGo won, playing in creative and unexpected ways, with moves that left even its creators 'pretty shocked':

'Now machines are starting to be built to think for themselves, creatively and unpredictably. Some emergent, non-linear shit is going on. And humans are having a hard time figuring out not only what the machine is up to but how it's even thinking about it, which strikes me as a relatively new development in our relationship.'

It was also interesting to read about how Deepmind used 'reinforcement learning' (defined as where the machine learns its behaviour based on reward feedback from the environment so it might automatically determine the ideal behaviour within a specific context, in order to maximize its performance'), setting two programs playing Go against each other so that they might learn strategies that they wouldn't have originated on their own and adapt accordingly. Fascinating. Makes me wonder what would happen if you took the same approach in a different scenario but used human teams.