10 March 2016

Google's AlphaGo beats a Go champion

For awhile now artificial intelligence has been really good at generating winning game play based on the rules of games. Tic tac toe, checkers, sudoku, chess; in just a few short weeks, one of my students in DSP made an AI that was very good at connect-4. (It didn't beat me-- I was only willing to play it once, though!) However, checking the outcome of every possible move will never be possible for a game like Go unless a completely new paradigm of computing comes about. Still, Google has designed a player that can win. Instead of checking the outcome of moves, AlphaGo learns how best to move based on historical data of real games and its own simulations of games. It keeps trying out new games and it sees what happens. Then it keeps a memory of what to play based on the board configuration using deep learning networks, a sort of dimensionality reduction technique that encodes this experience in several layers of equations that can take any input and give the output of what to play next.

This discussion of what AlphaGo means to the future of AI takes several perspectives on what the implications of a Go-winning AI really are. I agree most with Professor Brunskill. She says, "Go is a fixed game: The rules, possible moves and observable information about the game are all prespecified. AlphaGo is not allowed to invent a new move, nor gain new insight by quizzing its opponent. Fortunately the real world is not like this. From the Hubble telescope to vaccinations, people constantly invent new ideas that allow us to transform how we monitor and shape the universe and achieve previously unimaginable outcomes."

I would add that furthermore, not only because humankind can innovate and create new realities do the rules of life change out from under us. New challenges face us every day, like the disappearance of the Malaysian Airlines flight MH370 two years ago this month, or the Zika virus, or locked cell phones of terrorists. Humans have evolved over centuries to face this adversity head-on and adapt for survival; this is exactly why we do innovate, create new measurement technologies, new drugs, and new security protections. Hopefully computers will be able to help us with this process in the not-too-distant future. But before that, machine learning research needs to face the hurdle of learning in a dynamic world.