Compared to chess, Go has been a far more difficult game for computer engines to master. The larger search space (up to 381 possible moves per ply to search) has meant that brute force techniques haven't been that useful. At first this problem lead to researcher only looking a smaller boards (9x9) but then a modified Monte Carlo method provided the next leap in strength.
Now researches at Google have developed a program strong enough to beat the European Go champion. Using neural nets the researchers were able to train the program to recognise positions and moves that were likely to lead to a favourable position, and from there the program's strength grew. As mentioned in this report, at first the program was very week, but after examining 30 million moves, and playing against itself, it twigged to what was needed to be a strong player. Having conquered Go in Europe, the big challenge would be for the program to take on the best players in Asia. I am not sure whether this will happen (although I do not think Man v Machine matches in Go are banned, as they are in Shogi) as the researchers are talking about moving their research in other directions.
Friday 29 January 2016
Subscribe to:
Post Comments (Atom)
5 comments:
The European go champion is about the equivalent of Japanese/Korean chess champion.
This is essentially right. It beat a 2-dan pro (somewhat dubious, that's a new European title, kind of like spreading titles around in FIDE), which is much behind a 9-dan pro, which is much behind the world best, though Korean dan and Japanese dan are not the same dan. They have bought up a match against Sedol in March (world #1), and the hype machine is rolling as fast as Google can get it. What the terms will be, is not known. Clearly they can ramp up by 100x or 1000x the hardware, but who knows how much that will help?
There is an analysis from a Korean 9 dan (pro) here. https://www.youtube.com/watch?v=NHRHUHW6HQE
Clearly, if AlphaGo does not improve, it will lose unless underestimated.
There was a funny comparison. The human uses about 20 watts (brain power) over maybe 20000 hours of learning (10 years fulltime), while the computer uses thousands of watts over maybe 3000 hours, though 24/7 so it's closer to 6 calendar months. Humans win, by the power-learning curve!
In Denmark we are more and more losing our Grandmaster Peter Heine Nielsen to this little game...
Henrik Mortensen
Post a Comment