In case you weren’t aware over the past week there has been an Go tournament running between 18th world champion Go player Lee Sedol and Google’s AlphaGo program developed for their DeepMind neural network. Go for those not aware is an ancient chinese game of area control where players take turns placing stones on a grid and attempting to create a territory that occupies more than half the board to win. What’s interesting about Go compared to other abstract strategy games like chess is the number of possible different games is staggering (10761 compared to that of chess which is estimated at 10120 possible games).
In 1997 IBM’s supercomputer Deep Blue successfully won a chess tournament against the reigning world champion Garry Kasparov winning 3½–2½ . Deep Blue used largely a brute force method to determine moves that went and searched down a possibility tree to a typical depth of 6-8 moves (sometime up to 20 or more depending on the situation) in order to pick an optimum move. Looking back to Go the number of possible games is far to big to attempt a brute force attack so instead AlphaGo has a large database of 30 million moves which it will search through and select a few promising moves, it then will work those moves down a Monte Carlo tree and evaluate each move with a “value network” and “policy network” thereafter selecting the move that has the highest success of improving it’s odds in the game. Interestingly enough AlphaGo doesn’t necessarily choose the moves that will increase its margin of winning, if it is negligibly lower risk to pick a move that wins by a small margin then one that would win with a higher margin it will pick the move with the less risk. (For a more information on AlphaGo watch this video)
So far 4 games have been played between AlphaGo and Lee Sedol with the first three been won by AlphaGo and the fourth by Lee. I’ve watches parts of the first game between them and also the third on the weekend during the live stream. While I haven’t played too much Go (I’m beginner level to amature at best) The livestream is quite easy to follow with a running commentary of the game by a Go professional that shows possible future plays and evaluating each player’s position throughout the game. I’d highly recommend you go watch some of the 15 minute summary videos if you’re at all interested.
A livestream of the 5th and final game starts today at 1pm here.