

You can now learn to play Go like a supergenius with AlphaGo Teach, a new tool created by Alphabet Inc.’s DeepMind team that cracks open the mind of their undefeated artificial intelligence, AlphaGo.
AlphaGo Teach offers up AlphaGo’s analysis of the 6,000 most popular opening sequences in Go, which are based on data from 231,000 human games, as well as the 75 games AlphaGo has played against real people. The analysis provides a bit of insight into AlphaGo’s thought process, showing where human players would place their pieces and where AlphaGo would place its piece.
DeepMind’s tool also shows the value AlphaGo places on different plays by predicting the win rate of each potential move, and this is where we learn something unexpected about the AI: It does not always choose the move with the highest win rate. Even more surprising, AlphaGo might not choose the same move in two different games even if the boards are identical.
“AlphaGo’s preferred move does not always have the highest value,” DeepMind explains on the AlphaGo Teach website. “This is because each move’s winning probability was computed by running an independent search of 10 million simulations from that position. AlphaGo has some randomness in this search, meaning it might pick different but similar value moves if we run the search again.”
This unpredictability explains how high ranking Go player Myungwan Kim could say that AlphaGo “plays just like a human.” Kim said at the time that AlphaGo’s playstyle was both patient and fearless, but he did not believe that the AI would be able to beat legendary player Lee Sedol. AlphaGo proved him wrong just one month later.
DeepMind decided to retire AlphaGo from competition after it defeated reigning Go champion Ke Jie, but the AI’s story is far from over. In October, DeepMind unveiled AlphaGo Zero, a sort of reincarnated AlphaGo that started learning from scratch, but this time the AI taught itself everything it needed to know without any human intervention or historical data. It took AlphaGo Zero just three days to surpass the AI version that defeated Sedol, and in just over one month it surpassed all versions of the previous AlphaGo.
AlphaGo Zero later demonstrated its versatility by also learning the games of chess and shogi, reaching “superhuman performance” in each game within 24 hours.
THANK YOU