The stunning defeat of Alphago to Lee Sedol today is in some ways, even more fascinating than its previous wins. We now can finally see some of its weakness's and gain insight on the whole Monter Carlo deep learning algorithm itself.
The game proceeded as the previous 3, and by the mid game, Lee Sedol was a significant disadvantage. In the face of defeat, Lee Sedol spent a good 40 minutes to come up with what 9 dan pro Gu Lee named the 'move of god'. What is telling are the following observations:
1. Lee Sedol's following comment. "This was the only move I could see that worked, there was no other move I could have played.""
2. The placing of the move is very unexpected.
Evidently this worked to Sedol's advantage. As AlphaGo's policy network assigned the move a low weighting (Due to #2), and its being the only move that looked correct made the position appear very good - allowing AlphaGo to fall into the trap.
AlphaGo's predicted win rate dropped massively 9 moves later. After which, a second weakness is revealed.
AlphaGo is dreadfully impatient. It needs to optimized win probability. Thus, will all reasonable moves have low win probability (as she is losing). AlphaGo will be pushed to play moves that are more 'likely' to win - that is moved where it can reverse the game unless the opponent plays at the exact right point of the board. E.g: Capture Races, Ko Threats, and threatening cuts - even if these moves will always lose points when the opponent responds correctly.
In a way, it is funny. The black-box behaviour almost looks like kid throwing a tantrum. The pro commentators were a little confused, but anyone who's about to beat a bot in KGS would see the same behaviour!
So how to beat AlphaGo? Play a divine move in an utterly bleak situation.
Lets see if Sedol can repeat this!