4
$\begingroup$

We can read on wiki page that in March 2016 AlphaGo AI lost its game (1 of 5) to Lee Sedol, a professional Go player. One article cite says:

AlphaGo lost a game and we as researchers want to explore that and find out what went wrong. We need to figure out what its weaknesses are and try to improve it.

Have researchers already figured it out what went wrong?

$\endgroup$
5
$\begingroup$

We know what Lee's strategy was during the game, and it seems like the sort of thing that should work. Here's an article explaining it. Short version: yes, we know what went wrong, but probably not how to fix it yet.

Basically, AlphaGo is good at making lots of small decisions well, and managing risk and uncertainty better than humans can. One of the things that's surprising about it relative to previous bots that play Go is how good it was at tactical fights; in previous games, Lee had built a position that AlphaGo needed to attack, and then AlphaGo successfully attacked it.

So in this game, Lee played the reverse strategy. Instead of trying to win many different influence battles, where AlphaGo had already shown it was stronger than him, he would set up one critical battle (incurring minor losses along the way), and then defeat it there, with ripple events that would settle the match in his favor.

So what's the weakness of AlphaGo that allowed that to work? As I understand it, this is a fundamental limitation of Monte Carlo Tree Search (MCTS). MCTS works by randomly sampling game trees and averaging them; if 70% of games from a particular position go well and 30% of games from another position go well, then you should probably play the first move instead of the second move.

But when there's a specific sequence of plays that go well--if, say, W has a path that requires them playing exactly the right stone each time, but B has no possible response to this path--then MCTS breaks down, because you can only find that narrow path through minimax reasoning, and moving from the slower minimax reasoning to the faster MCTS is one of the big reasons why bots are better now than they were in the past.

It's unclear how to get around this. There may be a way to notice this sort of threat, and then temporarily switch from MCTS reasoning to minimax reasoning, or to keep around particular trajectories in memory for consideration in future plays.

$\endgroup$
  • 2
    $\begingroup$ Also, we knew that AlphaGo was better at playing White than it is at playing Black. This is why Lee suggested (and Deepmind agreed to) Lee playing Black on the last game, rather than the coinflip that it was originally planned to be; Lee wanted to see if the same strategy that worked against AlphaGo's weaker side also worked against its stronger side. (Not well enough; he lost that match.) $\endgroup$ – Matthew Graves Aug 9 '16 at 15:34
  • $\begingroup$ What's the difference between playing black or white?Does one of the colors start first? Even if it does, it seems weird to notice a difference in the performance. $\endgroup$ – Ovi Apr 25 '17 at 20:30
  • $\begingroup$ @Ovi Yes, Black always goes first. This is known to have the long-lasting effect of putting White in a position where he usually struggles to play as efficiently as Black in local battles where Black is up a stone (White is given a small score handicap at the end to compensate). Perhaps AlphaGo is better at exploiting human mistakes than it is coming up with creative and flawless strategies? $\endgroup$ – Feryll Apr 30 '17 at 20:03
  • $\begingroup$ @Feryll no it's not, note the sequence recommended here alphago-games.com/view/eventname/leesedol/game/3/move/37 (B after whites "brilliant" move 78) That is to say Lee Sedol's "brilliant move" may not work Idk anything about A.I. , I just find it so hard to believe that alphago did not find this counter $\endgroup$ – Hao S Feb 12 at 0:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.