9

One of the running themes in the commentary I've read on AlphaGo's games is that, from time to time, the AI would make a sound play that was nonetheless completely unfamiliar to its opponent (and to the spectators). Consequently, the opponent was unable to fully appreciate the move and counter it effectively. (or were even thrown off their game entirely for a few moves)

This leaves me to wonder; how important is this effect to AlphaGo's success? Or put differently, when humans have learned from the new sorts of plays they've seen AlphaGo play, will AlphaGo still enjoy anywhere near its current level of success?

  • Aside: the ability to find new, sound tactics is still a great success for the AlphaGo team, even if it does turn out that its victories are more due to novelty than skill, and I do not mean to diminish that feat. – Hurkyl Jul 4 '17 at 22:16
  • One of the things Ke Jie and Deep Mind mentioned after the second match at the future of Go summit was that it was the closest game anyone played against AlphaGo so far, for the first 100 moves. For the first 50 moves he played perfectly according to AlphaGo. Then Ke Jie mentions later that he got excited about how well he was playing and it might have lead him to make some mistakes or less than ideal moves... – snulty Aug 30 '17 at 20:42
  • ...Long story short, it would've been interesting to see how well Ke Jie could've played on if he didn't get too excited in the second match at least according to AlphaGo :) – snulty Aug 30 '17 at 20:44
12

I would argue that AlphaGo's advantage cannot be significantly attributed to the novelty of its moves.

The original public AlphaGo games were those against Lee Sedol, the second ranked player in the world, in March 2016. At that time, as mentioned, several of AlphaGo's moves were novel, and surprising to Lee Sedol and observers.

Then, after players had had nine months to study AlphaGo's play style, in December 2016 it played 59 unofficial online games under the names "Magister" and "Master", including against several of the top ranked players. It won every game. I haven't seen much commentary of those games, but my understanding is that some of the players used moves that had first been observed in the original AlphaGo games. So, at this point I think they were less novel, because they were becoming parts of human players' strategies.

Finally, in May 2017, AlphaGo played a variety of games in the Wuzhen Future of Go Summit, including a match against Ke Jie, the top ranked player in the world. The commentary of those games did not indicate that AlphaGo made a significant number of surprising or difficult-to-answer moves, but rather that its sizeable advantage came from its ability to accurately determine the values of moves, trades, and territory, and its ability to play across the whole board instead of concentrating too much on the currently active area.

  • 1
    In the interests of improving this answer for future users, Michael Redmond and Chris Garlock do a lot of the reviews of the Master series here on AGA's youtube. Also those games were very quick so the ai gets a bit of an advantage there :) I agree though that the future of Go summit showed the strength of AlphaGo, and Michael and Chris have started another series reviewing some of the AlphaGo vs AlphaGo matches released after the summit. – snulty Aug 30 '17 at 20:23
  • 1
    Actually I don't think I meant improving as in this answer could be better, but rather here's a link if future users (or op) is interested in commentary on the master series. It's a good answer as is :) – snulty Aug 30 '17 at 20:46
1

AlphaGo plays moves that are "novel," and that is the key to, but not the reason for its success.

The reason that the moves are considered "novel," is that they have been examined and rejected by human players. So then the question is, does Alpha Go's follow up movies (and variations) prove that the moves are sound or not? The almost invariable answer is the Alpha Go's moves are sound, and that human players had missed the correct followups, some of which are counterintuitive. For instance, it was realized in retrospect that AlphaGo would break off sequences a move earlier or a move later than a human player would, thereby maximizing the sequences' value and beating out alternate lines used by human players. Put another way, AlphaGo could calculate more deeply than human players.

The program does seem to have a few glitches that occasionally make it play moves that are (slightly) unsound in hindsight. Humans have studied these glitches attempting to beat it, so far without success. The reason is, while these moves might be off by a point or two, AlphaGo wlll during the course of a game improve over human play by say, five points, on one or more occasions.

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.