Hacker News Reader

Comment by ktRolster | Hacker News Reader

Will he learn from the matches against Lee Sedol? Will Alpha Go become even more advanced (ie, have more hardware added)?

View on HN

Considering he's beaten Lee Sedol and Gu Li, who knows how this will play out.

But no doubt he'll talk to Sedol about his take away from playing AlphaGo, especially since part of playing against any computer is reverse engineering its decision tree.

There were some games with Gu Li? Could you point me to them? I can't seem to find them.

Gu Li and Ke Jie played each other (in a jubango -- a ten game series -- in 2014: an excellent series of games!). AlphaGo did not play Gu Li unfortunately.

Ah, I read that wrong, I thought he meant Alphago not Ke Jie.

Exciting times, I can't wait. I guess the later the game will be, the less chance Ke Jie has.

The jubango was between Gu Li and Lee Sedol. Not Gu Li and Ke Jie.

I assume Alpha Go has been trained using a large number of games played by Go masters so the games it played itself probably won't provide much more information, unless it can learn its own weaknesses from those games.

They said it was trained with the games of 5 dan and stronger players from KGS, plus self playing. No professional games and no training with the games of a specific player, because there so few of them (60-70 per year).

Although this is not the main reason, there is also a matter of intellectual property of Go game records. The official position of Korea Baduk Association is that Go game records are copyrightable.

On the other hand, it is not recognized as such in Korea, let alone in other countries. There has been multiple attempts to change the Korean copyright law to explicitly list Go game records as an example of work that can be copyrighted. All such attempts failed so far.

The official position of Korea Baduk Association is that Go game records are copyrightable.

Who owns the copyright?

Presumably, it'd be a joint copyright. When two people write a book together, both hold the copyright.

I'm guessing that AlphaGo has been doing more learning. Studying Sedol's games is probably a good platform but certainly not a plan to win - Jie needs to go in expecting the unexpected. Just look at how different the games were between Hui and Sedol.

To answer your question more directly: giving AlphaGo more hardware would only give it more CPU cycles to work with (faster decisions, not necessarily good ones). The true improvement will likely come from the very same source that humans get it: practice and mastery.

Doesn't faster hardware allow it to navigate more of the search tree, in effect giving it more time to find better decisions?

Yes, the online hardware used during the game, vs. the offline hardware used for training the neural network, is mainly there to run a ton of monte carlo tree search (MCTS) iterations. MCTS tends to produce better estimates as you throw more compute power at it. In perfect-information, deterministic games, it theoretically converges on optimal play given infinite compute power, although in games with very large state spaces in practice that guarantee isn't too useful on its own, and the heuristic guidance provided by the NN is crucial.

In the paper they state that they tried giving it more power than the version that played Sedol, but found that it had rapidly decreasing marginal value, so they decided on a level that was pretty close to optimal.

Reply on news.ycombinator.com