Lies Horrible Lies

[Subotai]

http://www.usgo.org/news/2015/11/google ... racked-go/

I refuse to believe that go has already been solved.

With the strongest programs playing at around a 5d amateur level, to have jumped from that to the top is ridiculous.

Thoughts?

[DrStraw]

It was misquoted. What he said was that he had cracked his go board when he threw it against the wall after getting frustrated at its intractability.

[Krama]

It was misquoted. What he said was that he had cracked his go board when he threw it against the wall after getting frustrated at its intractability.

This!

However I did watch the game with DCNN against fuego was it?

The Neural Network played really beautiful moves that you could probably see a dan level player playing however since it didn't have any way to actually calculate moves it failed in easy tsumego problems.

If they somehow fix the poor tactical calculations and put more power in NN I think a high amateur dan - low pro level would be achievable.

But computers will still be far from Gu Li/Lee Sedol at their prime.

[Uberdude]

Also note that "solved" has a specific meaning for games. Even chess, where the best engines are now far better than the best humans on home hardware, is far from solved. Draughts/checkers is probably the most complex famous game to be solved, but that is only weakly solved (the main line of optimal play from both sides is known, but if the opponent plays a suboptimal move the bot may then draw rather than win) not strongly solved (best move known for all possible positions). So I would be very extremely strongly sceptical that Go has been strongly solved, not only because of it's absurdly huge game tree, but because the approach of Deepmind with their neural nets is exactly the opposite of the crunching the game tree you need for strong solving. As for making a better-than-human Go engine, just highly sceptical, but I would like to be surprised .

[CnP]

I'll remain sceptical until they actually do something. btw I think super-Go playing AI's would have a positive influence on Go.

[Uberdude]

Deepmind have done something, namely published a paper about a neural net being about as good as an amateur mid-high dan at predicting the next move of a pro game (55%), though the actual playing strength of that approach was less clear or impressive. Something neural nets trained on pro games tend to suffer from is not knowing how to play against bad moves, for example there was one game I saw where the neural net played a crane's nest tesuji to capture some key stones and then when the opponent tried to escape, which should have been futile as any 15k knows, it let it escape as presumably pros never try such a futile attempt so it hadn't been trained how to respond to that pattern.

[Krama]

Deepmind have done something, namely published a paper about a neural net being about as good as an amateur mid-high dan at predicting the next move of a pro game (55%), though the actual playing strength of that approach was less clear or impressive. Something neural nets trained on pro games tend to suffer from is not knowing how to play against bad moves, for example there was one game I saw where the neural net played a crane's nest tesuji to capture some key stones and then when the opponent tried to escape, which should have been futile as any 15k knows, it let it escape as presumably pros never try such a futile attempt so it hadn't been trained how to respond to that pattern.

so much for "deep" mind