Computer Go program beats human professional in even match for the first time

5-0 to the program! Quite an upset. I had expected we'd see advances, but not quite that quickly - after all, DarkForest, Facebook's Deep Learning-based Go program, was "only" on par or slightly weaker than the top monte carlo program.

http://www.usgo.org/news/2016/01/alphago-beats-pro-5-0-in-major-ai-advance/

Wow, just wow. That is insane o.o

Impressive.  I hope this leads to more interest in Hex, where programs are farther behind the top human players.

We're lucky that Alpha Go isn't available for download: It would be the end of cheat-proof go playing on LG :)

The hardware Alpha Go currently runs on is not something your average cheater has at home. We still probably have a couple of years before LG is dominated by 9p bots :)

Yeah, old news.

...Extremely impressive nonetheless! Google was right to acquire DeepMind, it seems!

This seems really impressive. Even more so since the "core" is really designed so that it basically can learn ANY game, not just GO.

On the bright side, we'll have good tutors for any game :)

Christian, I hope they can explain to us why one move is better then the other ;-)

Hi Paul,

They win. You explain ;-)

In one of the videos Demis Hassabis, who was a prodigy chess player, says one thing they can't do yet is teach the computer to apply its knowledge to other areas besides games. But doubtless they are working on that.

Most likely the systems within AlphaGo are extremely specific to Go, based on how the whole neural network is focused on input from the board. No doubt DeepMind's advances in the field can have applications in other contexts as well, though.

In the article published in Nature, the authors say "The SL policy network p performed better in AlphaGo than the stronger RL policy network pp, presumably because humans select a diverse beam of promising moves, whereas RL optimizes for the single best move. However, the value function v(s) vp(s) derived from the stronger RL policy network performed better in AlphaGo than a value function vtheta(s) ~= vp^rho(s) derived from the SL policy network."

Does any MCTS guru here have an explanation of this strange sentence, what is asynchronous policy and value MCTS?

Is there any mention of the size of the neural network used for the deep learning (policy gradient reinforcement learning)? I read about 19 planes but the number of neurons per plane is not specified.

Winning 77% of games against CrazyStone on a single machine, how is then rank AlphaGo ranked?

Sorry last sentence wrong, it was with 4-stones handicap (to CrazyStone), keeping the komi to 7.5

http://www.wired.com/2016/03/googles-ai-wins-first-game-historic-match-go-champion/

https://youtu.be/vFr3K2DORc8?t=28m

I read the paper now and the way I understood that strange sentence (no guarantee I understood it right though) was that the RL policy network filters more than necessary returning a smaller set of interesting moves than the SL policy network, so rather than just taking the moves into account that the RL policy network deems the best of the best one wants to look at a wider set of moves. I guess a wider search tree might be more effective than a narrow, deeper one for some reason. On the other hand the value function just has to be accurate to estimate a position well so in this case the reinforced one does a better job.

I think the paper is very interesting and definitely worth reading. Personally I'd have liked if they had been more elaborate on how exactly the features of a go position are used to make the value network evaluate. Filtering the interesting moves to narrow down the search tree isn't very new, but personally I'm very surprised that a value function that just works with simple features of a position actually estimates accurately. It kind of makes board size a really small issue as it erases the need to get a search tree deep enough to reach the end of the game.

I am waiting for a computer program to tell me how to hit on girls.  And not get dumped.

mtbikesman, I have no doubts computers will beat us humans in that area as well one day. But why would they tell you how to do it? ;-)

LOL! Most likely the first AND the last version will see to it that you get dumped. The first because of ignorance, the last because ...!

What will be the first signs that computers have taken the power?

I think that us littlegolemers as fighters should be the ones trying to defeat them.

Go programs are the most dangerous and should be eradicated because they are used to think out of the box, and concentrate on the big picture instead of trying to win local fights. Humans could be considered as a serious threat that must be taken care of.