Life In 19x19
http://www.lifein19x19.com/

The ceiling of human Go has been raised
http://www.lifein19x19.com/viewtopic.php?f=10&t=12799		 Page 1 of 1
Author:  jeromie [ Thu Mar 10, 2016 9:38 am ]
Post subject:  The ceiling of human Go has been raised
Watching the Lee Sedol / AlphaGo games has been a lot of fun (kudos to the AGA for providing some great commentary!), though like Kirby I was a bit disappointed to see AlphaGo convincingly win the first game. One of the things that helped me move beyond my initial reaction is thinking about how this might elevate the play of top humans.

One of the limiting factors for improvement at the highest level is finding an opponent that can consistently push you to improve and/or introduce new ideas into your game. The game two commentary by Andrew Jackson and Myungwan Kim highlighted how AlphaGo is already doing exactly that. Several moves made the professional players react with surprise, but as they took a deeper look they couldn't find anything wrong with the computer's play. It's opening up the possibility of creative new lines of play that just haven't been considered and bringing back older strategies that have fallen out of favor. Current and future professionals will undoubtedly incorporate the lessons they learn from high level computer play into their own games. Imagine a tool like this in the hands of a player like Ke Jie!

This advancement in computer go doesn't just mark the end of human superiority at an ancient game. It also represents a new tool that humans can use to surpass the best efforts of past generations. I'm not sure how this will impact future styles of play or, more importantly, the culture surrounding the game, but I am sure that one of humanity's great strengths lies in its ability to adapt to changing environments. Humans may never again be stronger than computer players at playing go, but they will almost certainly be stronger than they have ever been before. It's an exciting time to be a part of the go playing community!
Author:  Pio2001 [ Thu Mar 10, 2016 10:43 am ]
Post subject:  Re: The ceiling of human Go has been raised
The games between Lee Sedol and AlphaGo are indeed exceptional. And seeing these strange moves used by AlphaGo can reveal limits of the current go theory.

But on the other hand, it doesn't necessarily mean that we may improve our understanding of the game. Humans are good at evaluation, while computers are good at enumeration.

I once wrote an article in the french review of Othello (another abstract game) about a move played by the computer playing an Othello game, that seemed to be the worst possible one according to the theory, but that actually worked.
After hours of analysis, we came to the conclusion that we were just in the 1 % of cases where the move works, and unfortunately, we couldn't find any property or principle that would explain why we were in these 1 %.
The only way to prove that the move worked was to enumerate patiently a dozen of variations from the resulting position, and, at the end of all of them, conclude that "this situation looks correct".
Repeat for the next variation, etc.
At the end, all the variations having been proven correct, we could conclude that the computer was right in playing this move : it worked.

Despite the excitement of seeing these new incredible moves, we must keep in mind that, after all, they might just be computer moves with no explanation behind them. When a given kind of move is bad in 99 % of cases, the computer can easily see if the current position belongs to the 99 bad ones or to the 1 good one. Evaluating 100 different follow-up is nothing for the machine.

If this is true (thourough analyses of these new moves shall tell) the only new thing would be that for the first time, a machine would have crossed the boundary where brute force analysis becomes efficient on a 19x19 board.
Author:  Kirby [ Thu Mar 10, 2016 10:53 am ]
Post subject:  Re: The ceiling of human Go has been raised
Pio2001 wrote:
But on the other hand, it doesn't necessarily mean that we may improve our understanding of the game. Humans are good at evaluation, while computers are good at enumeration.


Traditionally, this might have been true.

But isn't its "value network" one of the big distinguishing features of AlphaGo? As I understand, one of the neural networks they've trained is for the purpose of estimating who is leading at a given board position, rather than searching to the end of the game.

So while we could before say that computers are good at enumeration, but at least we have evaluation... Now, perhaps, we've seen that computers can surpass us at evaluation as well.
Author:  RobertJasiek [ Thu Mar 10, 2016 11:26 am ]
Post subject:  Re: The ceiling of human Go has been raised
None of Alphago's unorthodox moves is new.
Author:  Calvin Clark [ Thu Mar 10, 2016 12:01 pm ]
Post subject:  Re: The ceiling of human Go has been raised
Someday soon we may see an article like this written about go: http://www.businessinsider.com/anand-on ... ss-2013-11

Due to the nature of go, I suspect go players are less vulnerable to what in chess are called "novelties," although once in a while a prepared move shows up that yields a minor advantage.

Go endgames are really boring and tedious to analyze. While computers may assist greatly in doing this faster, it remains to be seen whether it will improve human play, other than making players more aware that they are dropping points in the endgame.

Maybe some new invasions will be unveiled by the computers and will be adopted as standard patterns.
Author:  Fedya [ Thu Mar 10, 2016 2:23 pm ]
Post subject:  Re: The ceiling of human Go has been raised
I play a lot of unorthodox moves, and nobody praises me for it. /sarcasm
Author:  jeromie [ Thu Mar 10, 2016 2:38 pm ]
Post subject:  Re: The ceiling of human Go has been raised
There are certainly lines of play that a computer might choose that will be unworkable for humans, but that doesn't mean people can't learn from any of the moves. Otherwise I couldn't learn anything from professional play. :)

Robert may be right that AlphaGo's moves aren't strictly new. I think that encourages the notion that humans can learn from its play. They are either rare or out of favor, though, so they will at least encourage professionals to explore lines they might not have considered. (And there were 1 or 2 moves that, while not strictly new, seemed to genuinely surprise some of the commenting pros with their timing.)
Author:  Mef [ Thu Mar 10, 2016 11:38 pm ]
Post subject:  Re: The ceiling of human Go has been raised
Pio2001 wrote:
I once wrote an article in the french review of Othello (another abstract game) about a move played by the computer playing an Othello game, that seemed to be the worst possible one according to the theory, but that actually worked.



Please correct me if I'm wrong, but Othello as a game is basically capped a 60 moves long? And I think computers can read 30+ moves deep at times? Because the combination of those two things is a bit frightening in an opponent (e.g. perfect play from middlegame onward....).
Author:  Pio2001 [ Fri Mar 11, 2016 4:47 am ]
Post subject:  Re: The ceiling of human Go has been raised
Mef wrote:
Please correct me if I'm wrong, but Othello as a game is basically capped a 60 moves long? And I think computers can read 30+ moves deep at times? Because the combination of those two things is a bit frightening in an opponent (e.g. perfect play from middlegame onward....).


Yes. The search depth was about 20 moves, and the endgame was perfect. Computers are way stronger than humans at Othello. The game on the smaller 6x6 board is even solved.
The computer was about my level with a search depth of 2 moves. To review my games, the right setting was 12 moves. I could see my mistakes and the computer was pointing the right moves.
With a search depth of 16 or 20 moves, the choices of the computer were becoming strange and I could no longer understand their meaning.

If the level of AlphaGo is just above the top pros, we should learn something by studying its games. If it rises too far above, it will be more and more difficult to learn something from it.
Page 1 of 1 All times are UTC - 8 hours [ DST ]
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group
http://www.phpbb.com/