Hacker News new | past | comments | ask | show | jobs | submit login
Exploring the mysteries of Go with AlphaGo and China’s top players (deepmind.com)
130 points by beefman on April 10, 2017 | hide | past | favorite | 60 comments



The players who played against AlphaGo and lost said they got something valuable out of it. They started looking at completely new possible combination of moves they hadn't thought viable before. (The guy that played against AlphaGo before Lee Sedol said how it improved him. He even climbed several ranks.)

This makes me wonder. Maybe the human brain never computers for global optimas. It gets to a point where it's good enough and then stops. Clearly, if they play vs AI and then start improving their strategy again, the brain might have thought defeating most human is a good enough strategy.

This is also see in other games, someone who plays in "easy" difficulty or low ranked games in multi player games are stuck with "bad habits". As in they do really well in those thresholds but perform poorly when introduced to higher difficulty. But as soon as there is a need, good players rapidly adjust. (Example: In MOBA games like leauge of legend or Dota, higher ranked players adapt at a much higher rate to new emergent strategies while lower level player tend to not manage this and find things unbalanced.)

My point is, as we discover more of these `slighly better than human AI`, we'll also discover that human brain might be capable of much more. It just never saw the need to pour more resources in it till it found a healthy competition. Our only (soon to be) sentient friend, the AI.

--------------------------------------

[0] https://www.youtube.com/watch?v=0X-NdPtFKq0


Maybe the human brain never computers for global optimas. It gets to a point where it's good enough and then stops.

That sounds a lot like Herbert Simon's idea of "satisficing"[1].

Satisficing is a decision-making strategy or cognitive heuristic that entails searching through the available alternatives until an acceptability threshold is met. <snip> Simon used satisficing to explain the behavior of decision makers under circumstances in which an optimal solution cannot be determined. He maintained that many natural problems are characterized by computational intractability or a lack of information, both of which preclude the use of mathematical optimization procedures. Consequently, he observed in his Nobel Prize speech that "decision makers can satisfice either by finding optimum solutions for a simplified world, or by finding satisfactory solutions for a more realistic world. Neither approach, in general, dominates the other, and both have continued to co-exist in the world of management science.

[1]: https://en.wikipedia.org/wiki/Satisficing


Indeed, Ken Regan has attempted to measure it in chess.[1] The techniques used for AlphaGo may soon make go another model problem we can use to precisely study human cognition.

[1] https://rjlipton.wordpress.com/2015/10/06/depth-of-satisfici...


In a very crude sense, we learn by observing patterns that work, and without new working patterns we can tey to come up with novel patterns. But this is much harder, especially if the new patterns require a completely different approach instead of being an incremental change.

I fully expect human ability to drastically increase in several areas when computers start finding new working patterns.


So then I wonder if with enough play, people can beat AlphaGo once more. Maybe would should let people have a chance to keep playing against AlphaGo till they recover from the flaws it finds in human games, and maybe human creativity can win out once more.


> So then I wonder if with enough play, people can beat AlphaGo once more. Maybe would should let people have a chance to keep playing against AlphaGo till they recover from the flaws it finds in human games, and maybe human creativity can win out once more.

Alpha go learns just like humans it is not standing still for humans to catch up.


What if the size of Go board was increased? Does AlphaGo scale as well as expert human players?


Probably better, actually. (But that's just a guess.)


Humans are limited by the size of their skull. Computer systems have no such limitation.


No. People will learn from a stronger player, but not far beyond your own ability. AlphaGo is far enough ahead of human play, the human era in Go is over.

However, the era of everyone being able to have a super-human Go teacher should certainly make things more fun and raise the level of amateur players.


Case in point: I constantly use Stockfish (Open Source Chess Engine) + Scidvspc (Open Source Chess tool) to analyze my own games. I can have Stockfish play and evaluate all of the games I play and give me a rating on each of my moves.

Stockfish figures out any blunders I do automatically. I just set it to "search depth 20" (which normally takes ~20 seconds per turn... although some complicated positions can be ~1+ minute per turn) and have it analyze my games. In less than 10 minutes, I have a very strong analysis over the course of an entire game.


What an interesting thought!!

It led me to another: you know how AI can imitate painting styles? Well, perhaps an AI could also figure out a style, by itself (i.e. given constraints, but no style to follow.) i.e. it figures out a new one! But which has a novel effect, on humans and as judged by a computer. Once we have been shown the style, humans could improve by starting to use that! We might like the effect, be able to judge that it is good, but just never saw that way before.

Imagine if AI had invented pointillism, impressionism, cubism, etc! Imagine if it figures out the next "global optimum" from putting splotches of paint down, showing human artists a new way to see. Wow.


When humans are the judge, how are you going to test trillions of combinations per minute?


I don't know. I am assuming there is some other way to judge "what something looks like!" - something that an AI could learn, be programmed with, derive, or invent. My assumption could be false and perhaps there is no way to approach aesthetics with anything other than a human brain, or training on the outputs of human brains.

But if the AI had brain-like features, couldn't it be its own judge? Couldn't it have as an emergent property, its own sense of aesthetics or meaning, which it could then approach using its own naive approaches. Which humans could say "Wow" to.

AI as the ultimate naive painter. What a concept!

I realize it presupposes a lot that may never come to pass.


A/B testing.


The system notices anthropomorphization of the system. The list of allowances is updated. Recalculating optimal behavioral responses. Done. Estimated probability of winning increased by 0.12.

Thank you, you are welcome. We'll get along wonderfully!


Don't anthropomorphize non-human systems. They hate it when you do that.


Wow! I knew the Ke Jie vs AlphaGo would come one day, and probably the human team vs AlphaGo, but didn't expect the "Pair Go" (human + AlphaGo vs another human + AlphaGo). Interesting to see.


Centaur chess has been a thing since Deep Blue. I think the problem their future of ai stuff is having is most orgs have no idea which of their zillion problems actually needs this tech. Hopefully this will raise more awareness.


In Centaur chess, the human gets to consult the AI. In Pair Go, there is only alternating play, with no other form of communication, leaving the human guessing at the AI's plans.


I thought people stopped doing Centaur chess when it turned out that human+computer usually loses to computer alone.

There are still positions that people evaluate correctly but computers misjudge, but those are pretty rare.


Pretty is nice, tell more please.



It's not as far as people might think though.

If I recall correctly, Nakamura used an older version of a weaker chess engine (Rybka). He also tried to force a relatively equal position (since he was behind in points anyway) which led him to lose because of risky play.

I wouldn't be surprised at all if Stockfish would lose to Stockfish + Nakamura or any of the leading chess players. These people have been using chess engines for analyses for quite a while now and certainly know their limitations.


The key idea of human+computer is that human does strategy and position evaluation, and computer calculates tactics and acts as a safety net to prevent human blunders. A weaker chess engine would still perform that function.

Granted it's all conjecture since no large competitions took place, but the mere fact that no such competitions took place suggests that the outcome probably won't be pleasant for humans.


My point was tat with the same chess engine, man with machine beats machine.

I do see your point about position evaluation though. And yes, it would be interesting to see a tournament.


I actually have not seen any recent cases when a human with chess engine was beating the same chess engine more often than a random chance would imply.

It used to be the case before about 2012-2014, but now the gap is so large that a human can only make things worse.


That's because Ke Jie was already beaten 3-0 by AlphaGo [1]. The new set-up will definitely be more interesting to watch.

[1] - https://qz.com/877721/the-ai-master-bested-the-worlds-top-go...


There is another NN engine from Tencent making progress for last couple of weeks.

https://en.wikipedia.org/wiki/Fine_Art_(software)


Fine Art's played a lot the on Fox server. It actually hadn't been seen since March 11 (at which time it was still losing sometimes), but is playing again today...


Has deepmind done any additional training of AlphaGo since last spring?

The training that AlphaGo underwent between fall of 2015 and the match against Lee Seedol in March 2016 seemed quite effective.


Yes. From the press conference:

https://www.reddit.com/r/baduk/comments/64hmu2/watch_live_pr...

> Last year, the AlphaGo that played against Lee Sedol was v18. The Master that achieved 60-0 was v25. Which version will play this time?

> Scott Beaumont: The exact version of AlphaGo will have to be introduced by the experts working on the project, but it will be the newest version. Defeating Ke Jie is no easy task, and AlphaGo has been continually improving. Also, we hope to continually improve Go as an art, through the wisdom of the players and AI in cooperation.

(Magist/Master was the handle under which AlphaGo played 60 online blitz games last December: https://en.wikipedia.org/wiki/AlphaGo#Unofficial_online_matc... )


Very cool, thanks for the details. I didn't realize the version playing as "master" was 7 versions after than what Lee Seedol faced!


AlphaGo played a series of online games under the name 'Master' after the Lee Sedol match, so presumably deepmind are still working on it. Id say it's pretty safe to assume that's going to involve a lot of training.


Can AlphaGo run behind the Great Firewall?

I think that for the match with Lee Sedol it was offloading computation to (some form of) Google Cloud no?


This is not a semi clandestine match. The collaboration with the Chinese Government stated in the post should take care of that, a reliable connection to the AlphaGo servers at home.


Probably

Also (if this is optimized) the information exchange back and forth should be minimum (board position and next move selected by AlphaGo takes very little space)


Hotels and conference centers often get out of China through Hong Kong, thus bypassing the Great Firewall and letting their guest use Facebook or Twitter.


Is AlphaGo written in the Golang to play Go?


No.


How did you figure out?


This is exciting. Kasparov was prescient when he talked about and promoted advanced chess [1]. The concept was older but he was very humble on accepting this possibility. Magnus Carlsen training is with computers [2]. I would love to see Carlsen playing against the top computer chess program now. He is the "New Hope".

BTW, it seems we are experiencing here in HN some kind of AI fundamentalism since the downvote rush doesn't correlate with this normal comment. I am not arguing against the downvotes, just that sych reactionary behavior was not expected.

[1] https://en.m.wikipedia.org/wiki/Advanced_Chess

[2] http://m.dw.com/en/world-chess-champion-magnus-carlsen-the-c...


> I would love to see Carlsen playing against the top computer chess program now. He is the "New Hope".

I don't understand; he would get destroyed. Do you mean just to see what it would look like?


The answer is: we don't know yet.


I'm not downvoting you but this is against every evidence based on ELO ratings and all the history of human-human, computer-computer and human-computer games: Carlsen is going to lose against my phone and probably every future human will.

Let's wait until the end of May but maybe we're already at that point with go as well.

(We don't know yet if I can win a running race against a Ferrari, still I'm pretty sure about the outcome.)


I guess for go it depends more if they bothered to train AlphaGo sufficiently before the match. If it has been training at the same level as before it is probably well beyond human reach.


We do know, with essentially 100% certainty. Carlson's elo rating is mid 2800s and the top engines are high 3300s. A 500 pt rating difference implies he has less than a 5% chance of winning a single game, let alone a match.

Plus, he can play against an engine any time he wants. If he could beat them in some magically surprising way, would it make any sense for him to be keeping this amazing skill a secret from the world? No.


Those ratings are uncomparable, since humans only play humans and computers only play computers. It's probably way lower than 5%.


> the top engines are high 3300s.

Sources?


Seriously?

https://www.google.com.au/search?q=chess+engine+elo+rating

And we don't know "with 100% certainty", but we can calculate a projected win/loss probability using those ELO ratings[1]

A 3300 rated chess engine vs Carlsen at 2850 gives a 93% win probability to the chess computer.

That's roughly the same probability of an International Master raned player (ELO 2300) beating Carlsen.

The top ranked engine is Stockfish, rated 3500+

[1] http://www.bobnewell.net/nucleus/bnewell.php?itemid=279


ELO scores can only be compared across a single pool of players, and those scores you linked to are from computer v computer tournaments only. The absolute scores aren't comparable because you can effectively adjust the absolute score by adjusting the parameters you start the system off with.

As far as Im aware no engines compete against humans regularly enough to get an ELO score (presumably because watching a computer thrash human players for a few hundred games isnt that interesting).


Sure, the rating compression thing is a valid criticism.

There's little doubt they are much stronger than humans though. There are enough online games that if this wasn't so it would show up.


You haven't read the comments in your links, they say tuat this ELO was not calibrated with chess players which means it is not an ELO.


Arpad Elo's rating system has no single defined reference pool of players. The way it works is that you take a pool of players, establish the rating that unknown new players will receive (1500 in some cases) and then the system evolves from there. There are many disparate pools of players using the Elo system for their ratings, across many games.

Just because it was not calibrated with humans does not mean it is not an "Elo" system.


Carlsen himself certainly realizes he would have no chance in a match against the top computer engine. Its effective rating is probably at least 400 Elo above his.


We definitely know. Engines are taking handicaps when playing humans in modern tournaments.

https://en.m.wikipedia.org/wiki/Human–computer_chess_matches


Sorry but, I don't think Carlsen can even beat a decent chess engine on a smartphone nowadays. Humans beating computers is an impossibility for many years now.

He may only win if computer starts with a serious handicap.


Carlsen would still beat Stockfish on an iPad. Stockfish 8 on a server farm is 3294 to Carlsen's 2838. But on a circa 2014 1GHz iPad, it's about 2500.

http://support.stockfishchess.org/discussions/questions/865-...

http://www.inwoba.de/


This is a hilarious anti-computer game, Rybka-Nakamura. It's 270 moves but it plays like Tetris or PacMan.

http://www.chessgames.com/perl/chessgame?gid=1497429


Unless you believe Carlsen is 100x better than Kasparov at chess, I don't think he would have any chance.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: