DeepMind Produces a General-Purpose Game-Playing System, Capable of Mastering Games Like Chess and Go Without Human Help (ieee.org) 124
DeepMind has created a system that can quickly master any game in the class that includes chess, Go, and Shogi, and do so without human guidance. "The system, called AlphaZero, began its life last year by beating a DeepMind system that had been specialized just for Go," reports IEEE Spectrum. "That earlier system had itself made history by beating one of the world's best Go players, but it needed human help to get through a months-long course of improvement. AlphaZero trained itself -- in just 3 days." From the report: The research, published today in the journal Science, was performed by a team led by DeepMind's David Silver. The paper was accompanied by a commentary by Murray Campbell, an AI researcher at the IBM Thomas J. Watson Research Center in Yorktown Heights, N.Y. AlphaZero can crack any game that provides all the information that's relevant to decision-making; the new generation of games to which Campbell alludes do not. Poker furnishes a good example of such games of "imperfect" information: Players can hold their cards close to their chests. Other examples include many multiplayer games, such as StarCraft II, Dota, and Minecraft. But they may not pose a worthy challenge for long.
DeepMind developed the self-training method, called deep reinforcement learning, specifically to attack Go. Today's announcement that they've generalized it to other games means they were able to find tricks to preserve its playing strength after giving up certain advantages peculiar to playing Go. The biggest such advantage was the symmetry of the Go board, which allowed the specialized machine to calculate more possibilities by treating many of them as mirror images. The researchers have so far unleashed their creation only on Go, chess and Shogi, a Japanese form of chess. Go and Shogi are astronomically complex, and that's why both games long resisted the "brute-force" algorithms that the IBM team used against Kasparov two decades ago.
DeepMind developed the self-training method, called deep reinforcement learning, specifically to attack Go. Today's announcement that they've generalized it to other games means they were able to find tricks to preserve its playing strength after giving up certain advantages peculiar to playing Go. The biggest such advantage was the symmetry of the Go board, which allowed the specialized machine to calculate more possibilities by treating many of them as mirror images. The researchers have so far unleashed their creation only on Go, chess and Shogi, a Japanese form of chess. Go and Shogi are astronomically complex, and that's why both games long resisted the "brute-force" algorithms that the IBM team used against Kasparov two decades ago.
Re: (Score:3)
Re:How much longer? (Score:5, Funny)
What's next; Parcheesi? Tiddlywinks? Backgammon?
Global Thermonuclear War
Re: (Score:2)
what side do you want?
Re: (Score:1)
Global Thermonuclear War
what side do you want?
Whatever side (human, AI or otherwise) that's smart enough to play the winning move.
Re: (Score:2)
"Whatever side (human, AI or otherwise) that's smart enough to play the winning move. "
Define "winning move".
I'll prefer the one that plays the move that makes me survive.
Re: (Score:3)
Re: (Score:2)
BUELLER!!
Oh, sorry, wrong movie.
Re: (Score:2)
Global Thermonuclear War
...or CalvinBall as Randall Munroe puts it in this XKCD:
Game AIs [xkcd.com]
Re: (Score:2)
Re: (Score:2)
"It's really only shocking it took so long."
No, it isn't. These kind of plays have a staggering number of combinations, so brute attack approaches are absurdly costly.
And then, while it looks "easy" to optimize a program for a single game, this one works against all of them -that's quite high grade stuff.
Re: (Score:2)
As mentioned in TFS Go and Chess are complete information games, and that's what AlphaZero is good at. What AI is struggling with for now is incomplete information games, like Poker.
And every job that I know off that isn't already being replaced by robots are VERY incomplete information games. If you want to convince yourself, just read a real life instruction manual or specification document, or even worse, ask a customer what he wants.
Re: How much longer? (Score:2)
Look up Libratus. Computers don't struggle with Poker any longer.
Re: (Score:2)
Re: (Score:2)
Uh huh. When is this going to happen? Computers are good at playing games. We get it. Apparently the "AI researchers" aren't able to figure out any other applications.
Re: (Score:2)
1. Falken's Maze
2. Black jack
3. Gin rummy
4. Hearts
5. Bridge
6. Checkers
7. Chess
8. Poker
9. Fighter Combat
10. Guerrilla Engagement
11. Desert Warfare
12. Air-To-Ground Actions
13. Theaterwide Tactical Warfare
14. Theaterwide Biotoxic And Chemical Warfare
15. Global Thermonuclear War
Re: (Score:3)
If it can win at Secret Hitler [secrethitler.com] against humans, I'll start worrying. Think about how that can apply to social network bots, public comments, and graph search.
Hilariously, that game tries to reinforce their point that if you're not liberal, you're fascist but instead shows that fascists are always the people claiming to be liberals.
It's an eye-opener, but not for the reasons the game makers think it is.
Re: (Score:1)
Liberal means favoring individual liberty. Liberalism is antithetical to authoritarianism. It's not the same as left- vs. right-wing, or progressive vs. reactionary. Like most authoritarians, fascists oppose liberalism [wikipedia.org], believing the individual must suppress his own "selfish and materialistic" desires in favor of the "superior morality" of his leaders.
While it is true that the world is not simply divided into fascists and liberals (there are many forms of authoritarianism besides fascism), for the sake of a
Re:How much longer? (Score:4, Insightful)
If liberalism is a negative it is because the term is being used incorrectly, like favoring the "rights" of corporations and capital over those of persons, or preserving the "liberty" of one individual to deny the liberty of another.
Then perhaps those who used to call themselves liberal (like myself) should distance themselves from a term that now means in favour of safe-spaces, authoritarianism, affirmative-action, etc.
For example, I now am very careful to distance myself from any sort of toxic group, even if I think they "hijacked" the word for their own uses. The term now means "authoritarianism", whether one likes it or not. If someone doesn't want to be seen as expressing support for authoritarianism then perhaps they should distance themselves from the word "liberal".
Re: (Score:1)
Can it do trade negotiations with China?
Re: (Score:1)
Not your point, I know, but there's still an active worldwide chess community, and the best players make a quite comfortable living. Even a cell phone can easily beat them, but that hasn't mattered.
Re: (Score:1)
Re: (Score:1)
Re: Correct me, but there is no 'self-learning' (Score:1)
Demi's Hassabis openly recognizes that DeepMind's stuff has major limitations in comparison to the human brain (or any animal really). He's done this in interviews and otherwise. AlphaZero however does just what I said. Reinforcement Learning is a branch of machine learning with somewhat understandable and reproducible results. DeepMind's currently best at it. I agree that RL has limitations, and I don't use it. I use primarily convolutional neural networks, since it's readily applicable to lots of imagery
Re: (Score:2)
Re: (Score:2)
Nice to see your stuck in the 60's with your thinking and knowledge. Get back to me when you actually can contribute to conversations instead of being so negative and limited all the time.
Re: (Score:2)
Neural Networks have literally been around since the 1960s and have very limited uses.
Well, so do you, so where's the problem? ;)
Re: (Score:2, Informative)
Re: (Score:2)
You AI nutters thinks neural networks are some sort of magic. It isn't
Well, it is "magic" in the sense that it works, as in this case where their machine learning algorithm taught itself to be the world's greatest chess/go/shogi player with nothing more than the rules of the game coded in to it, but no-one has any idea how or why it works.
Whether that is a good or bad thing is yet to be decided...
Re: (Score:1)
[...] but no-one has any idea how or why it works.
That is severely incorrect. Artificial Neural Networks are very well understood and their properties are mathematically described for the common cases. So are the methods for training them.
The concept of ANN:s being mysterious and not understood is just a myth spread by media because it sells better, just like AI sounds more advanced or scary than algorithmic inference or whatever would be a better term in each case.
Re: (Score:3)
What I mean is once the network is trained the "thought processes" that the network uses to come up with an answer are not understood.
This seems to be especially true of image recognition networks, but as they don't talk about AlphaZeros' reasoning in the open access paper I'm inclined to think it's also true of their network.
If you can link to a paper or post from an AI researcher that details how these kinds of networks are actually coming up with their answers I would be very interested to read it.
Re: (Score:1)
Re: (Score:2)
You AI nutters thinks neural networks are some sort of magic. It isn't.
Agreed! ANN:s is just an algorithmic tool.
NN have been around for many decades, and it is a dead end.
No, it is a very useful tool for function approximation in some cases.
However, in its current shape, it is most likely a dead end on the road towards AGI.
A NN is nothing like the human brain.
The human brain *is* a neural network. An *artificial* neural network, ANN, on the other hand, has to my knowledge not been implemented even close to similar to a human brain.
Just mentioning it since you enjoy ranting on colloquial abuse of terms on the subject instead of focusing on the contents.
Re: (Score:2)
From all the articles about this I've read, all they did was program the rules of the games (valid moves, win conditions etc.) and then let their machine learning algorithm teach itself actual play technique by playing games with itself.
As it was never shown example games or taught existing techniques I think it's fair to say it taught itself how to play.
Seems to me if it can learn Chess and Go (Score:2)
Most military systems are more complex and costly due to the human element and the protection of life. Removing humans and maybe one Abrams tank can be out fought by 100 trucks with auto guns/launchers? Just wondering?.
AI wise! If it can be done! It will be done! By someone!
Just my 2 cents
Re: (Score:3)
Re: (Score:2)
Re:Seems to me if it can learn Chess and Go (Score:5, Insightful)
Maybe we can take it a step farther - not fight the war at all, just simulate the fighting using computers. Then, depending on the enemy’s simulated tactics, we can calculate which of our citizens need to report to the disintegration chambers.
Re: (Score:3)
An Xmas sale... maybe. (Score:2)
Let me ask (Score:2, Informative)
Re: (Score:3)
Re:Let me ask (Score:4, Interesting)
Almost no systems in the real world are deterministic. That's why stochastic approaches to AI (develops a statistical model based on multiple repetitions - e.g. fuzzy logic, machine learning) have been much more successful in real world tasks.
Please unleash it on the ... (Score:2)
Re: (Score:2)
more games than you think (Score:5, Informative)
Thank you for spamming the entire thread with your imperceptive and unenlightened comments.
There's nothing odd about the choice of chess and Go whatsoever. Humanity has thousands of years of experience with these games. We know they aren't trivial, and we know they're not so complex that we can't understand progress, when we see it.
Additionally, the large literature of expert games was a useful hand-rail between hand-crafted and fully autonomous.
Quite apart from the neural network portion, Monte Carlo tree search (MCTS) is a fundamental algorithm in computer science, and this work demonstrates that MCTS is ready for prime time, having defeated from scratch exceptionally strong chess programs that have been painstaking hand-tuned over five decades and hundreds of man years. MCTS exists within the large and growing theory of multi-armed bandit problems. These are fundamental problems in many important industries (such as drug discovery, to name just one).
Multi-armed bandit [wikipedia.org]
Recurrent self-learning is another important algorithm in computer science and machine learning.
And finally, the neural network portion is far closer to the human brain than the vast majority of algorithms used in computing. Without any human instruction, these neural networks are learning to detect patterns of almost arbitrary complexity (so long as they seem to help in winning games).
I was reading Galileo in the original last night (English translation, but his original prose). He knew about Kepler, but wasn't sold on elliptical motion. Then he carefully observes four previously unknown moons of Jupiter and correctly determines that they can't all be in circular orbits. The word he used (in English translation) was "oval". But he still didn't choose to accept Kepler's work (apparently, he felt that Kepler's ellipse and his oval were not the same thing).
Galileo was a giant in the history of science. But still a little wooden headed on a few points, nonetheless.
I think Odd Buster Spamalot is nuts to criticise Galileo for not being Newton. Only because Galileo sorted enough of the fundamentals out in the first place (about the proper concerns and methods of science), was it even possible for Newton to become Newton (and he knew it, himself, and he's famous for having said so).
The computers we now apply to neural networks are roughly a factor of one billion times more powerful than the computers of the 1960s (thirty doublings over 45 years gets you there at the traditional pace of Moore's law).
You could complain that neural networks are only good at this one thing, but actually no: they are now state of the art in image classification (IC), speaker-independent large-vocabulary continuous speech recognition (CSR), and machine translation (MT), as well. All of these endeavours also date back to the 1960s, and have thousands of man-years of deep research behind them. Then DNNs come along, finally on a sufficiently powerful computer, with a few small tweaks to the algorithms, and simply cleans up the state of the art with nothing more than a small team of graduate students doing a quick project within the scope of their degree program to push this along (the subsequent move to industrial scale was immediate and brisk). Traditional MT research programs would have hundreds of professional researchers, slaving away for decades, at least, and never accomplished as much.
We're all of ten years away now from the day where no competent doctor ever reads an x-ray (or other radiological image) without computer assistance (definitely including a powerful NN component).
Watson was a bit idiotic, right from the beginning. The problem was Jeopardy, itself, which was always rather facile in the nature of the questions asked, and fundamentally more a test of ridiculously wide and shallow
Re: (Score:2)
Thank you for spamming the entire thread with your imperceptive and unenlightened comments.
There's nothing odd about the choice of chess and Go whatsoever. Humanity has thousands of years of experience with these games. We know they aren't trivial, and we know they're not so complex that we can't understand progress, when we see it.
Impressively long comment, but it doesn't change the fact that what computers are good at is games - things with well defined, and complete sets of rules.
Yes, you can gamify a lot of things (e.g. factories, to some extent), and very profitably, but you can't gamify all of existence. It still isn't general AI, which is the point.
Re: (Score:2)
I'm still going to count going from "given the rulebook" to "world's best player" in 9 hours as impressive
Re: (Score:2)
Slashdot has become a very sad place. This used to be the kind of tech-frontier topic that might make for interesting conversation.
Now, I can predict pretty accurately which people will honk out their same, sad, banal complaints for any given headline. I don't know what they get out of this - whether they actually feel that strongly about the subject, whether they like arguing or attention, or whether it's just a kind of reflexive contrarianism.
What's clear is that none of the more frequent posters have a
Big Yawn (Score:2)
Yeah, they built a huge database of moves and then they read it back while playing. That's exactly how humans play these games, isn't it?
For bonus points, they embody that database in a format that they can't interrogate in any useful way outside of actually playing the games.
But can it play... (Score:1)
But can it play Tic Tac Toe?
Where's the source? (Score:2)
ntr
Obligatory: Elon Musk warning of AI. (Listen!) (Score:2)
https://www.youtube.com/watch?... [youtube.com]
Re: (Score:3)
Would anyone with QM and human brain (Score:1)
Re: (Score:2)
And if you can compress "multiple human lifetimes" worth of practice into a few hours and get actual usable results from it that's rather interesting isn't it?