Slashdot Mirror
Google's AlphaGo AI Defeats the World's Best Human Go Player (engadget.com)
It isn't looking good for humanity. Google's AI AlphaGo on Tuesday defeated Ke Jie, the world's number one Go player, in the first game of a three-part match. The new win comes a year after AlphaGo beat Korean legend Lee Se-dol 4-1 in one of the most potent demonstrations of the power of AI to date. Adding insult to the injury, AlphaGo scored the victory over humanity's best candidate in China, the place where the abstract and intuitive board game was born. Engadget adds: After the match, Google's DeepMind CEO Demis Hassabis explained that this was how AlphaGo was programmed: to maximise its winning chances, rather than the winning margin. This latest iteration of the AI player, nicknamed Master, apparently uses 10 times less computational power than its predecessor that beat Lee Sedol, working from a single PC connected to Google's cloud server. [...] The AI player picked up a 10-15 point lead early on, which limited the possibilities for Jie to respond. Jie was occasionally winning during the flow of the match, but AlphaGo would soon reclaim the lead, ensuring that his human opponent had limited options to win as the game progressed.
Playing games is not AI. A game has strict rules. These are easy problems for computers to solve. Computers love strict rules. It isn't intelligence. And don't say "well you cannot do a depth first traversal of same Go states becuase it is so huuuuge". That doesn't make any difference: just use a different algorithm. It still isn't AI.
Mathematically, which is harder to solve for, Go or Chess? Is this some sort of diversity thing that they've started using Go over Chess?
Adding insult to the injury, AlphaGo scored the victory over humanity's best candidate in China,
There is no insult to losing in China. The appropriate response is, "Thank you for allowing me to win."
The real question is: when two identically trained systems compete against each other, what are the underlying mechanisms of competition leading to one winning?
AlphaGo at its core is an MCTS
For such a thing, one needs (I think) to do some unexpected moves to constantly force machine into sparsely probed regions.
And, during discovery stage, one needs doing it "off-line" to avoid google's retraining. Thankfully, space is big enough to ensure that google can be forced quickly enough into deep woods.
For a match like this - one needs to use different precalculated prologs for all games (won or lost).
It's more like hacking than playing...
Same for the human player ...
It isn't looking good for humanity.
Purpose built machines have been able to, or be used to out do humans for a very long time. A lever can be used to lift more weight than a person alone can. But we're not being ruled by sticks. Cranes can lift even more.
Cars are used to move people further and faster than they could on their own. Computers can do many more calculations per second. These things make life better for humanity as a whole.
Unless AlphaGo figures out a way to keep a person from unplugging it, I'm guessing that humanity will be just fine.
Is AlphaGo programmed in Go?
I think you miss the point. Due to the complexity of Go in the sense that any turn can be played on dozens if not hundreds of spaces, computers could not brute force their way to victory. The reason this is important is because A: it shows a computer using something other than brute force to solve a logistical problem, and B: the program has the ability to be self taught beyond learning the basic rules (and rule sets don't get much more basic than Go). Yes, a computer beat a human, but this is a much different victory than winning at chess.
Why doesn't Google actually apply this to solve some REAL world problems, huh?
Like figuring out how to end war, share resources and be peaceful? Or whom to vote for as president?
That's right - Google is bringing useless things to the table.
My problem with these things is that they are purpose built by a team of engineers, and updated massively in between matches. It's valid to say "AI did X", but given how the AI is redesigned, tweaked, and special cased for each new opponent, it's much more accurate to say "a massive well funded team of specialists beat this one old guy at his life's work", because the AI that is developed will never be made available in the manner that an AI or other fixed product would be. If the top guy is allowed to play against the AI a thousand times, will the AI still win?
In other words: is it the AI doing the work, or is it all the tweaks and redesigns that happen between matches, done by the combination of game specialists and engineers, that have rigged up some kinda trick that holds out for a few matches?
Not only that; heuristics are horribly complicated in Go. For example, it is possible to score moves in Chess simply by assigning value to pieces and evaluatiing the current state of the board. On Go a seemingly innocuous early in the game can be decisive in determining a match later on.
Someone call Boston Dynamics.
Fuck.
I'm a good cook. I'm a fantastic eater. - Steven Brust
I'll be impressed when a computer wins Hungry Hippos, that game is obviously rigged towards anyone younger than 7. My kid beats me in it all the time.
Sent from my TARDIS
Thank you for that explanation, which is helpful and has some points I hadn't considered. But I do still wonder if whoever said "meh" isn't partly right too. Is this really an AI kind of approach? Different than how computers win at chess, yes, but is it really AI? The whole reason that chess was an AI problem for computers to solve was that early on people thought that to beat humans you'd have to learn how to think. That ended up being wrong. Computers playing chess against humans now is unfair because the computers are basically taking an open book test against humans who have to memorize and think and can't consult an open book. I guess it's impressive for Google in that they aren't doing brute force but are they still doing lookups? Is trying to maximize winning chances really nothing more than limiting which "book" the program consults?
It might be important to note that the rules has been adapted to fit a computer player. "Win" condition in classic Go is not as straight forward as the modern computer friendly version.
Nah, it's basically the same. Chess was solved, not by brute force (because, just like Go, the branching factor is too big), but by brute-force combined with some pruning algorithms.
Go is, as you mention, even bigger. So it was solved by brute force, much more powerful computers, and improved pruning algorithms. That's basically it.
"First they came for the slanderers and i said nothing."
Definitely, this is kind of a Big Deal(tm) --- If you watch last week's episode of VICE (or was it the week before...?) on HBO, they delve a bit into the state of AI technology and actually mention a Go match with AlphaGo and a Chinese master player, I don't remember if it's the same guy though. In that episode (edit: ahh there it is, it was two weeks ago; "Engineering Immortality & Robot Revolution" talks robotics + AI with Hammilton Morris, my favorite VICE guy) they explain a little bit why defeating these Go champions w/ AI is a bit more impressive than it seems initially. If you have HBO or HBO GO (lol), you can watch the episode here ---> http://www.hbo.com/vice/episod... -- or just torrent it like a normal human being.
640k ought to be enough for anyone.
Alpha go is trained by reinforcement learning, like a person would be. They let it watch some historical games until it gets the basics, then it plays itself to refine its game.
It's debatable whether chess is a simpler game or not, but chess can be effectively played with standard look ahead and tree pruning techniques. Those work poorly in go. The reinforcement learning used for alpha go could be used to teach it to be an unbeatable chess player too. And originated with deep mind for teaching the computer to play Nintendo games. Any Nintendo game.
One of the neat things about reinforcement learning is that you define the outcome you want (highest score, winning the most games) and what inputs are allowed (placing a stone, pressing buttons on a controller) and that's it.
But I do still wonder if whoever said "meh" isn't partly right too. Is this really an AI kind of approach?
It's a clear example of "weak" AI, not "strong" AI. It's a clever solution to the problem with a solution inspired by human intelligence, but the machine is not learning in the general sense.
"First they came for the slanderers and i said nothing."
Go is, as you mention, even bigger. So it was solved by brute force, much more powerful computers, and improved pruning algorithms.
No, that's not at all how Alpha Go works.
Yes, yes it is. It's primarily a tree searching algorithm. On top of that, they use a heuristic (the neural network, which doesn't learn while it plays) to figure out which branches to search down. Then it also uses a monte carlo algorithm to prune the tree to a manageable level.
"First they came for the slanderers and i said nothing."
Humans probably can't beat computers in chess even with all the open books in the world though, assuming the "books" aren't actually computer programs.
...Go have no strict rules,
Go most certainly does have strict rules. You can only play one stone in a turn, for example; you can't play another stone until your opponent plays; and you're not allowed to just pick up all the stones that you don't like and throw them into the trash.
yes there are some basic rules describing player moves however just applying them doesn't play a good game.
Sure. In chess, too: it's easy to learn how to move the pieces. Moving the pieces to win a game is hard. Saying that "following the rules doesn't play a good game" is not a subset of saying "Go doesn't have strict rules."
http://www.geoffreylandis.com
Wow, I didn't think anybody still remembered "Game of Life".
http://www.geoffreylandis.com
I just wonder what would happen if the next tournament play with a board that is double the x,y size
A 38x38 positions board
Would Alpha Go maintain the lead ?
Maybe we adapt faster ?
Any GO expert care to explain why this is feasible or silly ?
What rule would you change? Each player gets a handgun?
It was already lost in the 1950s, when the original researchers tried to make an end-run around perception, which required a density of compute Not Available Soon.
You can't not deliver on your founding conceit for six decades and then expect no social erosion of your Tokamak grandeur. Of course, they made lots of progress, but hardly any of this happened under their original Tokamak brand.
Those who still care about the One True Meaning usually trot out the term AGI (artificial general intelligence).
The next rung up the ladder from AGI is MHHAGI (muhaha artificial general intelligence). At this point, it's clear to everyone that we're finally and truly talking about penises as endowed by God himself, so King Arthur's quest for the master term usually ends here.
Any GO expert care to explain why this is feasible or silly ?
It is silly. If you train a neural net to differentiate a photo of a dog from a photo of a cat, it can learn to do that. But it is then silly to expect it to recognize a picture of, say, a horse. That is NOT what it was trained to do.
Likewise, Alpha-Go was specifically trained to play on a 19x19 board. Any other size, such as 18x18, would not even be recognized as valid input.
On the other hand, if you trained it on variable sized boards, then it could adapt to that.
Here is an actual example: Deepmind trained a NN to play a wide variety of video games. When it was introduced to a new video game, it could used its existing training to play and master the new game much faster than even the best humans.
Go is played on 9x9, 13x13, and 19x19 boards. On the smaller boards, tactics (joseki) is more important. On bigger boards, strategy (fuseki) is more important, and apparently innocuous early moves can have far reaching effects much later in the game. On a 38x38 board, strategy would likely be even more important, and winning the game would require a profoundly different style of play. My gut feeling is that an AI, trained by playing against itself, could master that new style much faster than a human.
I know Slashdot isn't what it used to be, but I think you're reading the wrong site.
It's primarily a tree searching algorithm. On top of that, they use a heuristic to figure out which branches to search down.
This is also what human players do. Since humans can't search as broadly, they prune more aggressively, but the basic algorithm is the same.
Yeah. The human pruning algorithm is actually really amazing in how quickly it works. Tree searching is slow and hard for humans, but oh well.
"First they came for the slanderers and i said nothing."
It seems to be about goalposts and definitions. One could have the same discussion with "Does peg legged Pete have an artificial leg?" Some would say "Yeah, sure. Artificial leg." Others would say" No way. It's just a piece of wood, driftwood even, that he uses to hobble around on. An artificial leg is something else" and they would keep saying that even after we'd have the "six-million dollar man" legs. Same with artificial life. Is it life, but not really, just a good approximation because it is "artificial". Or is it "life" that did not arrise through the ongoing natural evolution process. When talking A.I, we're not talking artificial humans or even "general artifical intelligence" (whatever that would be) but something that gives the impression that the thing on the other side (of the screen, the board, the table, the whatever) is intelligent. For the champion Go player, if he was to play against AlphaGo without knowing so, he would not be able to tell if he was playing a human or a machine. So, AlphaGo, in that incarnation ( it can learned to a totally different skillset) is real A.I for Go in a sort of successfull but limited Turing test. Same for Big Blue for an even more limited chess Turin test. Now that we know Big Blue can beat world champions and AlphaGo can beat human Go champions, we kinda say, meh, yeah, we know it's better than humans but it's not real A.I. Things will progress piecemeal in this fashion untill we have natural conversations with our digital assistants (on phones? Tablets? Robots?) knowing that they are not real humans but acting as if they were. And still people will say "Yeah, but that's not real A.I"
Doesn't the AI cheat to win in games such as CIV?
I don't think you're correct at all with the every video game in existence beating a human player.
Wow, sent an e-mail as suggested when clicking on "use classic" banner, and got a fast response that addressed my msg
Funny. As a libertarian I get that all the time. I bring up reasonable objections to a post and get down-modded as a Troll.
So, down voting people with a different opinion is not simply the province of the "right-wing."
If you're scared of your govt then you need to further restrict its powers
Vote 3rd Party in 2016 and beyond
It's clear that, for national security reasons, this technology should be trained and deployed to assist with foreign relations. Particularly, since it should theoretically be a master of game theory, it should be trained on a set of prior foreign relations incidents. In order to deal with North Korea and other rogue nations, it must be taught brinksmanship. In order for this to be effective and to prevent the enemy from calling our bluff, it must be given direct control of our nuclear arsenal. The only question left is whether to call it 'Joshua' or 'Skynet'. /s
Corruption is convincing someone that the selfless ideal is the same as their selfish ideal.
All I have been able to glean so far is that the rewritten version uses around 10% of the computing power (both to train its neural networks, and during actual play) to achieve much improved play compared with the original AlphaGo used to beat Lee Sedol. Thus far, although promised, the architecture behind the rewritten version is unpublished. Later this week, some insight is going to be provided.
The old version was based on a combination of techniques (primarily multiple neural networks, combined with Monte Carlo techniques). The interesting thing about the way it operated was that it could tell you which move was likely best, but could not explain why. The same is actually true of human Go players. While locally best moves can be identified, the human selects the globally best move based to a large extent on feel. The game is too complex (both for humans and AIs) to use calculation on a board wide basis. Both the old and new AlphaGo systems appear to demonstrate characteristics we would refer to as "intuition" and "creativity" if seen in humans. How similar is it to human instinct and creativity? We really do not know.
I am extremely interested in learning how the rewritten system works. I think the twinning matches (between two teams each with one human expert and AlphaGo collaborating with each other) will also be extremely significant. The short to medium term promise of AI involves humans and AIs working together. As the AIs become increasingly complex, and the manner in which it comes up with recommendations ever harder to comprehend, this is a critical challenge to be addressed.
It depends how exactly it was trained. Most likely it was trained on a fixed-size board, so the architecture of the neural net wouldn't match a bigger board without modification. That's not necessarily true though, there are neural net designs that can take flexible sized input. Alpha Go could certainly be trained to do it.
The question is interesting. Given a human and computer that had never played on a larger board, which would do better? Go is a game that is fairly local (unlike chess) so it doesn't change radically if you make the board bigger, which means a lot of your strategy can be transferred... for both the computer and the human.
It's worth keeping in mind that although Alpha Go has played more games (against itself) than any human has, it has played for less clock time than any human champion has.
No one gets banned. ... easy.
However the modders might get banned from the modding system.
You inly need to complain to an admin
Cost free eBook I read (by iBook/Kobo/Amazon/ObookO/Gutenberg etc.): "The Green Odyssey" by Philip Jose Farmer.
How were they simplified? You'd think the rules about dead groups would be easy enough to computerize.
"When you have eliminated the unacceptable, whatever is left, however improbable, must be the truthiness" - Holmes
John Brunner made me curious as he is one of my favorite writers.
The story however is from Fritz Leiber.
Oops!!
Shows I should never trust my memory, and should always look things up even if I think I know them. You're right, the John Brunner chess story was the novel "Squares of the City."
I stand corrected. Fritz Leiber it is.
http://www.geoffreylandis.com