Slashdot Mirror
DeepMind's Go-Playing AI Doesn't Need Human Help To Beat Us Anymore (theverge.com)
An anonymous reader quotes a report from The Verge: Google's AI subsidiary DeepMind has unveiled the latest version of its Go-playing software, AlphaGo Zero. The new program is a significantly better player than the version that beat the game's world champion earlier this year, but, more importantly, it's also entirely self-taught. DeepMind says this means the company is one step closer to creating general purpose algorithms that can intelligently tackle some of the hardest problems in science, from designing new drugs to more accurately modeling the effects of climate change. The original AlphaGo demonstrated superhuman Go-playing ability, but needed the expertise of human players to get there. Namely, it used a dataset of more than 100,000 Go games as a starting point for its own knowledge. AlphaGo Zero, by comparison, has only been programmed with the basic rules of Go. Everything else it learned from scratch. As described in a paper published in Nature today, Zero developed its Go skills by competing against itself. It started with random moves on the board, but every time it won, Zero updated its own system, and played itself again. And again. Millions of times over. After three days of self-play, Zero was strong enough to defeat the version of itself that beat 18-time world champion Lee Se-dol, winning handily -- 100 games to nil. After 40 days, it had a 90 percent win rate against the most advanced version of the original AlphaGo software. DeepMind says this makes it arguably the strongest Go player in history.
I, for one, welcome our new Go-playing robotic overlords.
General Relativity: Space-time tells matter where to go; Matter tells space-time what shape to be.
It feels nothing.
It has no algorithms to 'feel' anything.
Cost free eBook I read (by iBook/Kobo/Amazon/ObookO/Gutenberg etc.): "The Green Odyssey" by Philip Jose Farmer.
... It can deal with hidden information.
So you mean something like poker? AI beats pros at Texas-Hold'em.
That's just what an AI would say, Mr. Schumann
They could have mined tons of bitcoins instead with that computing power.
Come now, lets be honest, it enjoys playing with itself just like everyone else.
Sorry, I disagree.
That sentence isn't really clear. But the fact they are specifying "version of itself" suggests that they have a save state of the version that won those games but have not stopped improving AlphaGo AFTER it beat the world champion. It may well be able to beat that older version of AlphaGo 100% of the time but only be able to beat the latest and greatest AlphaGo 90% of the time. This is a fork, it might be so close as to be a major release revision number variation of the same software but it's had completely different nurture and experience and is a completely different mind than the first much like identical twins are different people.. creepy people yes, but distinct.
Not impressed, doesn't prove anything, and why should anyone even care?
Maybe because they're not trying to prove anything? Maybe their actual goal is to improve general purpose algorithms by an iterative approach? Like it says in the article. Which you read of course.
Though it's true that we're neural networks, we aren't the same neural networks as these are. Remember folks, "neural network" in the sense of AI is a marketing term, it does not in any way imply that it functions in a manner similar to how our brains work. Fact is, we have no idea how our brains work. We know what certain parts are responsible for, but no idea how they do it. If anybody claims to know, then please ask them to describe in detail how memory is encoded in our brains, and have them demonstrate by altering a memory in a predetermined way.
There is also something to be said for diversity in those games. Beating AlphaGo doesn't automatically translate into beating the people AlphaGo's strategy defeated.
Ouch, I misread it. I got dumber instead :).
If it were all about tells there would be no online poker. Poker IS about reading other players but you can read a player from their play.
I've decided that this accomplishment -- a dizzying milestone in artificial intelligence that not long ago was though impossible or at least decades away -- is actually meaningless and doesn't prove anything and they should clearly have been working on some other problem. I have no idea how their system works, but I'm confident that their approach is just "brute force" (or something, I clearly have no idea what even that means) and won't generalize to any "real" problem solving (with my definition of "real problem" subject to change without notice).
I will only admit that any progress has been made towards artificial intelligence when computers perform exactly equivalent to humans in all tasks with no human intervention. I mean, I won't really, because I have weird quasi-spiritual hangups about believing computers can be intelligent, but that's where I'm putting the goal posts for now. Digital computers can't think, but I can because reasons. Free will or quantum mechanics or something else that I haven't thought about at all, probably.
Also, cotton gins and blacksmiths, therefore computers will never take our jobs. Amen.
Let's not stir that bag of worms...
Although to me this is clearly a case of A.I., parent has a point that not all types of problems lend themselves to be treated as reinforcement learning tasks, which seems to have been the key to success in this case.
Don't be too hard on him, manufacturing this sort of outrage is a time-honored tradition and puts bread on a lot of tables.
Outrage doesn't just grow on trees you know. Without his efforts the outrage deficit would lead to a world awash with harmony.
The answer is no, and provably so, because it is not Turing complete.
"First they came for the slanderers and i said nothing."
The essence of intelligence is that it enables one to predict the outcome of a unique situation based upon an understanding of its essential elements.
Starting with only the rules of Go, Zero explored a variety of combinations, learning that some were more likely to give a satisfactory result. It developed a sense of what types of moves are best. Thus, without playing or studying an infinite number of games it could know the type of move that should be best in each unique situation.
Theoretically, a vast intelligence, given only the facts of the Big Bang, could anticipate most of the resulting evolution of our universe. Zero has taken the first small step.
You are making quite a few assumptions. One, that somehow a "game" that has a goal (e.g., a "winner") is the same as predicted
an open ended problem. Two: that somehow AlphaGoZero developed a "sense" of what types of moves are best.
First, because of the limited rules and state space of the game Go, and the fact that there is a "winner", the Go universe is certainly closed and quite bounded.
In contrast, the real universe has a much larger state space and the rules are unknown (although some approximate rules are known by the current state of physics, they are known to be somewhat inconsistent) so although perhaps a theoretically vast intelligence, given our *estimate* of the state of the universe at the *assumed* Big Bang could apply our current approximate rules of the universe and potentially extrapolate the evolution of an *idealized* universe, since we know our current rules are inconsistent, this would likely be equivalent to garbage-in, garbage-out.
Similarly if AlphaGoZero wasn't aware of the full laws of physics of Go (say the 19x19 board size limitation, or perhaps was not taught the rule of Ko) and learned to play using simplified rules, it might not have learned anything essential about playing a "real" game of Go.
Secondly, it is unclear if a "sense" of what types of moves are best is being learned (it may simply be better at out-computing humans and building/"learning" its own huge dictionary of game outcomes). I think AI will only truly be useful (and trusted) when it also learns how to "teach". Anecdotally, I often find out how much I know about a subject when I try to explain it to someone else. Until, I do that, I might be able to be okay at muddling/faking my way through it, but I know that I don't really know it until I can successfully explain it to someone else.
FWIW, AI constructions like "GANs" (Generative Adversarial Networks), are a small step in this direction for certain things, but still have the potential loophole that it can "fake-it" as long as the discriminator can't be taught to detect it). Similarly, we can examine the moves that a program like AlphaGoZero makes and look at the probabilities that it assigns as part of it's MonteCarlo tree search before it moves, but that's like looking at an MRI of a Go Master and trying to learn what they are thinking when they made a certain move. By *humans* analysing the games that AlphaGoZero makes (and people have done this already) we can infer what it is trying to teach, but it will be much better when we figure out how to know for sure as integrated as part of its reinforcement learning.
What happened was, Google made a version of AlphaGo that beat Lee Se-dol. Call this AlphaGo One. It won, but it was at least close.
Then Google updated it and had it play lots and lots of top players, and it trounced them all. Call this AlphaGo Two.
Then they did this new version, AlphaGo Zero. Zero, early in training, beat AlphaGo One 40-0. Late in training it defeated AlphaGo Two 90% of the time.
This graphic paints a different picture. Three days to beat AlphaGo "Lee Sedol"-version, 21 days to beat AlphaGo "Master"-version that beat Ki Jie + 60 top pros, after 40 days AlphaGo "Zero@40" is now crushing the "Master" version by winning 90% of the time. I think that means the journalist got two things intermingled, "Zero@3" only surpassed the "Lee Sedol" version while "Zero@40" wins 100-0 against it. It's done both, but not at once.
Live today, because you never know what tomorrow brings
This doesn't show we are winning at creating AI. It simply shows that the game of Go is more tractable than we previously thought. Claims about the number of positions in Go being vastly greater than the number of atoms in the universe (something like the number of atoms squared) completely miss the point: this is a straw man argument for why algorithms weren't good at Go until recently, since (obviously) humans are not searching the entire space of all possible board positions either. It stands to reason that once a sufficiently flexible fuzzy hierarchical pattern matching algorithm were produced, it would be able to play Go much better than a human.