DeepMind's Go-Playing AI Doesn't Need Human Help To Beat Us Anymore (theverge.com) 133
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.
ObSimpsons... (Score:5, Funny)
I, for one, welcome our new Go-playing robotic overlords.
Re: (Score:3)
Re: (Score:2)
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.
Like? Evolution is a form of reinforcement "learning". And for the nature vs nurture crowd, so is the passing down of knowledge and skills. Even in situations that aren't actually physically repeated we make a mental simulation of what we think might happen and play out scenarios, then it's corrected by reality as things do or don't work out as we thought. Of course a Go AI has it easy in that the rules define the outcome so it can just play itself a zillion times without external input. If you were trying
Re: (Score:2)
Not all task cab be reinforcement learning tasks but all task can be broken down into reinforcement learning subtasks. Your thoughts and deeds are not the singular actions you think they are, they are a composite of many inputs, feeding into many thought recurrent though processes, circulating through many decision trees, and that continuously composite of results, leading to actions to be taken. So each sub task can be learnt, the composition of sub tasks can be learnt, once each sub task has been learnt,
Re: (Score:1)
Re: (Score:2)
Who says it couldn't? You'd need to provide it with robotic actuators, a camera, lots of butter and bread, and a score function describing how sandwich-buttery is any given situation. But it's only fair that you provide a definition of the problem to solve, since a human would need it too.
Under these conditions, it's quite possible that the algorithm would learn how to make sandwiches. They've achieved it with playing video-games, after all.
so it got dumber? (Score:1)
Re: (Score:2)
i think they are saying that perhaps the "most advanced version" of alpha go is better than the version that beat Lee...
what i want to know is what happens when it plays with itself... also how does it feel when it wins.
Re: (Score:3)
It feels nothing.
It has no algorithms to 'feel' anything.
Re:so it got dumber? (Score:5, Funny)
Re: (Score:2)
I want to know if they used a pre-versus-AG version for the later plays, or if it learned from playing AG--because AG learned from humans, and playing against it transfers human knowledge to Zero.
Re: (Score:2)
Re: (Score:1)
The 100 win rate games were played against AlphaGo Lee, the version that played Lee Sedol in 2016.
The 90% win rate games were played against AlphaGo Master, the version that played against Ke Jie in 2017.
Re: (Score:3)
Re: (Score:3)
This graphic [googleapis.com] 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.
Re: (Score:2)
No. It happened to have an initial win-streak of 100 games to zero. That doesn't guarantee a win is going to happen every single time.
There's no information given on how long it took to play the first 100 games. It might have been hours, or a few days, or a week. Continuing to play against the original software for 40 days, it eventually trended towards an overall winning percentage of 90%. We don't know how many games were played across those 40 days, it could have been thousands. Also perhaps the o
Re: (Score:1)
No. It happened to have an initial win-streak of 100 games to zero. That doesn't guarantee a win is going to happen every single time. There's no information given on how long it took to play the first 100 games. It might have been hours, or a few days, or a week. Continuing to play against the original software for 40 days, it eventually trended towards an overall winning percentage of 90%. We don't know how many games were played across those 40 days, it could have been thousands. Also perhaps the original software was improving its own play at the same time?
Why speculate? The paper says there is AlphaGo Lee, which AlphaGo Zero defeated after three days of training and AlphaGo Master, which took forty days of training to beat 90% of the time.
Re: (Score:3)
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.
Re: (Score:3)
Re:This is cool, but I'll be more interested when. (Score:5, Informative)
... It can deal with hidden information.
So you mean something like poker? AI beats pros at Texas-Hold'em [scientificamerican.com].
Re: (Score:2)
Poker is a very bad application of 'AI'. Poker is not a game of human strength, but rather a game of exploiting human weakness. A poker playing computer can be superior to a human simply by virtue of its programmers choosing not to program it to have any tells when it bluffs. A computer raising when it has a four of a kind is indistinguishable from a computer raising when it has high card 10.
Re:This is cool, but I'll be more interested when. (Score:4, Insightful)
Re: (Score:1)
"programmers choosing not to program it to have any tells "
Actually, creating and exploiting fake tells is exactly how an AI poker can win.
1. A simple computer has 4 of a kind and always goes All-In.
2. An advanced computer has 4 of a kind and sometimes calls and sometimes goes all-in, just to mix it up.
3. A good AI computer has 4 of a kind and has carefully built a fake trail of typically acting a certain way with a good hand, but this time decides to simply bet 18%, which is calculated based on what it has
Re: (Score:1)
I'm the parent AC.
I know about the poker AI, but the thing is that solution is very non-general.
AlphaGo's inner workings are based on the MCTS algorithm (which is a fairly general algorithm that can be applied to many games). It would be more interesting to see MCTS properly developed to deal with hidden information. There are improvements (ISMCTS, MO-ISMCTS and MT-ISMCTS) but they are all extremely limited in one way or another.
A general solution would be awesome.
Re: (Score:2)
A general solution would be awesome.
General solutions require Strong AI [wikipedia.org], which, for now, is science fiction.
Complaining that this self-learning Go program isn't general purpose is sort of like complaining that a better electric car battery won't help your Tesla reach Warp 9.
This is incremental progress, not a revolution, but it is still an interesting advance.
Re: (Score:2)
Re: (Score:2)
Such as Global Warming[TM]?
Oh you're denialist too!
Kind of obvious in hindsignt given the nature of your sig.
I can already see the next generation of Illiberalas taunting skeptics to "win a game of Go against this AI before questioning Climate Change".
Welp, if you've run out of things to be outraged about, just make some up and get outraged about those! Darn librulhs.
Re: (Score:3)
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.
Re: (Score:2)
you still don't get that whole "evidence" thing, do you?
as in, tons of evidence of one thing over here , no evidence over there, yet somehow you want to pretend the two possibilities are equally scientifically probable, or even that the one with no evidence is more likely to be true regardless of the complete lack of evidence supporting it.
Re: (Score:2)
Re: (Score:2)
That's the skill in writing AI for computer games. You don't just want to have the game always at skill level infinity, you want to have the whole range from skill level 0 to infinity. Skill level 0 is simply making moves at random. Sometimes, it's up to looking ahead a dozen moves, other times, it's not constantly making aggressive moves.
Re: (Score:2)
Re: (Score:2)
There is minimal benefit in being the absolute best Go player that could exist.
This is a research project. People want to know how far they can push it.
Dumbing it down to provide a useful challenge for humans is easy.
Re: (Score:2)
Re: (Score:2)
Humans are a ridiculously powerful AI's, much more powerful than AlphaGo, so why can it beat us? It isn't because AlphaGo never sleeps, in fact, AlphaGo does sleep it just spreads it among more frequent and smaller time scales. We can't beat it for the same reason we can beat AlphaGo at everything but playing go.
STILL not a real AI (Score:1)
Re:STILL not a real AI (Score:4, Funny)
That's just what an AI would say, Mr. Schumann
Re:STILL not a real AI (Score:5, Insightful)
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.
Re: (Score:2)
Well the TotallyFake non-AI are coming for your jobs regardless of how impressed you are.
Re: (Score:2)
More fake news bullshit machines ain't takin' nobody's JERBS STFU faggot
He posted angrily from the cab of a self-driving semi-truck.
Guys, imagine how ironic this is if the shills have some automation script to shitpost any article covering automation.
Re: (Score:2)
Drug Design and Climate models (Score:2)
Sorry, I doubt there is any chance that a neural network can be used in a meaningfull way in drug design or climate modeling.
Re: (Score:2)
Sorry, I doubt there is any chance that a neural network can be used in a meaningfull way
FTFY
I don't know, sounds like a great way to cheat at Go.
Re: (Score:3)
Sorry, I disagree.
Re: (Score:2, Insightful)
We're all neural networks designing drugs and climate models.
Re: (Score:2, Insightful)
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 demonst
Re: (Score:2)
Neural network in the sense of AI is in fact at it's core an implementation of a mathematical replication of at least part of how our brains work.
"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."
We can't even do that
Re: (Score:2)
Though it's true that we're neural networks, we aren't the same neural networks as these are.
Indeed. The machine's version is better, as demonstrated by its superior Go playing abilities.
Re: (Score:2)
Better at playing Go, but a neural network probably never will play Chess. ... nothing else.
And: AlphaGo only can do Go
Re: (Score:1)
I seriously doubt that a world class english major with a doctorate in english literature and accolades in their field can be used in a meaningful way in drug design or climate modeling.
But... a very smart person trained for 10 years in those fields might be able to contribute.
However, a.i. probably trained is already contributing in a meaningful way to drug design (even coming up with new drugs humans hadn't considered) for ... well.. several years now.
And while I don't personally know about climate modeli
Re: (Score:2)
AI != neural network
A neural network is a very specific kind of datastructure and algorithms.
It can do amazing things, but bottom line that will always be a kind of 'pattern matching'.
I see no way how something like this can be used to model chemical reactions or enzyme interactions aka modeling drugs.
And modeling climate, sorry: no way at all.
Re: (Score:2)
You mean a system that is pretty much defined as a series of interlocking simple equations? One of the few areas where the outputs of humans look like neural networks already?
Re: (Score:2)
A climate model takes an input state and computes an output state.
That output state can then be used as input for the the next step.
While that looks similar on the first glance like a NN, the computations behind are completely different.
Algorithms in a NN are 'just running the NN' in other words: they don't really change depending on the topic where you want to use the NN in.
A climate model is run by its equations, aka the algorithms change all the time when we know more or add another 'factor'.
Re: (Score:2)
I really don't see any difference, other than one is equations derived by humans, the other by a guided random walk.
And feeding the output of one state into the next is a very common NN technique.
Re: (Score:2)
I really don't see any difference, other than one is equations derived by humans, the other by a guided random walk.
Then you perhaps should read up what an NN/ANN is and how it works: https://en.wikipedia.org/wiki/... [wikipedia.org]
Sorry, perhaps I simply lack imagination. Hoever I see no way how that could be used for climate models.
An ANN is not modeling anything, it is recognizung things. Two things as far apart on the spectrum (of algorithms) as one can imagine.
Re: (Score:2)
I'm talking about the endpoint. The actual climate model isn't "modeling", its a bunch of math. A bunch of math that happens to be designed to model climate. And the NN is a bunch of math designed to pattern match climate. I don't follow the distinction you're positing.
And, when I said "I don't see any difference", I was making an implicit statement that I don't think you defined a difference. I'd be interested in seeing it, but I don't follow whatever you're thinking.
Re: (Score:2)
Sorry, you have to read up what an ANN is and how it works.
I lack englipsh skills to properly explain it in a forum (or would need two or three days of work, that time I don't have right now)
Obviously everything a computer does is a bunsh of equations, and applying them, solving them ... that does not make everything equal.
Re: (Score:2)
I think our miscommunication is not on ANNs, but on climate models [wikipedia.org]. In fact, like SETI@Home, you can donate cycles to trying to machine develop models [climateprediction.net].
I understand the methods are different, but I'm not sure the results are distinguishable.
Re: (Score:2)
Hm, i was not aware than an ANN can approximate arbitrary functions/formulas.
That mean, you could indeed use them in a climate model. But for what purpose? It would be much slower than using the real formulas.
https://becominghuman.ai/neura... [becominghuman.ai]
imagine the possibilities ... (Score:3)
Re: (Score:2)
Re: (Score:2)
No. For neural networks you need simple floating point calculations. For bitcoin you need to do SHA-256 hashing.
These are totally different tasks, each running on their own optimized hardware.
Re: (Score:2)
let's play global thermonuclear war! (Score:2)
What side do you want?
Re: (Score:2)
Doesn't matter. The only winning move is not to play against AlphaGo.
Slashdot AI commentary summary... (Score:5, Insightful)
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.
Re: (Score:2)
> Also, cotton gins and blacksmiths, therefore computers will never take our jobs.
They don't need to take all our jobs. Humans needs don't increase beyond shelter, security, food and water. The 'need' of entertainment is a bottomless pit in which you can throw money and out comes stories, sports, children and art, to name a few. Humans are definitely bright, but we are also routine and we like it when things at work don't change that much, so we can keep up.
When the majority of us like average, easy to
Re:Slashdot AI commentary summary... (Score:4, Interesting)
Never forget that the many of the people who did lose their jobs to industrialization died homeless of exposure and starvation after being put down by the military. They were provided neither training nor jobs on the new machines (which was their true issue).
Similarly, we could have a very rough 20 years where jobs are destroyed faster than they can be created and where workers over 50 (40?) can't afford train for the new jobs and there are more unemployed than society is willing to pay for (even tho our productivity is 100x what it was a hundred years ago so one worker should be able to completely support 99 unemployed with 200 square feet of living space and basic food).
It's coming. It could be better but it's probably going to be bad. Possibly even "great depression" or "financial panic of 18xx" bad. (they had a lot of financial panics in the 1800s that were pretty terrible.).
Re: Slashdot AI commentary summary... (Score:1)
Re: (Score:2)
This leads to a quick saturation of highly educated individuals still without positions. So they go into the only jobs on offer in the service sector, creating massive under-unemployment.
They (we?) will join the service industry, selling our time for $$$, in whatever way we can to pay for food. It's already underway with the massive number of vlogging, prostitution etc. The price of unskilled human labour will drop like a stone, welfare costs will balloon, and the ruling elite will eventually legalise things that currently make do as fiction - The Purge, Running Man etc.
Time to update that zombie-proof castle to human-proof?
Re: (Score:1)
Digital computers CAN'T think. Digital computers are nothing like how the brain works. You seem to be talking about yourself. You don't think at all. If computers playing games is impressive to you, then you must be easily impressed. Computers excel at games because they have strict rules. Computers LOVE rules.
Re: (Score:2)
Uh... dude... I was making fun of you. Well, you and Gweihir (sic?), who I assume is taking a day off to scream at pigeons.
Can it be generalized? (Score:2)
Re: (Score:3)
an exciting first step (Score:2)
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 inte
Re: (Score:3)
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
Re: (Score:2)
great reply slew;
"Secondly, it is unclear if a "sense" of what types of moves are best is being learned"
I understand that it can be hard for a programmer to accept a program that doesn't lend it's 'thinking' to analysis. But I believe that until computers can sense things, they are not intelligent.
There is a corollary in the classroom. The children have a math problem. Most of them repeat the steps they have been taught and come up with the correct answer. But one or two look at the problem with a fuzzy log
They need it to handle arbitrary size boards. (Score:1)
Boards may be 21x21 but may be other sizes as well.
This doesn't show we're winning at AI (Score:3)
Re: (Score:2)
That argument is meaningless because it always works:
1903: Heavier-than-air powered flight is more tractable than we previously thought.
1957: Putting objects in orbit is more tractable than we previously thought.
1969: Landing on the moon is more tractable than we previously thought.
2016: The game of Go is more tractable than we previously thought.
20xx: Human-level cognition is more tractable than we previously thought.
Yes, a lot of people are over-optimistic about AI at any given time, but it is moving for
Skynet birth (Score:1)
There is a catch. (Score:2)
I have programmed this type of learning algorithm in the past. About 30 years ago when computers were about 30000 times slower than now.
Anyway, you can have the program play itself for a while and it becomes quite good. But you won't know how it will perform against a human unless you try. It might be very good against those moves that the computer player will come up with, but very bad against moves thought up by a human.
30 years ago I tackled a simpler game than go.
Re: (Score:2)
30 years ago I tackled a simpler game than go.
Global Thermonuclear War? :)
Re: (Score:2)
Read the book The Promethean by Owen Stanley whereby a tech billionaire decided to build the perfect AI as a gift to humanity. Social Justice Warrior intervene and taught him the concept of social justice and the AI forked itself to infinity.
Re: (Score:2)
The way to pass Turing test would be to convert it into creationism as only human would believe it?
Oh wait, this should be easy, it was created...