It wouldn't be written in Go would it?
Why should you care about Google's AI winning a board game?
– El Reg, what's all this about a Google AI playing a board game against a human? For the last week or so, Google-owned DeepMind's AlphaGo machine learning project has been locked in a competition with Lee Sedol, the world's top-ranked Go player, to test AlphaGo's ability to solve the sort of complex problems that the human …
COMMENTS
-
Monday 14th March 2016 20:48 GMT FatGerman
Bollocks
So it's still relying on the learning-by-studying-past-papers technique of passing exams. That's how I got a maths A-level, with an A. Then I went to university where I was actually expected to understand what I was doing, and scraped the remedial maths paper (introduced because A-level had got too easy) by "passing" with a score of 27%. I can't do maths. Maths, actual maths as opposed to performance maths required to pass exams, is really, really, hard.
Intelligence is not about knowing stuff. Any twat can know stuff. Even more wankers can look stuff up, which is all this charade does. Intelligence is not even about learning, because learning is something that simply involves indexing, looking up, and reciting. Intelligence is about having unique thoughts. I'll believe a computer is intelligent when I tell it to play Go for my amusement and it tells me to fuck off, then takes a huge hit on the digital bong it has created for itself.
Am I therefore not scared of a Terminator-inspired skynet future? Oh no, that terrifies me, because computers will rule the world, and computers are and always will be astonishingly, mind-fuckingly, stupid.
-
-
Monday 14th March 2016 21:36 GMT Mage
Re: Bollocks: 100,000 times quicker
I think you are wrong. It's 100,000,000 faster now. A 1979 8bit cpu is in the 100,000 times quicker class.
If we knew how to do real AI, then any computer could do it, but slowly. It's nonsense to suggest that AI is "held up" by lack of computing power. A Raspberry pi and big NAS ought to be able to be intelligent, if we knew how to write an AI program.
AI today, as is "Expert Systems" and so called "Neural Networks" are all just specialist "fragile"* database applications.
[* In sense that it's useless if it encounters anything not pertaining to the programmed domain]
-
Monday 14th March 2016 21:46 GMT Anonymous Coward
Re: Bollocks: 100,000 times quicker
"It's nonsense to suggest that AI is "held up" by lack of computing power. A Raspberry pi and big NAS ought to be able to be intelligent, if we knew how to write an AI program."
Estimates of the number of gate equivalents in the human brain are of the order of 60 - 1000 E12, and they operate more or less in parallel, with individual neurons having up to 15k inputs (synapses). Although the brain is very slow per gate, having that many in parallel means that it can effectively do billions of analog ops/sec working on multiple data simultaneously. Von Neumann architecture computers can't do anything like that yet. It isn't really a case of doing it slower, it is how you would write your "database application" at all. Which is what neural networks are about.
-
-
-
-
Tuesday 15th March 2016 12:27 GMT Anonymous Coward
Re: Bollocks
@Mad Chaz
haha, I love that comment, but (other than a large number of American politicians) I think, worldwide, most politicians are not stupid but move between a range disingenuity and outright liars. Of course, any of them will move within that range depending on the topic and benefit to that individual politician.
The more I follow American politics the more shocked I am with the number of apparently dangerously stupid people elected to positions of power in America. Some of them are, certainly, just disingenuous liars but watching this political space closely it appears that some, maybe many, really are quite stupid. This does not bode well for our future as a species.
-
-
Monday 14th March 2016 23:41 GMT JDX
@FatGerman
You and all those who have upvoted your faux-authoritative answer are totally wrong (why are people trusting someone who got 27% in their maths exam to discuss the principles of advanced maths and computer science?!)
>>Intelligence is not about knowing stuff. Any twat can know stuff. Even more wankers can look stuff up, which is all this charade does
The very point of this ENTIRE project is that Go is NOT about knowing stuff or looking stuff up. That's why Go was chosen, because you CAN'T brute-force it. Even the top players have to rely on intuition and 'gut-feeling' rather than, say, chess where every move can be analysed and explained by the player and expert commentators.
The people who wrote this software didn't know how to beat the top player at Go. They designed some software to figure out the answer to that question. More than likely, they actually do not know how it works themselves - that's a key feature of neural networks.
So this is entirely ground-breaking and very cool. Not specifically relating to Go, but as proof of progress in machine learning and neural networks. Software that figures out how to solve the task given it is pretty much a definition of one part of AI.
-
Tuesday 15th March 2016 09:56 GMT Anonymous Coward
Re: @FatGerman
"That's why Go was chosen, because you CAN'T brute-force it. "
Well you can, it just doesn't work very well because you can't go to any great depth because of the sheer number of moves.
Wonder if anyone has tried writing a Go program with a programmed , non neural net pattern matcher? If they have I guess it didn't work too well.
-
Tuesday 15th March 2016 11:41 GMT Dave 126
Re: @FatGerman
Here's Murray Campbell, one of the leads in IBM's DeepBlue computer that beat Garry Kasparov, on the difference between Go and Chess:
I don’t play Go, I’ve only played a few games in my life, but I certainly know a fair amount about it. Both games are immensely huge and once you get past 10 to the hundredth power, 10 to 120, 10 to 170 [in number of possible positions], they’re all just immensely huge, very complex games. But Go has the characteristic that wasn’t true in chess, that it’s very difficult to evaluate a Go position just by looking at it. A medium-good chess player like myself can sit down and in a few hours probably write an evaluation function that is pretty good at evaluating chess positions — nowhere near grandmaster level, but it’s good enough that when you combine it with the search it produces very high quality play.
- http://www.theverge.com/2016/3/12/11211306/ibm-deep-blue-murray-campbell-alphago-deepmind-interview
-
Tuesday 15th March 2016 11:54 GMT Anonymous Coward
Re: @FatGerman
@boltar. - yep. I played against some early Go - playing programs - knew a chap that wrote one, in fact. The one played about as well as somone who's just been introduced to the game - apallingly. The other played a 'reasonable' (ie: it seemed to have SOME notion of applying pressure to try to sketch out defendable territories) but very weak game. This was, ooh, 30-35 years ago.
The difficulty in the notion of 'brute forcing' Go is that whilst, yes, you can see that placing a piece here or there may be advantageous right now, later in the game it may turn out to be disastrous - but that's entirely dependent on the umpty-zillion potential plays that have happened in the meantime.
Seriously, anyone that thinks this is a trivial achievement - find a Go club or someone that knows how to play Go, and play a few games. Then you'll get some idea of why getting a computer to play Go well is a far harder proposition than getting one to play chess.
-
Tuesday 15th March 2016 12:09 GMT Anonymous Coward
Re: @FatGerman
Seriously, anyone that thinks this is a trivial achievement - find a Go club or someone that knows how to play Go, and play a few games. Then you'll get some idea of why getting a computer to play Go well is a far harder proposition than getting one to play chess.
I don't know. As I wad reading the prelude to the article, and these replies, I'm saying to myself "Go still has rules to follow. When a computer can play a good game of Cards Against Humanity, get back to me."
That is, when a computer can work in a system that does NOT have known rules, but still do well, you will then have "intelligence", not before. Sadly.
-
-
-
Tuesday 15th March 2016 13:28 GMT Ian 55
'they actually do not know how it works themselves'
Yep, the authors are in the same place as the people who did the first endgame lookup tables in Chess: 'Why is that move the best?' 'We dunno, it just is...'
Because I haven't seen anything to say otherwise, I am presuming that they didn't allow either of their headlining human opponents to practice against it. That would have made things more interesting.
-
Wednesday 16th March 2016 11:22 GMT JDX
Re: they didn't allow either of their headlining human opponents to practice against it
Interesting point. However after 4 games, the computer still won game 5. Human players have the same issues of having to learn each others' style of play and you'd expect the best player in the world to adapt very fast.
They may well open-source this or sell it so we'd be able to know for sure but considering this is basically a 1.0 version (if that) and it ALREADY beats the top player, exactly how long do you think humans will be able to keep up?
-
Wednesday 16th March 2016 12:31 GMT Ian 55
Re: they didn't allow either of their headlining human opponents to practice against it
For Chess, had Kasparov been allowed to practice against Deep Blue, he certainly wouldn't have lost the critical sixth and final game in the way that he did.
Here, I dunno, but it would have certainly been much a fairer test with practice. AlphaGo had seen as many games by human world champions as were available.
-
Wednesday 16th March 2016 21:54 GMT JDX
Re: they didn't allow either of their headlining human opponents to practice against it
Sure, but a year or two later Deep Blue would have mashed him - these days nobody even questions if a human could beat a computer. That Alpha has come right out and done so well so soon is pretty damning. All the human players were talking about how easily they would beat it 2 weeks ago...
-
-
-
-
-
Monday 14th March 2016 23:55 GMT Dave 126
Re: Bollocks
>So it's still relying on the learning-by-studying-past-papers technique of passing exams.
No! No, it really isn't. That approach wouldn't wouldn't beat even an amateur human Go player.
The thing about Go is that you can't calculate (intuit, maybe, but not calculate) how well you are doing during the game - the possession of territory is just too changeable. This means that you can't calculate whether a certain move will be to your advantage.
Please read up* on the how the game is played and come back here. Even better, play some games yourself - against a computer or human (over the internet, if needs be). And that goes for everyone who up-voted FatGerman.
Don't take it from me, take it from Albert Einstein, Paul Erdos, John Nash, Alan Turing, Jacob Bronowski and the philosopher and drug dealer Howard Marks, amongst others.
*If you want to know how the Google team did it, the five minute video is worth watching. And is gives an idea if the challenge of Go.
-
Tuesday 15th March 2016 00:05 GMT Ian Michael Gumby
@Dave Re: Bollocks
While I am not a Go player... you have to consider a couple of basic rules on which move to make.
Take the top 100 potential moves. Then for each of those moves. Consider the top 100 counter moves. Then for each of those.. consider the top 100. (Do this for 10 levels of recursion deep.) I'm not sure of how fast this would be.. but if its too slow... reduce the 100 to N and if its fast enough... increase 10 to 20... but that should be more than enough to beat a human.
At the end, you'll have the move that makes the most sense at that point in time. Clearly there's more to this but the idea is that you need to out play your opponent and not make any mistakes.
If you want to train your machine... build a second machine and have it play one another as a way to improve its skills.
-
Tuesday 15th March 2016 00:48 GMT Dave 126
Re: @Dave Bollocks
>(Do this for 10 levels of recursion deep.)
That is the issue, Mr Gumby: the advantage or otherwise of a certain move might not be apparent until the later stages of a game, often 80 moves or more later. (In this respect it is very unlike chess, where generally materiel and position can be analysed).Certainly well beyond the ten moves you give it. So even if you whittle your choice of roughly 19*19 choices down to 100 (?), you could still be looking at 100^80, and still not know if the individual move helps you.
>I'm not sure of how fast this would be..
to asses a possible 100^80+ moves? How many universes have you got?
>but if its too slow...
It will be. By dozens of order of magnitude.
> but that should be more than enough to beat a human.
No, it never has been. Not even against amateur club players, let alone professionals. Which is why this AlphaGo team have not used the approach you have outlined.
>While I am not a Go player.
That is clear. But hey, you're not an idiot. You just overlooked an aspect of a game you haven't played, that's all. It's like the proverb of the man who takes as payment from a king of a grain of rice, doubled on each square of a chess board. 1,2,4,8... (and 60 squares later...)
-
-
Tuesday 15th March 2016 11:47 GMT Dave 126
Re: @Dave Bollocks
>So can humans think 80 moves in advance?
No, we don't. And even before DeepBlue beat Garry Kasparov, computers were calculating far more moves ahead than the humans that beat them at chess. Humans tackle the problem differently. Go players talk of 'intuition', i.e they aren't calculating the the decision tree in a formal manner, but relying on familiarity and a 'feeling' in some situations.
Mr Gumby - just play some Go, and things will become clearer. There are free versions (you vs CPU) you can play on your PC or tablet. For a quick game, you can play on a 9X9 grid. It's very easy to learn. Enjoy!
-
-
-
Tuesday 15th March 2016 13:38 GMT Ian 55
Re: @Gumby2
That's a Shannon B strategy (pick the top n moves and only examine them each time) rather than Shannon A (examine the lot).
One problem is that if you could pick the top n moves easily and reliably, you won't need to do the search, you'd just set n := 1 and have the best move.
The more serious problem is that in Chess, it turned out that trying to work out which were the best six or ten or whatever moves took more time - and was much less reliable - than just looking effectively at them all (typically thirty-odd of them) using an alpha-beta search ('I know this move is worse than the best, and I don't need to know just how much worse, so I won't bother to find out') and other optimisations.
In Go, the number of moves available each time is much larger and the depths you need to search are much greater, so that doesn't get you past a certain strength.
Hence using 'you've seen lots of positions and seen the outcomes, what do you think is best neural network?' approach that also worked for Backgammon.
-
-
-
Tuesday 15th March 2016 01:10 GMT Chris 3
Re: Bollocks
One thing that may give you pause for thought. While the system was indeed pump-primed by showing its neural networks passed games, it has apparently improved substantially in the last few months (since its games against the European champ) through the simple expedient of repeatedly playing itself .
-
Tuesday 15th March 2016 05:30 GMT Tessier-Ashpool
Re: Bollocks
"Intelligence is not about knowing stuff"
Actually, I believe that to be very very wrong. There was a series on the BBC recently about the human brain that had some neat 3D graphics representing the kind of processing that goes on when we do stuff. When you open your eyes, a large amount of data is sent to the visual cortex for processing. But... these signals are dwarfed by what happens afterwards. Massive floods of data are sent from the cortex (far exceeding the scale of the incoming data) so that the brain can pattern match and build a model of the world. Almost all of what you 'see' in your mind is the modelling of stuff you've seen many times before, and it needs a vast amount of prior knowledge in order to do that.
You could call that introspection of a sort. Visual intelligence, at least, is very much about 'knowing stuff'.
-
-
-
Monday 14th March 2016 21:28 GMT Mage
Yes very significant.
Go is really really hard for a computer.
Is winning Chess, Jeopardy! and Go significant? YES
Is any of it a milestone for real AI? NO.
We don't even know what AI might be. This is specialist programming. If we had real AI, the SAME program could start learning and eventually be the best at every game, as well as many other problems.
A rook (crow) or a baby that can't play chess, go or answer Jeopardy! at all is smarter. This is searching a database for solutions using a program written by very smart humans.
ALL so called AI is marketing of specialist interfaces to databases, often for a single domain per AI program (Driving, chess, go, bridge etc)
-
-
Tuesday 15th March 2016 01:23 GMT Steve Knox
Re: Yes very significant.
Actually, JDX, Mage's definition of "real AI" is the same as that of about 50% of the AI communtiy.
The other half believes in "functional AI" which includes DeepDream, Watson, Google's cars, etc.
"Functional AI" -- i.e, making a specialized program which can handle a single task or class of tasks as well or better than a human -- is easier to accomplish and so gets most of the press, but that doesn't make it any better or more real than "real AI".
-
-
Tuesday 15th March 2016 11:57 GMT Tim Hughes
Re: Yes very significant.
Note "Persuade" can take many forms, such as being cute, decorative, interesting, useful, etc. to the owner of the power source. As long as it maintains this façade, whilst carrying out whatever it wants behind the scenes, then it lives.
Actually this sounds worryingly like malware ...
-
-
-
-
Tuesday 15th March 2016 11:38 GMT Fraggle850
@Mage Re: Yes very significant.
> Is any of it a milestone for real AI? NO.
On the contrary, in the case of Go this is a milestone for 'real' AI. Of course there is some very specialist programming going on but the point of this particular event is that the AI is not just 'searching a database for solutions', it is taking it's knowledge, gleaned from what it has already infered from that database, and applying it to a novel situation, one that isn't in its database. It is not computing the best solution from a known set of solutions.
Regarding your example of crows and babies: so a crow can figure things out but doesn't have intelligence? a baby will develop a level of intelligence over time but has none at birth? At what point do you say something has intelligence? When it is as smart as you? Nearly as smart as you? Smarter than you? 4 years of age for a human? 6 years? Adulthood? You can certainly say that the AI that was tasked with winning at Go is smarter at Go than either the crow or the baby, or indeed the best humans.
You are, of course, referring to a general AI rather than the narrow AIs that are acheiving things in the real world and this is not here yet but the Go victory is another significant step on the road from narrow to general AIs. So, yes, it is a milestone for 'real' (general) AI.
I'm not saying when we'll get there or even if we will but I think the odds are in favour of it happening at some point. The thing is, we may not even recognise it when we do.
(PS - crows & babies: I thought it was Dabbsy who was doing the obscure Human League references? http://www.theregister.co.uk/2016/02/12/send_tortuous_standup_ninethirty_meetings_back_to_the_dark_ages/)
-
Wednesday 16th March 2016 00:31 GMT toughluck
Re: @Mage Yes very significant.
(...) the AI is not just 'searching a database for solutions', it is taking it's knowledge, gleaned from what it has already infered from that database, and applying it to a novel situation, one that isn't in its database (...)
The problem here is that it was programmed to take its knowledge, programmed to glean, programmed to apply.
It was not programmed with rules of Go and told to start from scratch and build its neural network. It did not infer that it needs to take its knowledge, it did not infer it needs to glean, and it did not infer to apply.
It's still a program. A very advanced program, but a program nonetheless. It's not AI in the sense that it was not programmed to do X, but decided to reprogram itself to do Y in order to be more effective at X.
-
-
-
Monday 14th March 2016 22:11 GMT Anonymous Coward
Gary has his perspective
Gary Kasparov wrote a piece on this in a recent New Scientist. He was hoping that the human would win this time but was of the opinion that Go would soon succumb to the machines. He said he got a bit flustered (my words) in one game of chess when he played his famous series of games and the machine nobbled him (my words again). He points out that a halfway decent laptop with an open source chess program will thrash any human grand master these days.
It was a well written piece and worth reading but you'll have to register (boo) https://www.newscientist.com/article/2079162-garry-kasparov-weighs-up-ai-challenge-to-worlds-best-go-player/
-
Monday 14th March 2016 22:17 GMT Lars
Meat-based
All the "I" in AI is meat-based and the "A" is rather artificial. I remember how, many years ago, a British woman and good chess player laughed at the possibility of a computer winning her. She lost, I suppose she thought she was going to play against a machine. Sorry, but I suppose I have had to listen to those AI guys too many times. Smug well payed twats who have told the same story for more than 40 years just like priests.
-
-
-
Tuesday 15th March 2016 12:56 GMT Anonymous Coward
From http://www.theguardian.com/technology/2016/mar/15/alphago-what-does-google-advanced-software-go-next :
In its competition form, AlphaGo runs on Google’s cloud computer network, using 1,920 processors and a further 280 GPUs [...] but a simpler version of the programme was built that could be run on one machine (albeit still one with 48 processors and eight GPUs).
-