Google DeepMind's AI Beats Humans At Even More Computer Games 96
An anonymous reader writes: Google DeepMind's learning algorithm has trumped human performance in an even greater range of games from the Atari 2600. The system's performance in classic games for the 80's games console has improved steadily since it was revealed in April last year (video) and a paper released yesterday shows it besting people in 31 titles.
80's console? (Score:4, Informative)
The Atari 2600 was released in 1977.
Re: (Score:3)
Re: (Score:2)
The Atari 2600 was released in 1977.
If you're going to be pedantic, at least be accurate. The Atari VCS was released in September of 1977. It was renamed the Atari 2600 in 1982 to distinguish it from the newly released Atari 5200.
So, to answer your question, yes, that date is within the decade known as "The 80s".
Re: (Score:2)
Color me shocked (Score:3, Interesting)
That a computer can beat humans at a computer game.
The real question is, can a computer beat a human at a human game? Chess, yeah. Go, not so much.
Hasn't reverse engineering been around for a while now? If a computer wasn't better and faster at that than a human, that would be the true surprise.
This just in -- maybe it doesn 't require "intelligence" to win most computer games, just good memory and fast reflexes.
Re:Color me shocked (Score:5, Funny)
Computers have been steadily beating humans at more and more games, "real life" ones or not. Yes, this includes Go. Ironically, humans still beat machines at things that any idiot could do, such as walking or talking or seeing. But even those things are they are getting better and better at (and we aren't), enough to beat us at various surveillance things like recognizing people or license plates.
Humans still beat computers at Calvinball, so there's that.
Re:Color me shocked (Score:5, Insightful)
I don't want my computer to play classic atari games better than me I want it to make my work easier so I can have more time to play classic atari games, just saying.
Re: (Score:1)
Well, eventually the computer will be so good at your job that your company won't need you anymore, and at that point you will have tons of time to play classic atari games. Just saying.
Re: (Score:2)
Re: (Score:1)
You assume that in a world where robots do all the work humans who do not own robots get nothing.
That assumption is correct, under a purely capitalistic society.
Unless there is also a move to a more socially oriented society, there will be a lot richer few and much poorer many, when the majority of all work is done my machines.
Re: (Score:2)
You assume that in a world where robots do all the work humans who do not own robots get nothing.
This depends on whether robots replace soldiers before replacing too many of the other jobs...
Re: (Score:2)
On Caprica *all* the humans got nuclear missiles, the Cylons didn't differentiate.
Re: (Score:2)
I don't need a super learning neural network artificial intelligence to automate 30-40% of all the jobs at my company.
Yes the TPS reports will be in on time with the new cover sheet... I just added it to the script that makes them for me.
Re: (Score:2)
Yes the TPS reports will be in on time with the new cover sheet... I just added it to the script that makes them for me.
Right, but if they can get a robot to know when and how to modify the script, they can downsize you also, and improve company profits that much more.
Re: (Score:2)
Here is the thing if every company has learning AI doing all the work and nobody has a job anymore who the heck is going to buy the companies products and services?
Re: (Score:2)
Presumably all those nouveau-riche AIs will have cash to spend... :)
Have Patience... (Score:2)
I want it to make my work easier so I can have more time to play classic atari games
One step at a time. This is just the beginning of "real" computer AI iRobot ( or Robot & Frank ) style. Sure, this seems a trivial application, who needs it. But you have to start someplace, and game decision making is a good place for many reasons.
Re:Color me shocked (Score:5, Funny)
Pedantically speaking, computers have been beating humans at videogames since they first appeared in arcades.
Re: (Score:1)
Re:Color me shocked (Score:4, Informative)
The real question is, can a computer beat a human at a human game? Chess, yeah. Go, not so much.
Neural nets are rapidly gaining on old fashioned hand coded algorithms. Here is a Go playing NN [arxiv.org], that can beat Gnu Go after only a few days of learning. Progress is rapid, and computers will overtake the best humans at Go within a few years.
It's all in the reflexes (Score:5, Interesting)
Computer with sub-millisecond reaction time and ability to perfectly calculate matrices, vectors and quaternions as well as predict positioning in x amount of seconds beats person. No-one should be surprised.
Re: (Score:2, Informative)
Deep mind is a neural network based computer. This isn't a competition to aim a laser at a brightly colored balloon. This is a competition to teach a deep neural network strategy and game mechanics. Highly abstract concepts which are not easily encoded in expert systems.
Re:It's all in the reflexes (Score:4, Insightful)
Well, sortof. From TFA:
"However, the system's continued poor performance in Ms Pacman exposes a weakness that DeepMind discussed earlier this year. The limitation stems from the DeepMind system only looking at the last four frames of gameplay, about one fifteenth of a second of the game, to learn what actions secure the best results." (my emphasis)
GP misunderstands the ML aspect of this, but it does come down to reflexes and precision in this specific project. It is nevertheless interesting to investigate which games the net performs badly on and which ones it doesn't.
In a way, this is also a manner of 'ranking' games: the harder it is for such a system to perform well at it, the more cerebral and less primitive/physical it probably is (although I don't want to imply that one type is better than the other)
Re:It's all in the reflexes (Score:5, Informative)
Computer with sub-millisecond reaction time and ability to perfectly calculate matrices ...
This is NOT about computers being able to play well. It is about computers LEARNING to play. The point of TFA, is that DeepMind was simply given the goal of "winning", and then learned on its own how to play the game and maximize the score.
Re: (Score:1)
Let's play global thermonuclear war with it. (Score:5, Funny)
What side do you want?
1. USA
2. USSR
3. China
4. United Kingdom
5. France
6. India
7. Pakistan
8. North Korea
9. Israel
10. NATO
11. Iran
Re: (Score:1)
12. Cylon
Re: (Score:1)
Re: (Score:2)
What I've found interesting is that neural nets are getting better at deterministic games. It's no news that neural nets play backgammon at and beyond world championship level, but if I understand the literature correctly, neural nets are now playing Go better than programs based on just calculating move trees. I also understand that there have been inroads into chess (even though chess already is played beyond world championship level by computers).
So no, Google has not found something at all new here, but
But can it win against a human players.... (Score:2)
Re: (Score:2)
Actually, a computer with access to Google's databases may be able to. CaH is partially luck-based, partially meme based but also based upon understanding the round interpreter's background, emotional reactions, feelings, humor etc.
It might know better than anyone which card selection would resonate best with someone's senses as well as the groups' senses (because there is also a group pressure subject based around the group's laughter and what the group as a whole thinks is 'funny').
Ever played with super-
Re: (Score:2)
And it's that understanding that makes me think a computer wouldn't do it very well. It might be able to gauge a popularity score with the general public for a response, but it ultimately depends on the personality of each round's caller.
I wouldn't want to.... it would cross a line, IMO. I can respec
I hate these stories (Score:5, Interesting)
I hate these stories. Games were designed (albeit evolutionarily, through generations of culture) to exploit specific human cognitive limitations in exhaustive search and look ahead, and thereby force us to fall back on things like heuristics and strategies. This makes games unpredictable and interesting.
But computers don't have those limitations. Of course they can out play us at games. They also add faster than we do.
This is all IBM's DeepBlue was, a massive, massive lookahead machine which used a little human-discovered / human programmed rules of thumb to reduce the search space and then human-discovered, human programmed rules of thumb for judging the relative goodness of each move.
The fact that computers are good at beating humans at something specifically designed to make humans perform badly is not an advancement in A.I.
Well, OK it is, but that's not saying much.
Re:I hate these stories (Score:5, Insightful)
But that's not at all the point of this article. The point of this article is that a computer program learned--in a manner SOMEWHAT analogous to human learning--through practicing how to play certain video games without having any game-specific special programming. AI opponents have existed as long as there have been video games (or close to it) and you're right, if that's what this article was about, it would be be boring. Neural net learning by examining visual output--now that's pretty cool.
Re: (Score:1)
It didn't learn how to "play" anything. It just mapped some inputs to some outputs. That's it. It can't plan, and it has no memory. It totally fails at anything needing a goal or series of steps, such as needing to get a key before opening a door in Montezuma's Revenge.
Re: (Score:2)
You're point is valid. WRT to neural networks, I am not impressed really. In a nutshell, I think it's a disguised way of doing statistics. An iterative, on-analytical way. With neural nets, after it's trained, no one can tell why the neural net functions as it does and no one can tell you when the neural net will do something completely insane.
There is no analytical framework which decomposes a neural network which has arrived at THESE weights on THESE node with THIS many layers using THIS algorithm to upda
Re: (Score:3)
In a nutshell, I think it's a disguised way of doing statistics. An iterative, on-analytical way. With neural nets, after it's trained, no one can tell why the neural net functions as it does and no one can tell you when the neural net will do something completely insane.
Just like training a biological brain. And yet, those seem to be somewhat useful.
Re: (Score:2)
OK that's the conceit of NN in a nutshell- just like a biological brain, so you said it. To me that's like saying a camera is like an eye. The brain is more than just neurons firing over synapses and reinforcing the ability to communicate across synpases. For example, nitrous oxide diffuses through the brain and is used in signalling. There are other things like that going on.
I am not saying that I think NN are worthless. I probably came across like that; a delay reaction from years of overexposure to NN ch
Re: (Score:2)
We all benefit from 80s-style expert systems every day. Even your GPS is a form of expert system.
The idea we'll soon be of replacing humans, which is causing a sort of mini moral panic amongst some engaged and intelligent but non-expert part of the population is totally off course.
Just focusing in on driving applications, lane awareness and accident avoidance are two great uses of AI but the thing is, they serve only to take minor, stereo-typical (get back in your lane! Break right now as hard as you can!)
Re: (Score:2)
AI is not going to parse out every possible set of real world events which a driver may encounter. Google is finding that out now. The situations are too varied, too unpredictable (the technical term for this is "fucked up") and engage too many independent actors whose reactions are unknowable but critical.
Actually, I know some guys on the self-driving car team, and Google cars already handle just about everything safely. What they're focused on now isn't so much handling strange situations, but optimizing the car so it behaves like a human driver, to avoid confusing other human drivers -- or being taken advantage of by them.
Re: (Score:2)
OK that's the conceit of NN in a nutshell- just like a biological brain, so you said it. To me that's like saying a camera is like an eye. The brain is more than just neurons firing over synapses and reinforcing the ability to communicate across synpases. For example, nitrous oxide diffuses through the brain and is used in signalling. There are other things like that going on.
Meh. So there are some additional interconnections. Are those actually essential? It seems unlikely to me, but they could certainly be added if they are.
It may be the start of a good way to model the actual working of the brain.
Irrelevant. Oh, I suppose it may someday be relevant to neuroscientists whose goal is to understand the brain rather than to create useful systems. But for the people interested in being able to create automated systems that can be taught to make complex decisions effectively, what really matters is that it seems to work very well. Sure, the fact that we do
Re: (Score:1)
The only game in town is the mating game. (Score:1)
Centipede (Score:2)
How could they have missed E.T.? (Score:1)
So, they haven't yet exposed it to E.T. The Extraterrestrial, or they already have, and DeepMind refused to continue playing?
Grand Theft Auto? (Score:2)
So I'm beat by a bot (Score:2)
/ That's a real Modern Warfare 3 name
huh? (Score:2)
Self-important nigilists (Score:2)
Looks like 90% of commenters in this thread are too proud of their superior human brains to even try and get the point of the experiment. Researchers made a computer which can learn to achieve goals with no instructions, and you mix it up with custom game AI or bitch about how it is not fair to compare scores with biologically limited humans. This is just depressing.
Go DeepMind!
Re: (Score:2)