Forgot your password?
typodupeerror
AI Games

Neural Net Learns Breakout By Watching It On Screen, Then Beats Humans 138

Posted by Soulskill
from the but-can-it-eat-doritos-all-day? dept.
KentuckyFC writes "A curious thing about video games is that computers have never been very good at playing them like humans by simply looking at a monitor and judging actions accordingly. Sure, they're pretty good if they have direct access to the program itself, but 'hand-to-eye-co-ordination' has never been their thing. Now our superiority in this area is coming to an end. A team of AI specialists in London have created a neural network that learns to play games simply by looking at the RGB output from the console. They've tested it successfully on a number of games from the legendary Atari 2600 system from the 1980s. The method is relatively straightforward. To simplify the visual part of the problem, the system down-samples the Atari's 128-colour, 210x160 pixel image to create an 84x84 grayscale version. Then it simply practices repeatedly to learn what to do. That's time-consuming, but fairly simple since at any instant in time during a game, a player can choose from a finite set actions that the game allows: move to the left, move to the right, fire and so on. So the task for any player — human or otherwise — is to choose an action at each point in the game that maximizes the eventual score. The researchers say that after learning Atari classics such as Breakout and Pong, the neural net can then thrash expert human players. However, the neural net still struggles to match average human performance in games such as Seaquest, Q*bert and, most importantly, Space Invaders. So there's hope for us yet... just not for very much longer."
This discussion has been archived. No new comments can be posted.

Neural Net Learns Breakout By Watching It On Screen, Then Beats Humans

Comments Filter:
  • I for one (Score:5, Funny)

    by fisted (2295862) on Friday December 27, 2013 @11:19AM (#45796219)
    I for one welcome our new virtual ass-kicking overlords.
  • by Anonymous Coward on Friday December 27, 2013 @11:20AM (#45796229)

    "I hope the lifter pilot doesn't get too bored." Jarvis is all chummy again.
    "There is no pilot. It's a smart gel."
    "Really? You don't say." Jarvis frowns. "Those are scary things, those gels. You know one suffocated a bunch of people in London a while back?"
    Yes, Joel's about to say, but Jarvis is back in spew mode. "No shit. It was running the subway system over there, perfect operational record, and then one day it just forgets to crank up the ventilators when it's supposed to. Train slides into station fifteen meters underground, everybody gets out, no air, boom."
    Joel's heard this before. The punchline's got something to do with a broken clock, if he remembers it right.
    "These things teach themselves from experience, right?," Jarvis continues. "So everyone just assumed it had learned to cue the ventilators on something obvious. Body heat, motion, CO2 levels, you know. Turns out instead it was watching a clock on the wall. Train arrival correlated with a predictable subset of patterns on the digital display, so it started the fans whenever it saw one of those patterns."
    "Yeah. That's right." Joel shakes his head. "And vandals had smashed the clock, or something."
    "Hey. You did hear about it."
    "Jarvis, that story's ten years old if it's a day. That was way back when they were starting out with these things. Those gels have been debugged from the molecules up since then."
    "Yeah? What makes you so sure?"
    "Because a gel's been running the lifter for the better part of a year now, and it's had plenty of opportunity to fuck up. It hasn't."
    "So you like these things?"
    "Fuck no," Joel says, thinking about Ray Stericker. Thinking about himself. "I'd like 'em a lot better if they did screw up sometimes, you know?"
    "Well, I don't like 'em or trust 'em. You've got to wonder what they're up to."

    • Reminds me of a story my AI prof told about creating an expert system for deciding when cheese was fully ripened.

      The cheese company had a person who's job it was to check the cheese and would go and poke the wheels to see if they were ready. The company created a robot to go and do the same job and trained it by having it poke fully ripened and unripened cheese. The problem was that when it was put it use it failed miserably at correctly and consistently telling if the cheese was ripe. The problem was the
      • You would be able to tell a real AI its mistake, and it would be able to figure out how to correct it. 3D print its own smell sensors...

      • by fatphil (181876)
        Cheese? That's nearly a solved problem in AI, here's some recent state-of-the-art work:
        http://www.youtube.com/watch?v=GzzOw2tmb3A
  • by Anonymous Coward

    Next up wee have Sky Net.

  • AI (Score:5, Insightful)

    by ledow (319597) on Friday December 27, 2013 @11:23AM (#45796253) Homepage

    For once, something based on proper AI (rather than human-generated heuristics).

    However - notice it's limitations: Where there is a direct correlation between where you need to be, and where something else is on the screen (basically a 1:1 relationship in Pong, for example), it can cope with going higher or lower as required.

    But when you put it into something that has more than a single thing to "learn" (move left/right, avoid bombs, shoot aliens, choose which aliens to shoot, don't shoot your own base, etc.) then the amount of training required goes up exponentially. And thus we could spend centuries of computer time in order to get something that can do as well as a simple heuristic designed by someone who knows the game (not saying heuristics don't have their place!).

    "Trained" devices require training relative to some power of the variety of the inputs and the directness of their correlation to the game-arena. And thus, proper AI is really stymied when it comes to learning complex tasks.

    But still - this is the sort of thing we should be doing. If it takes an infant two years with the best "computer" in the universe that we know of to learn how to talk, why should we think it will take a machine at even the top-end of the supercomputer scale (which can't have as many "connections" as the average human brain) any less?

    • Re: (Score:3, Interesting)

      by Anonymous Coward

      If it takes an infant two years with the best "computer" in the universe that we know of to learn how to talk, why should we think it will take a machine at even the top-end of the supercomputer scale (which can't have as many "connections" as the average human brain) any less?

      Because we're learning languages in the wrong way.

    • There is another aspect that limits this sort of machine learning, the need for direct positive/negative feedback. Games that don't have a score counter are pretty much unlearnable. Obviously you can base the feedback on something else like level reached but in some games, like kings quest for example, a neural network will not be able to figure out when it is doing well and when it isn't. Often a separate heuristic AI is needed just for this.
    • Re:AI (Score:4, Interesting)

      by StripedCow (776465) on Friday December 27, 2013 @11:38AM (#45796401)

      If it takes an infant two years with the best "computer" in the universe that we know of to learn how to talk, why should we think it will take a machine at even the top-end of the supercomputer scale (which can't have as many "connections" as the average human brain) any less?

      Because neurons are much slower than transistors?

      • by Anonymous Coward

        A neuron is more like a network router with local integrated storage, packed in a density somewhat comparable to integrated circuits.

    • by s.petry (762400)

      If it takes an infant two years with the best "computer" in the universe that we know of to learn how to talk, why should we think it will take a machine at even the top-end of the supercomputer scale (which can't have as many "connections" as the average human brain) any less?

      While I kind of agree with the point, it absolutely doesn't take 2 years to learn how to talk. It takes a few months to learn to talk (which would include learning that sounds have meaning). Just like it doesn't take over a year to learn to walk, it takes a couple weeks. Interestingly, they learn some very complex things all at the same time.

    • by tomhath (637240)
      This looks like pretty standard forward chaining. The twist they threw in was to let the program observe action/result and form rules to use next time. Once those rules were defined it's just boolean logic. As you say, the difficulty going forward is finding a domain where a machine can experiment and deduce positive/negative outcome when the domain is nontrivial.
    • Why not include a heuristic processor in the AI, that would override the statistical training in certain cases?

      So you could tell the program, in real time while its playing, something like "Watch out for bombs while moving left or right" and it would be able to ignore what its statistical training told it to do, in a context where the training told it to move right but that would send it into a bomb.

      • by syockit (1480393)
        Cool! One day we could ourselves play the role of GLaDOS while the AI finds its way out of puzzling mazes!
  • by Anonymous Coward on Friday December 27, 2013 @11:30AM (#45796315)

    This neural-net-combined-with-trial-and-error style of algorithm is typically referred to as a "JavaScript Programmer"-type algorithm in recent AI literature. (I'm being completely serious, too, in case you think this is a joke; it isn't.)

    The name derives from the similarity between how these kinds of algorithms work, and how JavaScript programmers tend to work.

    Both the algorithms and JavaScript programmers use a very basic, minute form of pseudo-intelligence.

    This small dab of pseudo-intelligence is then used to repeatedly attempt to solve a problem, followed by an analysis of the success of the attempt.

    In the case described in this article, it involves the computer trying to play the game, with the aim of winning.

    In the case of the JavaScript programmer, it involves the programmer repeatedly searching through Stack Overflow, finding code to copy-and-paste, and then hoping that it works well enough to trick the customer or employer into thinking the job is done.

    The summary should have probably mentioned this, but I suspect that the submitter may not be following the latest AI journals and research very closely.

    • Re: (Score:3, Funny)

      by cfulton (543949)
      Truer words were never spoken:

      the programmer repeatedly searching through Stack Overflow, finding code to copy-and-paste, and then hoping that it works well enough to trick the customer or employer into thinking the job is done."

      • Truer words were never spoken:

        the programmer repeatedly searching through Stack Overflow, finding code to copy-and-paste, and then hoping that it works well enough to trick the customer or employer into thinking the job is done."

        If it really works, if the specifications are met, and if it passes testing, then the job is done.

        Wisely leveraging the shared knowledge of others is a good thing to do.

    • by cascadingstylesheet (140919) on Friday December 27, 2013 @01:46PM (#45797611)

      This neural-net-combined-with-trial-and-error style of algorithm is typically referred to as a "JavaScript Programmer"-type algorithm in recent AI literature. (I'm being completely serious, too, in case you think this is a joke; it isn't.)

      The name derives from the similarity between how these kinds of algorithms work, and how JavaScript programmers tend to work.

      Funny, of course :)

      But, you got me thinking. The JavaScript programmer is generally trying to affect the appearance of stuff on the screen, therefore, he looks at the stuff on the screen, and tries to affect ... the stuff on the screen. So, it makes more sense than it might.

      Our new pong-playing overlords, on the other hand, if they are actually doing something important like remotely fighting wars or trying to save people or something, well, then we don't really know if they are looking at the right input, and it becomes much more important that they, and we, understand exactly how they are coming to their decisions.

    • In the case of the JavaScript programmer, it involves the programmer repeatedly searching through Stack Overflow, finding code to copy-and-paste, and then hoping that it works well enough to trick the customer or employer into thinking the job is done.

      And now, I'm wondering if there is another way for creating DOM manipulating Javascript. I mean, I can most of times make a Linux module by reading the documentation of a device and writting code that makes it work (but for some devices, it's the Javascript wa

    • "used to repeatedly attempt to solve a problem, followed by an analysis of the success of the attempt."

      The above is exactly how humans learn to play simple games. Sure, you learn a few rules beforehand but then you actively - and to an extent subconciously - engage in trial and error about what to hit/kick/click at what time in what scenario. Its called "practice". No one for example becomes a good football (soccer for the yanks) player by analysing angles of attack of other players feet - they just go out

    • Its also more formally called TDD. Create some code that tests the suitability of the existing code to solve the problem. Then randomly change the code until it passes the test, and all the others. Repeat, rinse, etc.

  • by Anonymous Coward

    Where did the researchers find the "expert Breakout and Pong players" to match their neural net against? Was it that same loudmouth kid down the hall who is always "beating the spread" on football?

  • The question is not, "when can a bunch of machinery beat a human at X." The question is "when can a bunch of machinery beat a team of humans _with access to similar computational resources_ at X." I don't see much progress there.

  • "Exterminate... Exterminate...."

    Actually, when they become advanced enough, we won't need to work anymore.

    I'll buy TWO. One to do my job and one ... just in case.

    • by Nyder (754090)

      "Exterminate... Exterminate...."

      Actually, when they become advanced enough, we won't need to work anymore.

      I'll buy TWO. One to do my job and one ... just in case.

      Dalek's are mutate life forms, riding in a machine. They are not robots or A.I.'s.

      • Dalek's are mutate life forms, riding in a machine.

        Now I have an image of the little squishy Dalek sitting inside in a little chair, turning a little wheel and going "wheeee!"

        (Username recognised)

    • by maharvey (785540)

      Nah, it won't be the machines taking over. When machines become advanced enough, the 1% will no longer need the rest of us humans to grow their food, to make their toys, to be their servants and chauffers. Why would they pay us to do nothing? Why let us use up food and oxygen? It will be time to exterminate the teeming masses. In the name of sustainability, no doubt.

  • by Anonymous Coward

    You were just asking for an oblig [xkcd.org], weren't you?

  • by wjcofkc (964165) on Friday December 27, 2013 @11:57AM (#45796557)
    Tetris, by nature, would prove most interesting. I myself never made it past level 10, and I've never seen anyone make it past level 20. I wonder what the breaking point for this neural net would be after a few days of practice. I would love to see a video of it starting from level one and making it's way to the insanity of level 50 - if it's up to the task. I imagine a super computer would have too much latency.
    • Re:Tetris (Score:5, Informative)

      by Sigma 7 (266129) on Friday December 27, 2013 @01:18PM (#45797335)

      Tetris is a solved problem if you're going for survival (assuming you don't get an extremely unlucky piece selection). Since AI has access to the current piece, the next piece, and can do a probability check on the next piece, it can basically last forever.

      I myself never made it past level 10, and I've never seen anyone make it past level 20.

      Tetris: The Grand Master: http://www.youtube.com/watch?v=jwC544Z37qo [youtube.com] - fast forward to 3:00 to see first majoor speedup, 4:45 for final speedup, and 5:01 for invisible pieces.

      That, and 999999 was done on a real NES within 3 minutes 11 seconds: http://www.youtube.com/watch?v=bR0BKCHJ48s [youtube.com]

  • by Anonymous Coward

    They should spin up two instances of the neural net and have it play itself

  • by portwojc (201398) on Friday December 27, 2013 @12:05PM (#45796629) Homepage

    The AI has another advantage over us human players with the Atari 2600. No blisters.

    • by GTRacer (234395)
      Also, it isn't tempted to pull the rubber cover off the stick and suction it to its forehead. Do you know how many forehead hickeys I gave myself with that stupid thing?!
  • by Stargoat (658863) <stargoat@gmail.com> on Friday December 27, 2013 @12:10PM (#45796677) Journal

    But has it learned to let someone else design Breakout and then steal a couple thousand dollars from him for his efforts? When it does that, it will truly be an intelligence. (And it will be a superior intelligence if it leaves off the black turtlenecks.)

  • by Anonymous Coward

    Show me an AI that can play minecraft; that would be impressive.

    • by Anonymous Coward

      Define the goals of minecraft. If you mean "An AI that can form its own aesthetic desire of what a nice house/castle/statue/dick should look like and creates it for fun," I'll agree with you, but if the standards are low, I can make an AI that plays minecraft by looking straight down and holding left click for a few minutes. I'll call it a zombie survivalist AI.

      • by Anonymous Coward

        Kill the dragon in hardcore mode.

  • by The-Ixian (168184) on Friday December 27, 2013 @12:15PM (#45796711)

    to farm gold for me.

  • So it can play Breakout, big deal.

    Wake me when it's giving the checkers-playing chicken a run for her money.

  • The interesting part of the slim article was the part left out. Why did not not perform as well on some of the games. There was not much detail on that issue. I'm not familiar with the poorly played game, but I would guess they introduce a level of visual complexity that overwhelms the AI?

    Other than that, simply astounding accomplishment.

  • ...Welcome the new King of Kong!

  • Winning means nothing unless you can enjoy it. Is the computer having any fun?
  • Neural Net Learns Breakout By Watching It On Screen, Then Beats Humans

    Women and children and nerds first!! The machines are coming!

    Oh, my mistake. I thought it said "neural learns to break out" and then something about beating humans.

    This is one reason why most people don't Capitalise Every Word In A Headline.

  • TFA does not describe an advancement in AI technology whatsoever.

    It is an external 'computer player'...We have had AI's that play video games virtually since we had video games.

    Take good ol' Tecmo Bowl...you play against an AI opponent that does absolutely everything this AI did and more.

    This is not an AI advancement, it is....an **application** of new and better **sensor inputs** for an external AI

    I don't see anything in this that would indicate we are some kind of 'step' closer to having Terminator kill b

    • Nah, not related. In-game AI is written specifically with that game in mind, often knowing more than the eye can see. This is a general-purpose thing that attempts to learn only with visual input, without direct access to the program itself.

  • "Would you like to hear the song I learned today while we play? Daisy, Daisy, give me your answer do ..."

  • ... they don';t need us creators any longer. Now if only we could see what they come up with about their origins after we are gone.

  • This is neat but it needs to be understood that mimicking what other players do without the understanding of strategy and deeper conceptual thought severely limits what the AI can do. This AI could never learn to play sophisticated games simply because it works by copy-pasting basic behavior. Even something with as basic rules as Go would be far beyond this AI.
  • Very regularly, someone writes a clever new algorithm to crunch a specific limited set of data more efficiently.

    Repeat it with me: "This is not an AI breakthrough".
  • How...About...A...Nice...Game...Of...Chess...
  • I didn't know anyone would still dare call themselves an expert Pong or Breakout player any more.
  • If computers could do that for video games, do you think they could perform the same tasks behind a real gun in real life in a real battle? It would help armies a lot if such technology could be used.
  • See http://www.arcadelearningenvironment.org/ [arcadelear...onment.org] for a few other approaches to this de-facto AI test.

  • "However, the neural net still struggles to match average human performance in games such as Seaquest, Q*bert and, most importantly, Space Invaders."

    There's the Singularity put off for another year.

Real Users hate Real Programmers.

Working...