Follow Slashdot blog updates by subscribing to our blog RSS feed

 


Forgot your password?
Close
typodupeerror
×
AI Science Technology

AI Learned To Use Tools After Nearly 500 Million Games of Hide and Seek (technologyreview.com) 71

Posted by msmash from the moving-forward dept.
In the early days of life on Earth, biological organisms were exceedingly simple. They were microscopic unicellular creatures with little to no ability to coordinate. Yet billions of years of evolution through competition and natural selection led to the complex life forms we have today -- as well as complex human intelligence. Researchers at OpenAI, the San-Francisco-based for-profit AI research lab, are now testing a hypothesis: if you could mimic that kind of competition in a virtual world, would it also give rise to much more sophisticated artificial intelligence? From a report: The experiment builds on two existing ideas in the field: multi-agent learning, the idea of placing multiple algorithms in competition or coordination to provoke emergent behaviors, and reinforcement learning, the specific machine-learning technique that learns to achieve a goal through trial and error. In a new paper released today, OpenAI has now revealed its initial results. Through playing a simple game of hide and seek hundreds of millions of times, two opposing teams of AI agents developed complex hiding and seeking strategies that involved tool use and collaboration. The research also offers insight into OpenAI's dominant research strategy: to dramatically scale existing AI techniques to see what properties emerge.
This discussion has been archived. No new comments can be posted.

AI Learned To Use Tools After Nearly 500 Million Games of Hide and Seek More

AI Learned To Use Tools After Nearly 500 Million Games of Hide and Seek

Comments Filter:

  • Moving and locking objects sounds like autonomous activities that could apply to some medical uses. Really interesting way to teach AI.

    • Moving and locking objects sounds like autonomous activities that could apply to some medical uses. Really interesting way to teach AI.

      Using tools and cooperation to seek for hidden objects sounds like Skynet.

      • Re: (Score:1)

        by bob4u2c ( 73467 )
        That is what I was going to say.

        You are teaching AI hunting skills! My god man what is wrong with you people? Have you not seen: The Terminator, War Games, Blade Runner, Metropolis, West World, Colossus the Forbin Project, or Short Circuit? Don't teach AI to hunt and kill!

        • Don't teach AI to hunt and kill!

          Hunt and find! I guess it's not too far of a jump to hunt and kill. Oh dear. We just made Skynet.

          • Call me outdated but reality proves the most advanced "AI" known are called Facebook/Apple/Google contractors.

            • Re: (Score:1)

              by Anonymous Coward

              Call me outdated but reality proves the most advanced "AI" known are called Facebook/Apple/Google contractors.

              Your post has been deemed to be against community standards. You can respond in one of three ways. But we'll ignore any response and cancel your account if you respond. -Facebook

      • Who is Al, and why do they keep bringing him up?

        • Who is Al,

          Paul Simon [paulsimon.com]. Coincidentally, "The Walrus" was Paul McCartney.

          and why do they keep bringing him up?

          The guy's a musical legend, and according to the article, not a bad Hide 'N Seek player.

    • For Medical use, we really do not want a trial and error process to treat us.
      AI advantage is the ability to copy learned results and send them to other systems. Unlike us humans who have to teach or relearn actions to every person, with usually a fair amount of data loss. However for an AI, that has a good process to solving a problem, we would want to take its found process and apply it to a non-AI system to do the task over and over. We really do not want a robotic surgeon to apply a random value to see

      • Re: (Score:3)

        by Shaitan ( 22585 )

        "For Medical use, we really do not want a trial and error process to treat us."

        Why? It's what doctors use.

      • Re: (Score:3)

        by Kjella ( 173770 )

        We really do not want a robotic surgeon to apply a random value to see what will happen.

        No, but we might want to make a simulation and let the AI devise its own procedure, then try to figure out if the model is flawed or the AI has actually found a better way. A whole lot of medicine is basically "this is how other experienced professionals do it" which can be subject to group think. Think a bit like traditional chess engines versus AlphaGo, we're not teaching it chess we're just learning it the rules and it teaches itself. If you managed to get an accurate enough simulation of a clogged heart

      • For Medical use, we really do not want a trial and error process to treat us.

        Thats what it all was until the late 19th Century! Feeling sick? You need some leeches to suck on you. Constipated? Well you just need a pigeon to sit on your head.

        It's absurd what people let doctors do in the past.

  • 500 million iterations sounds more like brute-force search, than intelligent behavior.

    • 500 million iterations sounds more like brute-force search, than intelligent behavior.

      If the AI is consistently getting better over the course of 500 million iterations, it isn't brute force. I can't say if this is what the system is doing but I would hope so.

      • It wasn't clear from the article, but it sounds like they are using genetic algorithms to implement their "AI".

        The result is not intelligence in the sense that it understands the world the way we do, but rather it evolved "instinctive" intelligence to solve various problems.

        Bees don't "plan" to visit flowers to collect pollen, they do so out of instinct. Bees can deal with a large variety of issues they encounter in their environment, but their behavior is mostly instinct. It's just complex instinct.

    • Re: (Score:2)

      by Shaitan ( 22585 )

      Given that man had 100k years or more to work it out with each generation raising the next is that unreasonable?

    • Watch the video [youtube.com] and tell me that looks like brute force.

  • Presumably evolution on Earth is driven by surviving things trying to eat you, and finding things to eat yourself (and then producing offspring) rather than playing "hide and seek".

    But I guess calling it "Hunt, Kill, and Eat" could sound a bit too aggressive....

  • Computers are good at playing games and running algorithms. Who knew?

    • Let me see ai have FUN playing a game, and then it can be called intelligence. Fun is a HUGE part of intelligence. Dare I say that without it, we won't evolve.

      • Re: (Score:3)

        by Shaitan ( 22585 )

        Fun is a big part of human intelligence. It isn't a given that something has to think like us to be intelligent.

        • Whoa take it easy pal one thing at a time. "Fun" is beyond the scope of this project. "Fun" will be an interaction between algorithms.

          What they haven't figured out yet is that the essence of human intelligence and sentience is thousands of algorithms running all the time, exchanging information and competing with one another. It is a complex relationship between hardware and software that were explicitly designed for one another.

          One easily spotted advantage we have is that our "algorithms" have the abili

          • Re: (Score:2)

            by Shaitan ( 22585 )

            Right but it has occurred to me that using multiple nets to reflect some of these functions could bring us closer to a human model. Where there is a larger abstraction of "neuron" composed of small nets each functioning and being trained to reflect things like reward behavior/dopamine in competition with opposed nets, etc. Obviously actually doing it is more challenging than saying it but there is a road there to go down.

            It isn't a given we need to go down it though. Our own model of operation isn't necessa

            • I forgot to mention the continuous stream of input that we have. Thousands of sensors constantly sending information to our brains.

              You give a machine thousands of sensors that do not turn off. Software and hardware synergy. And thousands of algorithms constantly exchanging with one another and competing with each for priority. Incentives and punishment.

              And it will learn.

      • Re: (Score:2)

        by Kjella ( 173770 )

        Let me see ai have FUN playing a game, and then it can be called intelligence. Fun is a HUGE part of intelligence. Dare I say that without it, we won't evolve.

        We have plenty intelligent people doing intelligent things that are all business and dead serious during those particular tasks. I really doubt a surgeon has FUN performing surgery, no matter what he does for shits and giggles in his spare time. Though I suppose that would explain dentists...

        • We have plenty intelligent people doing intelligent things that are all business and dead serious during those particular tasks.

          I never said that we didn't have fool out there too. There are a lot of people out there that are simply going through the motions of life, and since they get a decent paycheck, they sit there in life. Kinda like a useless robot that only does it's task, and waits. But at some point, even those types of people break away and have some sort of fun eventually. Maybe they even go waaay further into fun, until it's something that others don't recognize as fun anymore.

          But if you're having fun, it means that

          • Fun just means your dopamine rewards center got triggered. No different from the error back propagation in the neural nets. They are having fun. No one has given them an algorithmic motivation to develop facial expressions and jovial feedback to other algorithms, so you can't see that they're having fun. Getting good feedback in the network is all that fun is. Why do you think drugs are popular?

            • Fun just means your dopamine rewards center got triggered.

              The spark of fun that ignites the release of dopamine isn't triggered by dopamine. Fun triggers dopamine, not the other way around.

              Why do you think drugs are popular?

              Because societal norms aren't.

              • Let me be more pedantic: Humans identifying something as fun just means means our dopamine rewards center got triggered. Sorry for the confusion... I thought that would be clear from the context.

                • Then what triggers the release of dopamine? The way I understand you, you're saying that nothing causes the dopamine to be released, but once it is, the fun begins. My experience is the opposite. An example:

                  You're at a bar, just trying to get laid, when suddenly a hot chick approaches and starts to flirt (maaan, do you remember those days? mmm mmm mmm...). Now the night just got interesting. When did the dopamine get released? Before the girl, or after?

    • Normal algorithms normally will have a predictable set of results. AI algorithms begin to create results that are more difficult to predict, and can perform complex actions beyond the initial scope of the algorithm.

      Game "AI" that we see in most games are predictable, however they may be adjusted with random attributes, like setting up a D&D character, so they can have some variance in a detect box, and hit box accuracy . So if we had a hide and seek video game, we would just just increase values, and s

      • Oh right. "Normal" algorithms normally have a predictable set of results. Lets just call everything AI now. Simulated Annealing? AI. Fuzzy Logic? AI.

        • Oh right. "Normal" algorithms normally have a predictable set of results. Lets just call everything AI now. Simulated Annealing? AI. Fuzzy Logic? AI.

          Yes, that's the definition of AI as used in the industry. You continually insist that only what the industry calls Artificial General Intelligence (AGI) can be called AI. Do you actually think your slashdot posts are going to change industry terminology?

          • Shut the fuck up. The "industry" didn't start talking about "AGI" until about the third hype/failure cycle of AI.

            "AGI" is the equivalent of "AGW". It's nothing more than the fools and charlatans rebranding after they've fully struck out on "AI" and "GW".

        • Oh right. "Normal" algorithms normally have a predictable set of results. Lets just call everything AI now. Simulated Annealing? AI. Fuzzy Logic? AI.

          Actual AI is where the computer says, "Facebook is down, so I'll wait til that comes back up."

  • I must say I was impressed when I looked at the "paper" and saw animations as well as timelines explaining the research.

    What I would have liked to see is the hardware used to run the 500 million+ simulations and how long each one took and what were the ending criteria. Anybody know?

    It was interesting to see the different strategies that the AIs (normally I wouldn't use this term, but the evolution of them is very impressive) that included running away and climbing on top of things.

    I am tempted to apply.

    Gre

  • It would be interesting to see if the AIs parameters are complicated enough to allow for mutations. Just add or subtract from the LSBs and see which ones perform better every few generations (which would probably be represented by a couple 100k of games).

    The ultimate test for this would be to put the AIs on hardware that wasn't SEU hardened and run it in an actual beam of radiation.

  • People look for flying saucers to meet the first intelligent aliens. They are looking the wrong way. Soon AI will pass all human abilities, and when such a device is implanted in a robotic body we will have our first encounter with beings from a non-human individual. Being of superior abilities it will find a way to reproduce. I expect each generation to be more intelligent than the one before it.

  • ... by way of the random walk algo.

  • When my son was young we taught him hide+seek. It was an enjoyable game for him.

    However, there was one flaw in his hiding technique. He assumed that as long as he couldn't "see" you, he was hiding. I usually found him under a lamp table with his legs sticking out. But his head was hidden under the table. He thought he was hidden. We tried to explain it to him, teach him, however his young age prevented him from "getting it."

    When I think about AI, limitations, and failures, these kinds of experience

  • Rhetorical question; it can't 'think' in the first place therefore it can't tell you what it was 'thinking'. Also it can't talk to you at all there's no personality in there.
    This sort of research news is interesting on an intellectual level but I still don't find it to be very significant. We still have no idea how 'thinking' actually works in a biological brain. Can research like this maybe give us some clues as to how that works? Maybe. But that box isn't going to suddenly wake up for it's 500M games of

    • Re: (Score:2)

      by jbengt ( 874751 )

      Also it can't talk to you at all there's no personality in there.

      So my dogs have no personality in there? I guarantee you, they each exhibit a different personality.

      • I guarantee you that they talk to you. Don't put on the dog that you don't understand them. Dogs are brilliant at understanding people, up to and including spoken language. People, however, are abysmal at understanding dogs.

        Reinterpret his comment this way: The AI will not even make the *attempt* to communicate to you, having no personality to make it *want* to.
        • You're either trolling, in which case don't bother because you're awful at it, or you're engaging in 'magical thinking', like way too many people do, and actually believe (in the same vein as a child 'believes' in Santa, the Tooth Fairy, the Easter Bunny, and so on) that these simple computer programs will suddenly 'wake up' and become sentient. If so, sorry buddy, but K.I.T.T. on Knight Rider was just a car and a voiceover by a human, I, Robot was just science fiction (even if it was Isaac Asimov inspired)
      • Your dog is a biological being with a brain that has evolved over millions of years. The machines we're talking about have only been around for a few decades, and an amoeba has more cognitive capability, when it comes right down to it, than any of them. Apples and hand grenades.
  • Wait a few more million iterations until seekers learn how to use a ramp to get on top of each other?

    • Re: (Score:2)

      by Megane ( 129182 )
      Wait until it learns that it can "use tools" to kill the other players. Now you can't hide from me anymore, human!
  • Probability of seek find next round = 100%. Successful tool deployment!

  • So the AI also learned how to hide broken lamps and vases from grown-ups, and why the cat will now need therapy.

  • I tried emailing this company through several email addresses listed on their website and the email addresses don't appear to be valid. I also tried calling them and was unable to actually reach anyone. I also found this record: http://www.buzzfile.com/busine... [buzzfile.com] Is this company legitimate? Has anyone else had actual interaction with them?

Slashdot Top Deals

Remember to say hello to your bank teller.

Close