Slashdot Mirror


Machine-Learning AI Now Beats Humans At Super Smash Bros. Melee (qz.com)

"The AI is definitely godlike," one professional player told Quartz. "I am not sure if anyone could beat it." An anonymous reader quotes their report about an AI's showdown with the best players of Super Smash Bros. Melee: Of 10 professionals that faced the bot, each one was killed more than they could kill the bot... But the bot was once only as good as a mere mortal. At first, Vlad Firoiu, creator and a competitive Smash player himself, couldn't train 'Phillip' to be as strong as the in-game bot, which he says even the worst players can beat fairly easily. Firoiu's solution? He started making the bot play itself over and over again, slowly learning which techniques fail and which succeed, called reinforcement learning. Then, he left it alone.

"I just sort of forgot about it for a week," said Firoiu, who coauthored an unreviewed paper with William F. Whitney, the NYU student [who helped him] on the work. "A week later I looked at it and I was just like, 'Oh my gosh.' I tried playing it and I couldn't beat it."

Business Insider points out that their AI read the players positions, velocities, and states directly from the game's memory, so the AI responds six times faster than a human player. To compensate it played as Captain Falcon, the game's slowest character, but there was one crucial glitch. "One particularly clever player found that the simple strategy of crouching at the edge of the stage caused the network to behave very oddly, refusing to attack and eventually KOing itself by falling off the other side of the stage."

12 of 78 comments (clear)

  1. Not really a success for the AI by meerling · · Score: 4, Insightful

    Sure, you got a better AI than you started with, but it's still cheating, even if it is using the slowest character in the game.
    Now program it to emulate the time delays for using a controller and having to recognize what's happening on screen instead of the instant data i/o from direct machine & memory access.
    If you can reliably beat humans at that level, then you've actually done something worth talking about.

    1. Re:Not really a success for the AI by sunking2 · · Score: 2

      No, the purpose of AI should be that it can problem solve and adapt to a situation as well, or better than us. With an unfair reaction benefit it can actually problem solve worse, yet still win simply because it has an external advantage. That doesn't sound like a win for AI to me.

    2. Re:Not really a success for the AI by lgw · · Score: 4, Interesting

      His point was that this AI didn't use the same inputs and controls a human does, so it's not a fair test. Adapt this AI to use only the screen buffer, and give it input lag to match a mechanical controller, and you'll have something.

      --
      Socialism: a lie told by totalitarians and believed by fools.
    3. Re:Not really a success for the AI by lucasnate1 · · Score: 3, Interesting

      I am not sure how relevant that is. Part of what makes us intelligent is our inputs and outputs. It is possible that dolphins or some other animals are much smarter than us but because they don't have opposable thumbs we developed and they didn't. Don't you think it would make us more capable if we could direct digital input? imho, the only test for the quality of an AI is "what it can do", not how.

      Then again, I do agree that this project and projects similar to it are not exactly creating "intelligence", they are creating an expert system, good for one thing and one thing only.

    4. Re:Not really a success for the AI by Kjella · · Score: 5, Insightful

      No, the purpose of AI should be that it can problem solve and adapt to a situation as well, or better than us. With an unfair reaction benefit it can actually problem solve worse, yet still win simply because it has an external advantage. That doesn't sound like a win for AI to me.

      If a self-driving car can drive better than you because it's got 360 degree vision, millisecond reaction time and the capacity to focus on ten different factors at once is that "cheating"? I think that's a matter of perspective, limiting it to the wheel's turning rate and the pedals' actuation force sounds like unreasonably hampering the performance. Maybe that's not a "fair" fight, but I'd say we probably want the computer to play to its strengths and not mimic our weaknesses.

      --
      Live today, because you never know what tomorrow brings
    5. Re:Not really a success for the AI by Anonymous Coward · · Score: 4, Insightful

      No, this is like a self-driving car that only works in GTA because it has a pipe into the hard data for locations of obstacles and other vehicles etc.

    6. Re:Not really a success for the AI by ShanghaiBill · · Score: 4, Informative

      I remember when "AI" was defined as, "indistinguishable from human."

      AI has never been defined as that, at least not by people working in the field. There is a particular subcategory of AI focused on human-level performance, called "Strong AI" or Artificial General Intelligence, but few AI researchers are working on that, or consider Strong AI a realistic near term objective.

    7. Re:Not really a success for the AI by ShanghaiBill · · Score: 3, Insightful

      imho, the only test for the quality of an AI is "what it can do", not how.

      If I hire two men to dig ditches, and I give one a shovel and the other a backhoe, it is silly to say that the second is ten times as intelligent as the first. Intelligence is the ability to formulate an effective course of action, not the ability to execute it. Of course, the physical ability to execute is important, but it is not "intelligence".

    8. Re:Not really a success for the AI by crow5599 · · Score: 2

      Yeah, it's weird that this is being bragged about, considering Demis Hassabis and DeepMind trained their game-playing AI so that the only input it received was the pixels on-screen. You'd think advances in game-playing / learning AI would build on top of that, not go backwards.

      For anyone who hasn't seen this yet, here's footage of some of the technology behind DeepMind's AlphaGo (the AI that beat Lee Sedol at Go last year) learning to play old arcade games, eventually becoming superhuman at them. I jumped ahead to right before the demo: https://youtu.be/rbsqaJwpu6A?t...

    9. Re:Not really a success for the AI by ausekilis · · Score: 2

      I fail to see how what he did is really any different than "in-game A.I.". Sure, he didn't succeed at making a TASBOT with a controller and camera, but he did manage to best Nintendo's top computer-controlled player using a Neural Network. I'd say that's significant in it's own right. After all, enemies in Halo or GoW work on in-game memory and don't have controllers in-hand.

      Would it have been much cooler to have ROB with a gamecube controller-in-hand and some fancy kinect cameras in his head? Of course it would. But that's all for show really. It may be more realistic if he had introduced a delay between state recognition and action taken to simulate input perception lag. He didn't, and he managed to create an awesome computer opponent.

  2. Re:Fake News by belthize · · Score: 2

    And yet we can't seem to create AC posters who can pass the Turing Test.

  3. History in the making.... by WolfgangVL · · Score: 2

    Soon after, the emerging strong AI applications being developed by the primitive tech titans of the time began besting humanities brightest and most skilled players at various leisure activities. First, simple board-games, chess came early, then go, and soon the entirety of the skills based social board and card games of the 19th century. This was followed by the more modern trivia, grammar, and logic based social leisure activities. Video-games came next (the popular two 2 and 2.5D visual based games of the time) and finally, in march of 2021, (incidentally, nearly 35 years to the day, before the escalation of the Humanity First Treaty, which directly led to the great war 2057) the first paintball and laser-tag "bots" showed the world the killing potential of fully automated combat. Later that same year, Earths first fully robotics sports teams eclipsed humanities best athletes at nearly every skills based antithetic sport (with the exception of water-polo)

    --
    You are being ripped off every second of every day, so that advertisers can help rip you off even more tomorrow.