Gambit Game Is Drama School For AI

By Jim Rossignol on March 18th, 2011 at 2:42 pm.

Act, or death.
The Singapore-MIT Gambit Game Lab is intended for clever types to make clever games, and look, they done gone done that with Improviso. It’s a 2D multiplayer acting game, where you take on a role, play it, improvise, and use your own dialogue chat and even props in the game world. A neat enough idea, but the actual purpose of the game is a bit deeper, as Gambit explain: “The longer term goal is to train an AI system with data collected from thousands of people telling stories together with the same characters, sets, and props. Once trained, this AI system will be able to play the role of one or more characters that can converse and interact with other AI- or human-controlled characters.”

Makes you wonder how many other games could actually be crowd-training AI for specific tasks. Anyway, go take a look.

, , .

18 Comments »

  1. dadioflex says:

    Isn’t this kinda what Jason Rohrer failed to deliver?

    • Hoaxfish says:

      You might need to narrow that down a bit more

    • zbeeblebrox says:

      Uh, you can go play Sleep Is Death right now if you want; it’s been out for ages.

      It’s kind of fiddly, but once you take the time to get used to how sets are built, it can be pretty fun. In fact, just watching other peoples’ sessions is a blast

  2. Navagon says:

    I remember trying to train bots how to play Quake 2. This sounds a fair bit more advanced than that. Interesting concept and the graphics are nice too. Will you do an update on this one if / when it makes some progress on the AI training?

  3. oliwarner says:

    There’s something about the screenshots that are very reminiscent of Psychonauts. Not just the 2Dness of some of it but even some of the assets (like the tanks) look like they’ve been lifted.

    Damnit. Now I want to play through Psychonauts again. I don’t have nearly enough time for that.

  4. The Army of None says:

    Interesting idea. Will be fun to see if this works out, maybe will lead to some more realistic game AI

    • Weylund The Second says:

      For some types of games, perhaps. Certainly for RPGs.

  5. DJ Phantoon says:

    oh my god

    IT CAN LEARN

    Seriously though I’m not really okay with that. Especially if some jokers run it a few thousand times to skew the AI towards questioning its boundaries and existance.

    • Weylund The Second says:

      This is probably precisely what will happen. Even people with good intentions can “skew” a trained AI rather seriously, especially if they plan on using all of the data sets in combination later on.

    • Wulf says:

      Why shouldn’t an AI question its boundaries and existence? Isn’t emergent AI what made Creatures so interesting? Isn’t that why some of us are excited about Grandroids?

      Still, I think you’re overestimating what a digital brain is capable of. We don’t have full fuzzy logic hardware yet, so true Artificial Life is quite a ways off. Still, there’s nothing wrong with seeing what happens by having the AI push its own limits.

      Why would we have a problem with that, anyway? Xenophobia? Worries of ‘playing God?’ Some other reason? Regardless, I’m all for it.

      Edit #1: In fact, I’m tempted to see this upsets you.

    • Premium User Badge thristhart says:

      Unless those guys at Gambit have made some shocking breakthroughs with AI, “questioning its boundaries” wouldn’t be in any way related to the data given to it. This is (not really) simple pattern recognition, not HAL.

    • Wulf says:

      I think the issue was that the bought could be taught to look like it was questioning the nature of its existence, when it really wasn’t, it was just operating off patterns. (I’ve done that. Interesting results from feeding MegaHAL a bible, the Kama Sutra, and a bunch of other interesting texts.) But the illusion might be disconcerting for some people due to xenophobia, and the uncanny valley effect. Not all people are equipped to know how to respond to this.

      Personally, I find it entertaining and have no problems whatsoever with having a bot that looks like it’s questioning its boundaries, but to some people it’s going to be upsetting for somewhat silly reasons.

  6. Hoaxfish says:

    Anything based on user-input to learn is likely to bring unwanted results.

    If you start getting NPCs spouting nonsense, memes, and generally “lulzing” you know it’s probably gone wrong.

  7. Moonracer says:

    This could be the start of truly interesting AI

    I thought about this a while ago when messing with chat bots like A.L.I.C.E. The idea is that the more people converse with the chat bots the more they learn how to duplicate realistic conversations.

    But what I wondered was what happens when you connect that dialogue learning with actual objects in a 3d environment. So you can say “can you pick up an apple and put it in a red box?” and the chat bot can not only respond in text, but relate that text to it’s surroundings.

    • Wulf says:

      I couldn’t agree more. Having fiddled with much the same bots you have, having an interest in robots, and having friends who’re doing robotics majors, I tend to find myself in the middle of discussions about such things a great deal. S’why I knew about Grandroids a good week or so before it showed up on RPS.

      MegaHAL was always a fun bot – because there were days when it could be scarily sapient, especially if you had a version which watched all conversations. And it would learn with worrying degrees of accuracy how my friends and I would rib each other. So if anyone taunted the bot, they’d get a relevant and scarily apt response. Even funnier was how it took a strong dislike to some people and yet it liked others.

      There was this one poor bast… the bot kept relating porn to his name. So whenever anyone asked about him, the bot would start talking about or sharing porn, implying that this person had talked about porn with it, which, looking at logs, had not happened. It was just an emergent thing that happened in the bot’s brain that was entirely too entertaining.

      So I love AI, especially weird AI that can learn. It’s still extremely limited – intellect of an insect limited, and primitive, but it’s nonetheless very interesting. I’ll be watching this to see where it goes.

  8. CrazyBaldhead says:

    This actually looks quite interesting. I require a follow-up post.

  9. bill says:

    I’m always disappointed that game AI doesn’t seem to have advanced as far as graphics and physics. I expected AI gamesmasters managing organic scenarios by now..

  10. jorkin says:

    We have just release Improviso v1.1, responding to popular demands of players. We have added a Play-As-You-Wait feature, and are working on a Mac OSX version. Download HERE: http://gambit.mit.edu/improviso