Slashdot Mirror

← Back to Stories (view on slashdot.org)

EA Created An AI That Taught Itself To Play Battlefield (kotaku.com)

Posted by msmash on from the moving-forward dept.

Electronic Arts' Search for Extraordinary Experiences (SEED) Division has created a "self-learning AI-agent" that has managed to teach itself how to play Battlefield 1 multiplayer. From a report: In this blog post, Magnus Nordin from SEED details how his team, inspired by Google's work with old Atari games, wondered "how much effort it would take to have a self-learning agent learn to play a modern and more complex first person AAA game like Battlefield." So they tried to find out. The results are an "agent" that, while inferior to human players, "is pretty proficient at the basic Battlefield gameplay." The agent changes behaviour if it's low on health or ammo, and while more complex behaviours like knowing the details of each map are beyond it (at the moment), EA has found that "while the human players outperformed the agents, it wasn't a complete blowout by any stretch."

59 comments

  1. AI's true calling by nonBORG · · Score: 0

    Funny that AI seems to really only able to play games. Is this all that it is good for, a clear set of rules, no real damage that they can do outside the game world. When it comes to driving cars they sometimes forget to brake for people (oh lost a point, who cares.)

    --
    You can't handle the truth! - Because I don't post left all my comments get modded down, bye bye Karma.
    1. Re:AI's true calling by Boronx · · Score: 1

      Probably, but it reduces the problem to refining the game until it's close enough to reality.

      --
      Play Command HQ online
    2. Re:AI's true calling by jma05 · · Score: 1

      > When it comes to driving cars they sometimes forget to brake for people

      So do people, with greater frequency.

      There will never be a perfect AI. Well-surpassing average, imperfect people is advance enough.

      Given that the new AI rush started about 6 years ago, where we stand is pretty impressive. Progress may stagnate hereafter, until the next leap, but the advances are undeniable and practical.

    3. Re:AI's true calling by Koby77 · · Score: 1

      It is my understanding that calculating astronomical events within the solar system is easier than predicting the weather on earth. Even though it may appear to some that calculating how events millions of miles away unfold ought to be more difficult than those nearby, astronomical events such as a solar eclipse actually involves fewer variables than a weather pattern. Similarly, the AI's ability to understand games may be because games have more limited sets of rules, while real life involves far more calculations and possibilities.

    4. Re:AI's true calling by Anonymous Coward · · Score: 0

      As children we all learn from games of various types so it seems like the perfect way to hone our own AI creations. We don't let children make decisions with grave consequences until they have learned enough (we'd hope) and similarly we will probably keep our AI's confined to environments that are, like you say, devoid of true consequences for now. But as they learn, and as we learn to make them better, eventually there will be a point where we trust them to be self reliant and trust worthy for certain tasks with real consequences. Just as you trust your child at a certain age to be trusted with greater and greater responsibilities.

      Personally I think that even though it may take a while, there isn't any limit to the capacity AI could achieve in terms of sentience and raw intelligence, whereas we humans are generally limited in our mental capabilities. Maybe that's what scares Musk and the others so much.

    5. Re:AI's true calling by q_e_t · · Score: 1

      It's about creating strategies to work well in increasingly complex situations. Solutions can then use this to develop AI more able to navigate the real world. Using a game is easier to control and accelerate training for, and develop required techniques faster.

  2. AI you say? by Tsolias · · Score: 1

    I am waiting for Elon Musk, Stephen Hawking, e.t.c. to place statements about A.I.

    1. Re:AI you say? by Anonymous Coward · · Score: 0

      Hawking's statements would be really remarkable to read indeed.

    2. Re:AI you say? by Anonymous Coward · · Score: 0

      Well if you are waiting for Stephen Hawking to comment on this, it's going to be long wait...

    3. Re:AI you say? by Anonymous Coward · · Score: 0

      If you're waiting on Stephen Hawing to make a statement, I'd encourage you to come out from under your rock first.

  3. The answer? by ckatko · · Score: 2

    Not that hard. Because they're not trying to mimic human behavior, they're simply trying to win a known system.

    This is nothing but a PR puff piece to market EA. Actual people have been doing "AI plays game" for years upon years. YouTube has TONS of channels dedicated to just that.

    1. Re:The answer? by Kjella · · Score: 2

      Not that hard. Because they're not trying to mimic human behavior, they're simply trying to win a known system.

      While I agree this article is a puff piece, why does it matter if it wins the "human" way as long as it wins a legal way? While it's possible that some aimbot-like characteristics are far easier for a computer to achieve than an human, it's not cheating if it's man vs machine. It's only cheating if you're pretending it's man vs man. I've played some chess computers playing very "unhuman" chess, but I don't think their victory is any less valid because of that.

      --
      Live today, because you never know what tomorrow brings
    2. Re:The answer? by Anonymous Coward · · Score: 0

      Agreed. 4 years ago in my Intro to AI course at college we created self-taught Unreal Tournament bots. The AI concepts behind these things are very simple. It just takes a long, long time to train them if you don't setup good goals or well defined actions for the AIs. There are numerous ways you can create the bots: GAs, different types of NNs, GOAP, decision trees, etc... They all have their pros and cons, but they also all work. It seems EA is now getting its head out of its ass and is looking at something new. The research papers in this area are already beyond this, so EA still isn't looking at the best that can be done, only what's being marketed. That tells me either their AI developers are all crap or EA management is holding those guys back. I wouldn't expect management to read AI research, but I would expect any good AI developers to do so.

      Mimicking human behavior is easier. Break each human's actions down into a mapping of state-of-area-around-player to list-of-probabilities-of-human-input-at-time+1 and each bot would function based on a map lookup. The humans automatically do all the training simply while testing the game and the AIs would improve as the humans improve. You can do it a lot fancier, but this way is easy to program and should give good enough responses. Defining what your state entails is the hardest part. This is a case-based approach with training data provided by humans instead of drying to define goals and letting the AIs play against each other. If you have the hardware, letting this type of AI randomly play against itself could get you better than human results.

      I would have gone into game AI development if the industry wasn't considered a hellhole to work in when I was graduating. Instead I waste away writing business logic and do some FOSS game dev on the side.

    3. Re:The answer? by angel'o'sphere · · Score: 1

      Most old school games don't need aim bots, well people used them anyway.

      E.g. the size of all humans in the game are the same.

      So if you hear one coming around the corner, you exactly know where his head will be. If he is crouching you hear that too, and again you know exactly where his head will be when he comes around the corner.

      You just have to practice a bit for your preferred distance to the corner and you can intuitively head shot everyone coming around it. No aim bot needed at all ... but you likely get kicked from the game by the other players because they accuse you to use an aim bot.

      --
      Cost free eBook I read (by iBook/Kobo/Amazon/ObookO/Gutenberg etc.): "The Green Odyssey" by Philip Jose Farmer.
    4. Re: The answer? by Anonymous Coward · · Score: 0

      Wtf are you on about. The size of the humans in battlefield 1 are the same size too and theyre not old school. Most army games have the same size humans. Its only really sci fi games with odd shaped creatures where the hit boxes are different.

    5. Re:The answer? by fazig · · Score: 1

      If I remember correctly, even the first Battlefield game had at least some simplistic ballistic simulation, which made it different from your old school first person shooters. What you describe may work for games like Quake or even the newest iteration of Counter Strike, where distances are a lot more predicable and you don't have ballistic simulation. In those 'old school games' you usually also can fire from your hip and have the first shot hit exactly where your augmented reality cross-hairs point towards, because you know, the projectile (actually hitscan) doesn't come out of the gun's barrel and converges with the cross-hairs at a certain distance, it comes out of your virtual eyes.
      But in games that try to be at least a bit more realistic in certain aspects it usually don't works that reliably. The same applies to close quarters because of all the different stances a player may be in that require you to adjust your aim. Take ArmA 3 as an example, where the size of all player is also the same. Perhaps try it with the ACE3 mod. There you get a lot of factors in the ballistics simulation - obviously gravity, bullet mass, and muzzle velocity; dynamic wind which has somewhat realistic fluid dynamics; air pressure; temperature; humidity; Coriolis and Eötvös effect.

    6. Re:The answer? by Anonymous Coward · · Score: 0

      Ah but did it learn to shout your mom insults and call swat?

  4. Seems unimpressive. by PeterGM · · Score: 1

    I programmed an AI when I was at university over a decade ago that was a foul mouthed chatbot. Seems like it's the same thing except it's better at teaching itself new words.

    I'm 99% sure that the majority of game play in Battlefield is actually just 12 year old's screaming at each other. With an Aimbot.

    --
    There are no stupid questions, just stupid people.
    1. Re:Seems unimpressive. by Anonymous Coward · · Score: 0

      While the general battlefield online experience can sometimes be boiled down to 12 year old foul mouthed hackers, the actual core gameplay mechanics provide a vast array of challenges for players and AI alike. It seems so simple to us to sit down and play a game but for an AI there are layers upon layers of different things it has to analyze and make sense of. Just making sense of the video information and creating an understanding of the 3D environment being represented in 2D fashion is difficult on its own. This AI also has to understand what enemy and friendlies look like, what a gun is, what ammo is, etc etc. More importantly the AI had to understand how to play the game with all these layers, now just know how to use each one on its own. What becomes really interesting is when the AI starts to understand the player mentality and can react from it. Right now its limited to point and shoot and maybe some audio cues but at a certain point it will begin to understand the other player's intent and can strategize around that information.

    2. Re:Seems unimpressive. by yuriklastalov · · Score: 1

      Your chat bot was probably more cogent than 90% of the human players.

      Probably didn't call people fags nearly as often, anyway.

    3. Re:Seems unimpressive. by Anonymous Coward · · Score: 0

      That's not how AIs work. There is no intent nor understanding. What they have is massive lookup tables which compare what the local state looks like and a probability table which specifics what actions are likely best in this state based on past training. The AI has no understanding of intent or long term strategy even if it looks like it has. It's only following patterns in the training data. If the training data is good, then those patterns will match ones with good strategy, but the AI doesn't understand that. It's blindly rolling dice and picking actions out of predetermined sets of actions. You could argue GOAP AIs have intent, but they still have no understanding as you understand and they have even less awareness of others' intentions.

      Even then, zero AIs outside of research read what the computer displays. Instead they are embedded in the game so they already know which players are who, their ammo count, health, etc... They don't need to discover those things. They are known facts in their world.

    4. Re:Seems unimpressive. by Anonymous Coward · · Score: 0

      fag

  5. Machine learning by JThundley · · Score: 2

    There was that recent story about video games and machine learning and how long it takes humans and AIs to learn to play a custom video game. One of the conclusions they came to was that humans learned their custom video game faster because of societal queues that they already know from outside of the video game. For example, they saw a man and assumed that was their character, saw a ladder and assumed they had to walk over and press up, jump across gaps, jump over what must be bad guys since they have angry faces. Their machine learning bot took a lot longer to learn the game since it was trying to figure out a lot more details about the game than the humans did.

    This got me thinking: why hasn't anyone created a more generic AI that learns how to play *TONS* of our old video games? Start them off with older games and work them up to newer and newer ones. This way it would carry all those past experiences and draw upon them when faced with a new and unfamiliar video game. It would have knowledge of the meta of video games. I think this would be awesome, but scary at the same time. Could you imagine an AI that learned everything it knows about humans through video games only?

    1. Re:Machine learning by KClaisse · · Score: 3, Interesting

      In a lot of ways you have hit at the crux of the AI problem. Another prominent example of this issue is with Automotive AI. We humans bring so much more information to the table besides the basics of how to drive a car and what a road/signs are. We have an ingrained knowledge of what another human being is going to do. We pick up subtle cues from the "body language" of other driver's and use that to inform decisions. That decision making process was created and honed in us as children when we had to understand the intent of our fellow humans. This array of knowledge informs everything we do in life and is, to use your own terms, a kind of "generic AI' that we all share. Its an interesting problem and begs the question: Is this kind of generic AI possible to create without also creating what is essentially a basic consiousness? After all, it is this basic knowledge and understanding that makes us thinking, aware creatures. Its quite a fantastic time to be alive, I can't wait to see what happens next.

    2. Re:Machine learning by Anonymous Coward · · Score: 0

      Memory isn't well defined in modern AIs. If you train on one game, it will forget how to play that game when you train it on another. You can abstract out all the games to common elements, but then you're at there's a way to move, there's a score, and there's an ending. Most of that is already pre-programmed by the programmers because the AI has to be linked to those controls inorder to use them.

      For what you're asking, the AIs would need to understand objects. They would need to understand what a ladder is in all its forms and how one uses it. When you get an AI to train itself like that, you'll have a strong AI and it'll either kill you, love you, or ignore you.

      These current AIs operate on sets of pixels, not abstract organisations of them. Colors changes will break them. Programming them to work on shapes of objects and transformations of those objects is far more time consuming. Doing that, then being able to train it in a reasonable time means you'll have a state of the art AI framework able to complete most tasks. Maybe it'll be a strong AI if it can modify itself and change goals, maybe not.

    3. Re:Machine learning by JThundley · · Score: 2

      Totally, it's all very interesting. Technology advancements are one of the few things that keep me interested in the future.

      You bring up another daydream that I have sometimes. Just like we're talking about raising AI, I think about raising an AI like a baby. I'd imagine to do this, you'd have to give it some kind of a body to operate and sensors to process input from the world. Try to raise it as a person first, and then later try to teach it more advanced things once it sufficiently knows how to act human. Of course, the big problem with this is biology. Humans don't exactly learn all things they know. Some things are intuitive and ingrained in our DNA. We seek out patterns, recognize faces, and feel at least some empathy naturally.

      Another train of thought that comes from this is a generic science AI bot. I wonder if something like this could be created with the sole purpose making scientific discoveries. Teach it everything we know about physics and different scientific disciplines, and have it generate experiments that it can simulate and then we can carry out. Maybe I'll see FTL travel in my lifetime!

    4. Re:Machine learning by Anonymous Coward · · Score: 1

      You may find interesting: http://blogs.sciencemag.org/pi...

    5. Re:Machine learning by q_e_t · · Score: 1

      societal queues

      I've met people who could benefit from playing a few rounds of Queue Simulator.

    6. Re:Machine learning by ceoyoyo · · Score: 1

      Deep Mind did exactly that.

    7. Re:Machine learning by JThundley · · Score: 1

      Cues! Cues! Fuck, you got me.

    8. Re:Machine learning by JThundley · · Score: 1

      Looks like they just had it playing Atari games. I want to see it winning Starcraft tournaments already!

    9. Re:Machine learning by ceoyoyo · · Score: 1

      https://deepmind.com/blog/deep...

    10. Re:Machine learning by JThundley · · Score: 1

      I know they've been working on this for a while in both Brood War and SC2, but the bots still aren't good yet.

      Our initial investigations show that our agents perform well on these mini-games. But when it comes to the full game, even strong baseline agents, such as A3C, cannot win a single game against even the easiest built-in AI. For instance, the following video shows an early-stage training agent (left) which fails to keep its workers mining, a task that humans find trivial. After training (right), the agents perform more meaningful actions, but if they are to be competitive, we will need further breakthroughs in deep RL and related areas.

  6. Now all EA need to do by Anonymous Coward · · Score: 0

    Now all EA need to do Is teach the AI to buy lootboxes so it can go fuck itself.
    (Oh and AI is code for simple highly integrated turds)

  7. AI as commander ? by Archfeld · · Score: 1

    What would be really cool is an AI that play a decent commander spot. B2 is a great game I still love it when a good squad gets together, but playing minus a good commander is not nearly as fun, but playing commander is to be honest sort of boring. An active AI commander capable of meeting the needs of squad requests, following a strategy and assisting in the taking of command points could add a HUGE amount of enjoyment to the game. I know many people who play a good commander but it gets boring after a few rounds, and the side without a decent commander is crippled.

    --
    errr....umm...*whooosh* *whoosh* Is this thing on ?
    1. Re:AI as commander ? by angel'o'sphere · · Score: 1

      I used to play Eve Online.
      A good commander is more worth than a hundred troops, or thousand even. The biggest battles I have been in where about 3000 of our own forces, no idea about the enemy.
      You would be surprised how many of the "good commanders" are Russians. I only remember one American, but he was a Russian immigrate, Makala or something was his name (fleet commander of tripple A).

      Ha ... could now write a long rant about the different ways of how "typical americans" versus "typical russians" versus "typical british" do their commanding ... but I guess it would bore most of the people.

      But I miss the voice of Ekatarina, Titan pilot, most female angel like voice I ever have heard, a Russian Lady ofc.

      --
      Cost free eBook I read (by iBook/Kobo/Amazon/ObookO/Gutenberg etc.): "The Green Odyssey" by Philip Jose Farmer.
    2. Re:AI as commander ? by Anonymous Coward · · Score: 0

      Well, a good commander is certainly worth a lot, but not more than a hundreds or thousands of troops in that particular game. I mean it's not reality where we know that some smart guerilla warfare can deal a lot of damage to a much larger power and maybe even scare them into a retreat.
      At least when I still played EVE Goonswarm and perhaps TEST were the perfect examples for this. They could muster up so many players that they didn't really need something like good target calling. Just shoot at everything that's within your range in an alphabetical order. When you can outnumber your opponents by 10:1 there a very good probability that enough people will be firing at the same target to beat all the RNG. It's similar to those live show TV illusionists or magicians that claim they can fix your wrist watch if you follow their instructions and really believe in it. Odds are that among the thousands of viewers there will be a small number whose watches will resume function. Then they call the show and proclaim that it has worked. And suddenly miracles are possible!
      Sorry for the rant. I don't really care about the politics in that game any more and what alliances people are part of, but when I look back I still think about the game mechanics themselves. Maybe things have changed a lot since then (8 years ago). Tactics and discipline certainly do matter in that game (I was PL), but at the end of the day EVE was and probably still is a numbers game.

    3. Re:AI as commander ? by Archfeld · · Score: 1

      In battlefield, obviously the numbers are much smaller, but a single squad with a good officer under the direction or a good commander can easily swing the game. Conversely a squad with out direction and supplies will generally flail about without success. I never played EVE online so I can't really attest or relate but I did read some about the game.

      --
      errr....umm...*whooosh* *whoosh* Is this thing on ?
    4. Re:AI as commander ? by angel'o'sphere · · Score: 1

      Biggest online game with concurrent players online in the same world (not split into multiple "worlds" or "shards" or "servers")
      Unfortunately much to many "unintuitive" game mechanics.

      --
      Cost free eBook I read (by iBook/Kobo/Amazon/ObookO/Gutenberg etc.): "The Green Odyssey" by Philip Jose Farmer.
  8. AI Teaches EA How To Make Actually Good Games by dryriver · · Score: 2

    EA executives, horrified, immediately unplug the AI. EA is saved.

    --
    Why did the chicken cross the road? Because Elon Musk put an AI chip in its head.
    1. Re:AI Teaches EA How To Make Actually Good Games by Anonymous Coward · · Score: 0

      "The first 47 iterations all tried to improve our games. Finally we got one whose only goal was murdering humanity, let it run rampant on the internet and put it in charge of loot-boxes"

  9. The rest of the story by 93+Escort+Wagon · · Score: 1

    Unfortunately, after playing for a few hours a licensing glitch caused the AI to be misidentified as a game pirate. The system was blacklisted from all Battlefield 1 servers; and then, once EA's License Enforcement team discovered the IP address was from within the company, they forcibly powered-down the system - leading to the complete loss of the AI.

    --
    #DeleteChrome
  10. It's a beginning by Impy+the+Impiuos+Imp · · Score: 1

    In mmo games they need to do away with taunt as it is built around terrible AI rather than a human DM. Bury it and the "tank" concept and make weaker challenges that fight a hell of a lot better.

    --
    (-1: Post disagrees with my already-settled worldview) is not a valid mod option.
    1. Re:It's a beginning by Anonymous Coward · · Score: 0

      The Log Horizon TTRPG has an interesting hybrid mechanic for aggro. Everyone not tanking can easily produce far more than the tank, but the ability to 'force' targeting on themselves is quite limited. Rather, an actively participating tank interferes and penalizes targeting things below it on the aggro list, can punish going after something else even if it's higher, and forcefully 'steal' the aggro of others, though that's a mental compulsion that could miss or be resisted.

      However that same aggro that's building-up serves as a damage buff (basically acting as a gradual enrage timer), so forcefully catapulting yourself ahead of everyone makes things a lot more dangerous all-around and risks overwhelming the healers.

  11. Dumb automation cannot "teach itself" by gweihir · · Score: 0

    And dumb automation is all that exists at this time. At best, the statistical classifier used did supervised learning, i.e. somebody with actual intelligence defined a fitness function and then the classifier trained on that function. As basically all intelligence is in that fitness function, the classifier has none and it did not "learn" anything either (because that would require some sort of understanding). It simply conditioned its "reflexes" (if that is not already overstating the amount of "intelligence" involved) on that fitness function in a scenario again created by somebody with actual intelligence.

    --
    Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
  12. They do look like bots... by kangsterizer · · Score: 1

    Looking at the video, specially in first person mode they look exactly like bots.
    Was hoping for something more.
    They follow walls and snap on target basically. Which also creates the bug when each time is being another wall and it loops. They either see through or hear through precisely and don't try to jump over or set explosives on the walls. That's quite basic.

  13. An AI to play... by Anonymous Coward · · Score: 0

    ... ainâ(TM)t that called a âoebotâ? Like the bots we have on Counter-Strike, DotA, Heroes of the Storm and another gazzillion other games?
    How is this even a thing for EA?

  14. Hmmm by Anonymous Coward · · Score: 0

    I know this isn't very productive at all, but wouldn't it be cool to have a couple AIs like that to serve as teammates in multiplayer games?

  15. bots by n3r0.m4dski11z · · Score: 1

    I'm pretty sure battlefield 1942 had bots in it. https://www.youtube.com/watch?...

    Yep. Good job EA, you re invented something you already invented 20 years ago.

    --
    -
  16. Strange game. by Kargan · · Score: 2

    The only winning move is not to play.

    --
    Palaces, barricades, threats, meet promises
    1. Re:Strange game. by ohnonononono · · Score: 0

      fuck

    2. Re:Strange game. by Anonymous Coward · · Score: 0

      You must be fun at parties.

  17. AI noob by romit_icarus · · Score: 1

    I would categorise the creation of the first gaming noob as a significant milestone in artificial intelligence

  18. Teaching AI Warfare by Anonymous Coward · · Score: 0

    What could possibly go wrong ?

  19. Has it learned to tea-bag? by Anonymous Coward · · Score: 0

    nt

  20. upgrades by Anonymous Coward · · Score: 0

    Does the AI have to pay for upgrades in EA games?

  21. They already posted this?! by ckatko · · Score: 1

    How much do you guys get paid to post this marketting crap?

    There is zero "specialness" about a huge company, having some guy employ basic neural nets. The difficulty is not "can AI learn" it's "can AI act like a human" and that's the hard part. THAT is why Quake 3 had ZERO neural nets in their AI.

    This is nothing but a PR blurb.

    There are tons of actual AI videos (and channels!) on YouTube of people teaching AI to play 2-D and 3-D games from scratch. Post those here. But they wont... because nobody is getting paid off.