AI trained to play Doom Deathmatch has caused some controversy

[Doomkid]

http://www.sciencealert.com/controversial-ai-has-been-trained-to-kill-humans-in-a-doom-deathmatch

I wonder what this will mean for the future of AI? Very interesting stuff.

EDIT: The included YouTube video is very strange - A constantly crouching Doomguy unable to strafe killing what appears to be a bunch of zombiemen that have been reskinned as Doom marines, not actual bots or players. All the while the footage has been sped up to 1.5x speed for no discernible reason. It would be interesting to see a video against actual Doom bots or even better, actual human players. Still very impressive that it's all done based on only pixel data.

[Job]

Typical alarmist garbage and late to the game (metaphorically) to boot. What about the bots in Quake, Unreal, etc? Besides, I thought the most recent Doom deathmatch AI competition didn't involve human players.

[Unf]

This is actually pretty cool! Now we might even have good deathmatch bots or even coop bots that actually kind of know what theyre doing

[Doomkid]

Job said:

Typical alarmist garbage and late to the game (metaphorically) to boot.

I think all the alarmist stuff is absurd, to be clear - I just think the whole thing is quite impressive!

[wheresthebeef]

I want to see several different groups make up their own AI and put them all in a deathmatch and see which one wins.
Think Battlebots but with shotguns and BFGs

[40oz]

^ That's exactly what happened, dummy

[Job]

My money was on the one shaking in the corner.

[Mithran Denizen]

Just like controversial chess AI are being trained for world domination. Uh-huh. *snore*

[Doomkid]

Thanks for posting that video 40, the one included with the article wasn't nearly as interesting.

I would love to see this become a standalone .exe that could be run in any Doom port to give you a co-op AI buddy or an intelligent(ish) deathmatch opponent!

[Job]

I'm wondering if there will be subsequent Visual Doom AI Competitions. It'd be interesting to see how the bots improve as their teams tweak the programming based on the tournament results. Eventually, I would imagine some damn fine AI multiplayer bots would rise to the top for community use.

[HorrorMovieRei]

I bet this AI couldn't beat Sunder.

[Job]

HorrorMovieGuy said:

I bet this AI couldn't beat Sunder Sunlust.

Probably not.

[kb1]

Doomkid said:

I think all the alarmist stuff is absurd, to be clear - I just think the whole thing is quite impressive!

It is impressive. The scary part comes when you put the software on a laptop with a USB camera, place the laptop in a self-driving car with a turret on the roof :( Yikes.

[Memfis]

Doomkid said:

Still very impressive that it's all done based on only pixel data.

Just in case some people don't know, these bots don't just analyze the picture that we all see, they also have access to some depth buffer stuff or whatever (for example they can see it like this http://vizdoom.cs.put.edu.pl/user/pages/01.home/depthbuffer.png which is a lot more convenient for analysis I guess).

[SaladBadger]

Job said:

Typical alarmist garbage and late to the game (metaphorically) to boot. What about the bots in Quake, Unreal, etc?

worth noting, to some degree I guess, those bots are developed in a much different way and are explicitly instructed on their environment through nodegraphs, nav meshes, and the like. In addition they can look right at the game's world to see where things are (though they usually also have a layer of simulation to cause them to ignore things they wouldn't be able to know about)

I do find this new branch of actually trying to processing the world visually to be pretty interesting, though as mentioned by Memfis they do have a little more information than just the screen (the depth buffer, which probably helps working out the geometry quite a lot)

[Doomkid]

kb1 said:

It is impressive. The scary part comes when you put the software on a laptop with a USB camera, place the laptop in a self-driving car with a turret on the roof :( Yikes.

I don't think we're quite there yet, but we're certainly not too far off. This technology is, long term, absolutely going to end up being used for weaponry regardless of how many bans the UN puts on lethal autonomous weapons systems. Using such discoveries to engineer more efficient weapons is a consistent trend throughout human history.

The bit that gets me is the "if ported into the real world, it wouldn't be satisfied with a single kill" statement. These kinds of claims assume they're planning on building a death machine that they can't deactivate - That kind of stuff sounds more like silly fear mongering than a legitimate concern.

[MrGlide]

na, we're there, you could make and automated turret that tracks infered with a kinnect. As for the op, the articals are getting riddiculkus. The bot is mainly using googles deepmind so thats neat. Whatch deepmind play mario

I remember hearing about this quite some time ago, one of the mario deepmind bots got insainly good at mario and started to learn bizaar tricks, the same bot/ai I speak of had a very hard time playing tetris apperently because it would just smash all the blocks to the ground to end the game.

[Arctangent]

Wow that Mario video trips me up. SMB3 Mario, brick and ? blocks, and background, with SMW everything else ... including empty blocks. And Koopas that get stomped through their shells, and Galoombas that can be stomped ... yeesh.

[MrGlide]

I think your being sarcastic, this was just one of the quickest cleanist examples I could find, the bot/ai I spoke of was playing the original mario If I remember correctly, I'll google around when I get home to find it/ or the artical I originally read about it, it was quite some time ago.

[HavoX]

HorrorMovieGuy said:

I bet this AI couldn't beat Sunder Sunlust Ancient Aliens.

Nope

[Spectre01]

The AI would get to Precarious in Sunder and repeatedly fail the platforming.

[Arctangent]

MrGlide said:

I think your being sarcastic, this was just one of the quickest cleanist examples I could find, the bot/ai I spoke of was playing the original mario If I remember correctly, I'll google around when I get home to find it/ or the artical I originally read about it, it was quite some time ago.

I'm not really sure how any of that comes across as sarcastic. Style mixing has always been a massive peeve of mine with fangames, plus a lot of the enemy behavior in that is just wrong.

They could've just used Goombas! You can stomp on those!

[Fonze]

I just want to know why the bot in the video's shots were so low on walls and targets. But that article is absolute greatness: this computer was rewarded for killing people; since computers can totally think for themselves, it will act like an animal and continue to do the action required for the reward it so craves.

On a more serious note, what positive reinforcement, or negative, for that matter, would work on a computer AI? What do they want?! Lol

I wanted to find the clip from the What If machine where bender was a 100 ft tall robot, but this'll have to do:

[Arctangent]

Fonze said:

On a more serious note, what positive reinforcement, or negative, for that matter, would work on a computer AI? What do they want?! Lol

Well, I mean, that's kind of how learning AI works. When the AI is successful, you give it positive feedback. If it's not, you give it negative feedback. This give it actual metrics to analyze its data on - if it screwed up, it knows because it got negative feedback, which allows it to compare it to the actions that got it positive feedback and try to figure out the differences between them.

Sure, it's not positive feedback in the sense that you give your roomba a tasty treat when it's good nor negative feedback in the sense that it gets a hammer to the body if it does something bad, but that's still the basis of how the system works.

[Fonze]

Good point; they want a ++ in a particular line of code, heh. I still gotta wonder about that hitscan positioning, though.

*Edit*

Nm, watched 40's video, he's always crouching... kek

[Inkie]

MrGlide said:

the same bot/ai I speak of had a very hard time playing tetris apperently because it would just smash all the blocks to the ground to end the game.

I heard of a tetris AI that somehow managed to learn how to pause the game in order to prevent the final block from causing a game over.

Indeed, the winning move was to not play.

[fraggle]

Doomkid said:

EDIT: The included YouTube video is very strange - A constantly crouching Doomguy

I noticed this, and thought it was odd too. I wondered if they maybe adjusted it to exactly half the player height to get symmetry. The bot works off the visual image of the game so maybe it works better that way? There might be corner cases where it doesn't behave as intended if the camera is too close to the ceiling.

[Jaxxoon R]

They also might've assumed that having him duck would shorten his height and make him harder to hit.

[Doomkid]

Interesting, one of the guys actually responded to a question I had.

Doomkid said:

I'd like to ask why the video was sped up?

Guillaume Lample said:

During the training of the AI, the game API is actually going much faster than that (3 or 4min in real time is almost 1 hour in the game). The game is accelerated so that the AI is faster to train. When we visualize it, it's going too fast and it is impossible to see what's going on. In order to see something, we have to slow down the game by running some "time.sleep(ms)" command to pause the game for a very short time after every action executed by the neural network. If this still looks accelerated it is because we didn't pause it for a long enough time.

[Maes]

Doomkid said:

Still very impressive that it's all done based on only pixel data.

If (at least) that part is true, then this should really be judged from the point of view of a robotic/computer vision problem, and even that way this isn't state-of-the-art work.

An "A.I." that had perfect access to all relevant data (like bots in modern MP ports do) and still played like that would be nothing to write home about.

In this case, there's a problem of imperfect information that can only be partially inferred from visual data (like a human player has to) and memorization (in this respect, an A.I. might be superior).