You know how sometimes you look at a piece of research and think, I suppose its an interesting technical problem, but isnt teaching an AI to hunt and kill humans a pursuit fundamentally dangerous to the continued existence of mankind?
This is one of those times.
Some traitors to the species at Carnegie Mellon have applied the ever-applicable neural network approach to create an AI that is literally a killing machine. Or perhaps I should say fragging machine, becauseits only doing the killing in a Doom deathmatch. Now, you can of course say it doesnt count because the death is virtual, but why should that matter? To computers, everything is virtual. Think about it.
You have a question: How does this differ from the bots weve had in games since forever? Computer players have been around forever!
You are an acute observer, reader, but consider: Those bots are programs running within the game itself, aware of all the variables, coordinates, edges, the locations and specs of guns and health kits. Like any non-player character, they are programmed to act in certain ways in reaction to certain in-game variables.
The AI created by Guillaume Lample and Devendra Singh Chaplot plays the game the way we humans play: by looking at the screen, identifying our characters situation and orientation, finding our way around the map and shooting at anything that moves. Its essentially a level up from the AIs that use similar methods to learn the methods governing simpler games, like Space Invaders, and find the inputs that maximize score in those.
Here it is in action:
The network was trained mostly on pixel data that is, what you actually see on the screen but its creators had to cheat a little by giving it some basic insight from the game engine on whether there was an enemy or item on screen.
Its reinforcement strategy was this: It got attaboys for picking up items, moving a lot and racking up kills, but was reprimanded for taking damage and dying. There was also a light rap on the wrist for shooting, since otherwise the machine decided firing indiscriminately and waiting for enemies to wander into its crosshairs was the best technique. Reassuring!
The system that resulted from this setup outperformed the in-game computer players and humans alike. The former werent exactly advanced when Doom came out, so theyre pretty much cannon fodder.
Its actually split into two systems, or lobes if you will. Theres a navigation side, which drives while moving around and collecting things, and presumably learned how to interpret environmental imagery. Then theres a shooting side, which takes over when theres an enemy on screen, pointing the gun at the right place and pulling the trigger.
Are we witnessing the birth of Skynet? Or, perhaps, a startup that will soon pitch us with a universal adversarial AI for multiplayer games? Both are pretty scary.
Lample and Chaplots paper describing the FPS-playing AI is available for free on Arxiv.
The post Scientists teach machines to hunt and kill humans in Doom deathmatch mode appeared first on Machines.
source http://www.millermachine.net/scientists-teach-machines-to-hunt-and-kill-humans-in-doom-deathmatch-mode/
No comments:
Post a Comment