The machines are rising and will one day kill us all – one game at a time, as two students from Carnegi Mellon University decided to create an AI that can play DOOM like no one else.
NextBigFuture reports that this new killer bot (known as Arnold) uses a host of complex mechanisms to play DOOM, and to out-play any living human.
Guillaume Lample and Devendra Chaplot are the masterminds behind Arnold DOOM (as I’m gonna call it from now on). Old Arnold uses the same networking that allowed Google’s DeepMind to learn how to play Atari games.
Of course the 3D DOOM provides a very different challenge compared to a 2D Atari game. AI bots usually rely on information such as map data and variables that a human player wouldn’t have.
Impressively, Arnold was taught to play while only reacting to what was visible on screen at the time, thanks to a functional short term, long term memory to help it keep track of what it encounters.
But even after 50 hours of play, Arnold wasn’t making much progress (it often takes a long time for AI to learn and adapt). As such, its creators decided to use an API to access DOOM’s engine and help speed things along.
Arnold proved himself by taking second place at the Visual DOOM AI Competition. It managed to nab the highest kill to death ratio, but a bot with the less impressive name of F1 got more frags. Boo F1.
To those who might want to learn more about Arnold, you should check out Lample and Chaplot’s paper on the AI.
In the near future, when the robot overlords are exterminating our loved ones in front of us, it’ll be good to close our eyes and think back to little Arnold, the AI that just really liked DOOM.