Two Carnegie Mellon University computer science students have created an advanced AI in Doom that uses the same visual information as a human player – reading only what it can ‘see’, instead of cheating like a bot – but will still probably kick your arse in a deathmatch.
Seems like we need to brush up on the classics with our list of best old games, before the robots kill us all.
Deep reinforcement learning has recently seen AI perform well in Atari games by reading only raw pixel data, but this experiment is one of the first that’s seen the same techniques used for the AI to tackle a 3D environment.
The students cheated a little, using an API during the training phase to teach the AI level layout and items, but this is only the same as how a human becomes accustomed to in-game arenas through repeated play. When actually playing, it only uses the same visual information as you would.
In the impressive video above, you can see it facing off against bots in a deathmatch. Now these bots are traditional ones, and they can access much more information than what’s displayed in the pixels, but still the AI blasts them into dust.
“The fact that their bot could actually compete with average human beings is impressive,” said Ruslan Salakhutdinov, an associate professor of machine learning who was not involved in the student project. “Simply navigating a 3-D world, much less competing successfully in this game environment, is a challenge for such AI agents.”