Doom Bot Learns To Play Better Than Humans

Two Carnegie Mellon University computer science students have created the ultimate killing machine…if you face off against their AI in Doom’s deathmatch, that is.

As originally reported by NextBigFuture, this new bot (affectionately named Arnold) uses a host of complex mechanisms to play Doom better than any human possibly can. The students, Devendra Chaplot and Guillaume Lample, used deep-learning techniques to train the AI agent to negotiate the game’s 3-D environment, still challenging after more than two decades because players must act based only on the portion of the game visible on the screen.

Their work follows the groundbreaking work of Google’s DeepMind, which used deep-learning methods to master two-dimensional Atari 2600 videogames and, earlier this year, defeat a world-class professional player in the board game Go. In contrast to the limited information provided in Doom, both Atari and Go give players a view of the entire playing field.

But the jump from 2-D to 3-D provided unique challenges. Traditional AI bots rely on information that players can’t possibly have such as map data and other in game variables. Lample and Chaplot aimed to teach Arnold how to play while only reacting to what was visible on screen. It’s a bit complicated but the bot has a functional short term, long term memory to help it keep track of what it encounters.

Still, it takes a long time for AI to learn how to play games. Bots that iteratively learn through neuro-evolution can take days to make significant progress. A bot learning to play a single level of Super Mario World took 24 hours. Arnold made little progress after an initial 50 hours of play. His creators had to use an API to access Doom’s engine to help speed along the learning process.

Not only is Arnold a fast and an accurate shot, but it has also learned to dodge shots, making it hard to kill. Though Arnold placed second in both tracks of the VizDoom competition, it had the lowest number of deaths and the best kill-to-death ratio by a significant margin. In track one, the agents navigated with a map and only one weapon; in track two, they navigated without a map and with multiple weapons.

Arnold took second place at the Visual Doom AI Competition, even though it did manage to have the highest kill to death ratio. A bot named F1 managed to get more frags.

If you’re interested in the particulars, you can also read Lample and Chaplot’s paper on the AI. Me? I’ll be cowering in a corner, hoping that fragging enemies in Doom will be enough to state Arnold and his deadly skills.