Carnegie Mellon Students Build Doom Bot Based on Google Deep Q-Network

Two Carnegie Mellon students created a Doom AI agent that is capable of beating human players, as well as built-in AI agents, in the classic Doom computer game. Devendra Chaplot and Guillaume Lample used Google's DeepMind deep-learning technology to develop their Doom-playing bot, which they have nicknamed “Arnold.”

According to the story at the CMU website, Doom is harder than other games that are the focus of AI research because the player can only see part of the playing field. The game exists in a 3D world but plays out on a 2D screen. The built-in AI agents included with Doom cheat by consulting maps and other background data to compete. Chaplot and Lample's bot, on the other hand, uses visual information that would be available to a human.

Although other Doom bots exist in the world (Arnold actually took second in a recent world-wide competition), the Carnegie Mellon project is attracting attention as an application of a Deep Q-Network (DQN), part of the DeepMind platform, which is billed as Google's answer to IBM Watson and other similar tools.