The recent advances in deep neural networks have led to effectivevision-based reinforcement learning methods that have been employed to obtainhuman-level controllers in Atari 2600 games from pixel data. Atari 2600 games,however, do not resemble real-world tasks since they involve non-realistic 2Denvironments and the third-person perspective. Here, we propose a noveltest-bed platform for reinforcement learning research from raw visualinformation which employs the first-person perspective in a semi-realistic 3Dworld. The software, called ViZDoom, is based on the classical first-personshooter video game, Doom. It allows developing bots that play the game usingthe screen buffer. ViZDoom is lightweight, fast, and highly customizable via aconvenient mechanism of user scenarios. In the experimental part, we test theenvironment by trying to learn bots for two scenarios: a basic move-and-shoottask and a more complex maze-navigation problem. Using convolutional deepneural networks with Q-learning and experience replay, for both scenarios, wewere able to train competent bots, which exhibit human-like behaviors. Theresults confirm the utility of ViZDoom as an AI research platform and implythat visual reinforcement learning in 3D realistic first-person perspectiveenvironments is feasible.
展开▼