Clyde: A Deep Reinforcement Learning DOOM Playing Agent

Dino Stephen Ratcliffe, Sam Devlin, Udo Kruschwitz, Luca Citi

Abstract: In this paper we present the use of deep reinforcement learn- ing techniques in the context of playing partially observable multi-agent 3D games. These techniques have traditionally been applied to fully observable 2D environments, or navi- gation tasks in 3D environments. We show the performance of Clyde in comparison to other competitors within the con- text of the ViZDOOM competition that saw 9 bots compete against each other in DOOM death matches. Clyde managed to achieve 3rd place in the ViZDOOM competition held at the IEEE Conference on Computational Intelligence and Games 2016. Clyde performed very well considering its relative sim- plicity and the fact that we deliberately avoided a high level of customisation to keep the algorithm generic.