Deep Reinforcement Learning with DQN vs. PPO in VizDoom

Abstract

VizDoom is a flexible and easy-to-use 3D reinforcement learning research platform based on the well-known Doom first-person shooter. The challenge is to create bots that compete in the DeathMatch track, making decisions based solely on visual in-formation from the screen. The paper offers a com-parison of different approaches with reinforcement learning: Q-learning and policy-gradient algorithms. We explore the distributed learning paradigm in re-inforcement learning, and also discuss the differences in speed and quality of convergence when adding an object detection module.