Application of Deep Neural Networks for Decision Making in First-person Shooter

Reinforecement learning is the promising machine learning area that aims to solve hard real-world control tasks, through automatic learning from experience and rewards. This promise, though, comes with the unique challenges from the algorithmic and optimization perspective, often very different from the ones faced by supervised learning techniques. Moreover, the training procedure now involves multiple components such as environments, agents and learners. Choosing the efficient architecture for training this components in tandem while ensuring the best software and hardware performance can be difficult. In this work we develop one such architecture that is able to leverage hybrid CPU/GPU systems to train a reinforcement learning agent based on Q-learning. We compare our implementation with DQN on benchmark tasks from Atari and VizDoom simulator and show that it achieves faster training speed and better hardware utilization.