Scaling up Deep Reinforcement Learning for Intelligent Video Game Agents

We introduce the concept of Deep Reinforcement Learning (DRL) and describe the current state-of-the-art in subareas relevant to the author's research. We present previous and ongoing work done by the author in context of game engines, video game development and Machine Learning performance. We discuss our measurements showing the performance discrepancy between training DRL agents on game engines and end-to-end GPU-based physics simulators. We propose the use of external GPU-based physics simulators and transfer learning to accelerate the training of DRL models for game engines. As future work, we discuss the use of model decomposition in complex environments to further accelerate learning efficiency of DRL in addition to increased hardware utilization.