Transformer-based Deep Reinforcement Learning in VizDoom

Abstract

Transformers is a novel neural network architecture that is successfully used in natural language processing tasks and is starting to be used in other areas such as video processing and image processing. However, transformers are yet to be studied in different aspects of reinforcement learning scenarios. In this work we combine transformer architectures with reinforcement learning and train them in the VizDoom game environment, producing agents that play better in comparison to traditional neural network architectures.