A Survey on Deep Recurrent Q Networks

Reinforcement learning (RL), one of the branches of machine learning, enables a system to learn through trial and error. RL helps in solving control and decision-making tasks. Applying Deep Learning to Reinforcement learning has made it much better at solving many problems. Deep Reinforcement learning, a combination of deep learning and Reinforcement Learning is gaining a lot of interest and application in solving real-world problems. Among Deep Reinforcement learning, Deep Q networks emerged as a popular algorithm. While Deep Q networks have been successfully applied to a lot of scenarios, their application is based on the notion that the agent can completely perceive the environment. In real-time applications, this notion has a fallacy as complete observability is a difficult and sometimes impossible endeavor in real-time and the real world, therefore the use of recurrent networks along with Deep Q networks has been suggested for application to partially observable environments. This paper provides a literature review on various deep recurrent Q network applications. The paper first provides a brief introduction to the concept behind Deep Recurrent Q networks, and the various modifications to improve its performance and then proceeds to review its various applications in different fields.