A Systematic Study of Deep Q-Networks and Its Variations

Abstract

Recently, reinforcement learning has been at the forefront of the automation revolution. Reinforcement learning has proved to be a particularly powerful technique since it can understand the actions that will lead to ultimate success in an unfamiliar environment without any of the intervention of a supervisor. Due to the advent of Neural Networks, the unison of Deep learning and Reinforcement learning has led to unprecedented scaling of previously thought computationally expensive algorithms. This has caused an emergence of autonomous systems that can play games and automate robotics using visual cues, to name a few. In this paper, we intend to analyse a particular reinforcement learning algorithm called Deep Q-Networks and its numerous other variants. We cover several approaches including Improved Replay Memory, Distributional Outlook of Deep Q-Networks. Also, we discuss their advantages and disadvantages and conclude by detailing the current area of research in this field and future progress in sight.