An extensive review of applications, methods and recent advances in deep reinforcement learning

Abstract

Reinforcement learning is a learning problem where an actor must behave optimally in its given environment. A subtype of representation learning, deep learning approaches concentrate on extracting the required features for the problem. Deep Reinforcement Learning (DRL) is an extremely dynamic area that combines the two. It is also the most popular kind of machine learning since it can tackle a variety of complex decision-making issues that were previously beyond the capacity of a computer to think like a human. This paper first analyses the various case studies related to the applications of DRL. DRL applications involve robotics, video data analysis, transportation, industrial applications, Natural Language Processing (NLP) and so much more. DRL finds major success in applications involving finding out the best policy for an action to maximize an objective given a rich context in time. The second part of the paper will discuss different methods that are used to implement DRL concepts. These involve Deep Q Learning (DQN), Markov Decision Processes (MDP), and Deep Deterministic Policy Gradient (DDPG). Additionally, it discusses the two widely categorized types of methods: policy-based and value-iteration based. Lastly, this paper provides a survey of the recent advances made toward improving the performance of DRL-based implementations.