Exploring the Landscapes and Emerging Trends of Reinforcement Learning from 1990 to 2020:A Bibliometric Analysis

摘要

Reinforcement Learning (RL) becomes increasingly important in recent years as the huge success of AlphaGo and AlphaZero. However, this technique is a not a newly born research topic, which originates from the well-developed dynamic programming method. In this paper, we explore the history of RL from the bibliometric perspective for the last 30 years, to capture its landscapes and emerging trends. We conduct comprehensive assessments of the RL technology according to articles related to RL in SCI database from 1990 to 2020, and extensive results indicate that reinforcement learning research goes up significantly in the past three decades, including a total of 9344 articles covering 96 countries/territories. Top five most productive countries are USA. China, England, Japan. Germany and Canada. There are 4507 research institutes involved in the field of RL and among them the top five productive ones are Chinese Academy of Sciences, University College London. Beijing University of Posts and Telecommunications, Tsinghua University and Northeastern University and Princeton University. Besides, top frequently adopted keywords with strongest citation burst are Genetic Algorithm. Dynamic Programming. Q-Learning, Mobile Robot. Wireless Sensor Network, Smart Grid, Big Data, Inverse Reinforcement Learning and Cognitive Radio, which demonstrate the emerging trends of this field. We claim the results shown in this paper provide a dynamic view of the evolution of "Reinforcement Learning" research landscapes and trends from various perspectives that is able to serve as a potential future research guide, and the way we demonstrate could also be adopted to analyze other research topics.