Reinforcement learning algorithms using for agent behaviour modelling and researching

Abstract

Achieving autonomy and rational behaviour, especially when the environment is complex, partially observable, dynamic or unfamiliar, requires the inclusion of machine learning methods. This article presents the latest and most widely used model-based and model-free reinforcement learning algorithms and their applications. In model-based reinforcement learning algorithms, the learning agents have knowledge of their environment, which leads to faster learning and faster building of optimal policies. In model-free reinforcement learning algorithms, agents build optimal policies by interacting with the environment, which makes their training slower. On the other hand, their advantage is that they can work with unfamiliar and dynamically changing environments. Some hybrid reinforcement learning algorithms are also considered.