Nonlinear Filters

11Reinforcement Learning‐Based Filter

11.1 Introduction

Reinforcement learning is a branch of artificial intelligence that is focused on goal‐oriented learning from interactions with an environment [233]. Reinforcement learning algorithms aim at finding effective suboptimal solutions to complex problems of sequential decision making and planning under uncertainty [234]. Optimal decision making is a core concept in many disciplines such as engineering, computer science, mathematics, economics, neuroscience, and psychology. Therefore, there is a branch of these disciplines that deals with the same problems that reinforcement learning tries to solve. In this regard, reinforcement learning can be viewed as a multi‐disciplinary field with a wide range of applications.

The probabilistic models used to understand and analyze data have been enhancing in terms of complexity and scale to cope with big data, which is associated with five key attributes: volume, velocity, variety, veracity, and value (five Vs). Due to this trend, performing inference has become a more challenging task. Regarding the duality between estimation and control [235], these two fields can borrow algorithms from each other. For instance, by viewing control as an inference problem, control algorithms can be designed based on belief propagation. Similarly, learning the approximate posterior distribution in variational inference can be mapped to the policy optimization problem in reinforcement learning to design ...