Investigation of the Necessity of Past Input/output Information in Reinforcement Learning based Robust Control

Abstract

Reinforcement learning yields a feedback controller that achieves specific control goal (which is often translated as a reward function). However, it often suffers from the Sim2Real gap, and domain randomization is known to be a method to overcome this issue. In this paper, we demonstrate necessity of input/outpu history when domain randomization is employed by a formal example and a simulation result. This is equivalent to the necessity of dynamic feedback controller in terms of control theory. © The Korean Institute of Electrical Engineers