Agent-advising approaches in an interactive reinforcement learning scenario

摘要

Reinforcement learning has become one of the fundamental topics in the field of robotics and machine learning. In this paper, we expand the classical reinforcement learning framework by the idea of external interaction to support the learning process. To this end, we review a number of proposed advising approaches for interactive reinforcement learning and discuss their implications, namely, probabilistic advising, early advising, importance advising, and mistake correcting. Moreover, we implement the advice strategies for interactive reinforcement learning based on a simulated robotic scenario of a domestic cleaning task. The obtained results show that the mistake correcting approach outperforms a purely probabilistic advice approach as well as the early and importance advising approaches allowing to collect more reward and also to converge faster.