Two-Level Actor-Critic Using Multiple Teachers

Reinforcement learning (RL) has been successful in a variety of domains ranging from solving difficult games like Go [10] and drug discovery [4]. Most of these domains are characterized by high-dimensional states and continuous action spaces. However, sample inefficiency is a major challenge when applying these algorithms to real-world tasks such as robotics and healthcare care [6]. To address improved sample efficiency, rather than forcing agents to learn from scratch, domain knowledge from humans or existing agents can be leveraged in various ways [3].