Transferring Task Goals via Hierarchical Reinforcement Learning

Saining Xie, Alexandre Galashov, Siqi Liu, Shaobo Hou, Razvan Pascanu, Nicolas Heess, Yee Whye Teh

Abstract: In this paper we demonstrate how hierarchical reinforcement learning architectures with guiding rewards can be used to separate high-level task understanding from low-level task execution and to exploit this separation of concerns for transferring high-level task goals e.g. across bodies with very different properties and capabilities.