He, Jinke, Maximilian Igl, Matthew Smith, Wendelin Boehmer, and Shimon Whiteson. “Soft Option Transfer.”
URL
Transferring learnt options in hierarchical RL can yield poor performance when
they are even slightly misaligned to the new task. This paper introduces soft
option transfer: the given options are treated as a prior to learn task-specific option
posteriors. This combines the fast exploration of transferred options with the
flexibility to adjust them if need be. We investigate our approach in the taxi domain
with varying option applicability and exploration complexity. The experiments
demonstrate a clear advantage over flat policies and ‘hard’ options augmented with
primitive actions.