Abstract

Sparse-reward domains are challenging for reinforcement learning algorithmssince significant exploration is needed before encountering reward for thefirst time. Hierarchical reinforcement learning can facilitate exploration byreducing the number of decisions necessary before obtaining a reward. In thispaper, we present a novel hierarchical reinforcement learning framework basedon the compression of an invariant state space that is common to a range oftasks. The algorithm introduces subtasks which consist of moving between thestate partitions induced by the compression. Results indicate that thealgorithm can successfully solve complex sparse-reward domains, and transferknowledge to solve new, previously unseen tasks more quickly.

 

Quick Read (beta)