Abstract
Hierarchical reinforcement learning deals with the problem of breaking downlarge tasks into meaningful sub-tasks. Autonomous discovery of these sub-taskshas remained a challenging problem. We propose a novel method of learningsub-tasks by combining paradigms of routing in computer networks and graphbased skill discovery within the options framework to define meaningfulsub-goals. We apply the recent advancements of learning embeddings usingRiemannian optimisation in the hyperbolic space to embed the state set into thehyperbolic space and create a model of the environment. In doing so we enforcea global topology on the states and are able to exploit this topology to learnmeaningful sub-tasks. We demonstrate empirically, both in discrete andcontinuous domains, how these embeddings can improve the learning of meaningfulsub-tasks.