A core challenge in hierarchical reinforcement learning is skill discovery: how do agents independently discover useful higher-level skills for a problem? Skill discovery methods seek to solve this problem, and two categories of these approaches are graph-based and information theory. Graph-based approaches analyse the structure of a given problem to generate useful skills, whereas information theory approaches use statistical methods to learn a set of diverse skills. However, little work has been done in combining these approaches. My project aims to unify these approaches and develop methods that share both categories’ benefits and lack their drawbacks.
Artificial Intelligence, Hierarchical Reinforcement Learning, Skill Discovery, Transfer Learning
MComp Computer Science and Mathematics at the University of Bath
Prof Özgür Şimşek
Prof Guy McCusker
