While Reinforcement Learning has achieved significant success, agents currently struggle with long-horizon tasks requiring extended reasoning and planning. Hierarchical Reinforcement Learning offers a solution through temporal abstraction, yet existing methods are often constrained by static, environment-specific assumptions that limit scalability. This research addresses these limitations by developing novel, agent-centric algorithms grounded in the paradigm of Continual Reinforcement Learning.
By moving beyond traditional Markov Decision Processes, this project focuses on enabling agents to discover and utilize transferable skills within dynamic streams of experience. The objective is to produce computationally scalable methods that support continuous adaptation and robust generalization. Ultimately, this work aims to bridge the gap between theoretical HRL and real-world application, enhancing both the performance and interpretability of autonomous decision-making.
Machine Learning, Reinforcement Learning, Temporal Abstraction in RL, Skill Discovery, Continual and Lifelong Learning
MComp Computer Science and Artificial Intelligence, University of Bath
Prof Özgür Şimşek
Prof Mike Tipping
