Enhancing autonomy and efficiency in goal-conditioned reinforcement learning

Yang, Z. (2025)

Doctoral Thesis

Reinforcement learning is a framework that enables agents to learn in a manner similar to humans, i.e. through trial and error. Ideally, we would like to train a generalist agent capable of performing multiple tasks and achieving various goals. Goal-conditioned reinforcement learning is a step toward training such an agent. The goal-conditioned reinforcement learning framework comprises four steps: 1) defining the goal space; 2) selecting an interesting goal for the agent; 3) the agent learning to reach the goal; 4) the agent post-exploring. In the thesis, we proposed four research questions and focused on three parts of the goal-conditioned reinforcement learning framework.

• Goal-conditioned reinforcement learning
• Episodic memory

• All authors: Yang, Z.
• Supervisor: Plaat, A.
• Co-supervisor: Moerland, T.M.; Preuss, M.
• Committee: Batenburg, K.J.; Bonsangue, M.M.; Francois-Lavet, V; Liu, J; Spaan, M.T.J
• Qualification: Doctor (dr.)
• Awarding Institution: Leiden Institute of Advanced Computer Science (LIACS), Faculty of Science, Leiden University
• Date: 2025-02-26
• Title of host publication: SIKS Dissertation Series
• ISBN (print): 97865069975
• Name: 2025-10