Towards a fast and open-ended environment for curriculum learning in deep-reinforcement learning

Bruggen, Brian van (2025) Towards a fast and open-ended environment for curriculum learning in deep-reinforcement learning. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Bachelorsproject-10.pdf Download (681kB) | Preview
	Text Toestemming Van Bruggen.pdf Restricted to Registered users only Download (191kB)

Abstract

Reinforcement learning (RL) struggles with complex, sparse-reward tasks due to delayed feedback and the need for large amounts of computational resources. Curriculum learning and reward shaping, especially when guided by large language models (LLMs), help agents learn complex tasks by providing structured goals and intermediate feedback. In contrast, open-ended environments increase the challenge by introducing vast and unstructured goal spaces. This thesis presents craftax-wiring, an extension of the JAX-based Craftax-classic environment, designed to enable a wide range goals and curriculum learning for such goals. Inspired by Minecraft’s redstone and Terraria’s wiring systems, the environment introduces wires, logic gates, and inputs and outputs that simulate simple logic circuits. Experiments involving a half-adder construction task reveal that agents fail to learn without learning support, confirming the task’s complexity. When guided through a handcrafted reward-shaping structure, agents show successful patterns that indicate the learning of the task, showing the environment’s suitability for curriculum learning. Additionally, LLM evaluation confirms the high interestingness of the new goals. Performance tests show that the new environment runs at a similar speed when compared to the Craftax environment. Overall, Craftax-wiring shows promise for being a fast and open-ended environment for curriculum learning in RL research.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Pourcel, G. A.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	03 Sep 2025 09:58
Last Modified:	03 Sep 2025 09:58
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/36869

Actions (login required)

View Item