Login Sign up

GTCOARLab: Computable Construct as Curriculum for Reinforcement Learning

Nicholas Bollweg

Audience level:

Brief Summary

Come see our collection of wonderful things for learning about reinforcement learning, from theory to deployment. As a mybinder link, or cross-platform, offline-ready installer, you can follow (or customize) a self-guided course of study in our custom Jupyter environment which delivers hands-on tutorials, and cutting-edge research papers, along with open source reference implementations.


In GTCOarLab1 , we demonstrate an open source, adaptable pipeline for delivering self-contained learning environments accessible via free cloud resources, or distributed on "dumb" media and installed off-line. While our specific, motivating topic, reinforcement learning (a sub-discipline of machine learning) carries many domain-specific learning requirements, we explore delivering a broad range of learning activities in a single JupyterLab user interface, from traditional research skills like discovering, reading, and referencing research papers, to interactive experimentation and note-taking, to testing and packaging of research software, and finally to authoring of publication artifacts like papers, presentations, and posters.

This work is motivated by helping distributed learners and instructors overcome challenges in transferring mastery of a complex technical field. Current approaches often delegate this to cloud-delivered applications, which contribute to "learning lock-in", which then requires a base assumption of always-on, low-latency internet access, limiting the accessibility to under-served populations.