Learning Methods for Solving Astronomy Course Problems

Abstract

This work trains a specialized machine learning model to solve university undergraduate level Introduction to Astronomy course problems. We use Transformers and graph neural networks to predict arithmetic expression trees with three key contributions: (i) a new dataset of formatted Introduction to Astrophysics course questions; (ii) support of new mathematical operators; and (iii) inclusion of tabular data consisting of known physical constants and units. Perhaps most importantly, this work introduces the concept of turning questions into programming tasks and using a Transformer trained on both text and code, namely OpenAI Codex, to solve these quantitative questions. Comparing our specialized model trained only on Astrophysics with the generic Codex model, across all topics taught in the course, we find a classical trade-off between accuracy on a specific course and the ability to scale to many courses. Our specialized model achieves 92% though does not scale to any other course; whereas the generic Codex achieves 67% on Astrophysics and is able to scale to many different courses.

Tasks

Reproductions