Skip to main content Skip to navigation

Reinforcement Learning Tutor Better Supported Lower Performers in a Math Task

Project Overview

The document explores the integration of generative AI and reinforcement learning (RL) in education, particularly focusing on enhancing personalized instruction through educational software aimed at elementary school students learning math. It addresses the challenges of delivering effective personalized instruction amid resource constraints and advocates for the use of RL to automate and refine pedagogical approaches. By employing a narrative-based AI tutor, the findings demonstrate that RL can successfully tailor support to meet individual student needs, particularly benefiting those who start with lower performance levels. The emphasis on explainability in the AI's decision-making process is also highlighted, underscoring the necessity for transparency in how AI systems assist learners. Overall, the document illustrates the promising potential of generative AI and RL to improve educational outcomes and foster a more personalized learning experience.

Key Applications

Reinforcement Learning Narrative AI Tutor

Context: Educational software for elementary school students learning math concepts, specifically ages 9 to 12.

Implementation: Developed a narrative-based software where an AI guide uses reinforcement learning to adaptively provide pedagogical support based on student interactions.

Outcomes: Students with low pretest scores showed significant improvement in math learning and engagement compared to a control group.

Challenges: Variability in student backgrounds may affect the RL policy's effectiveness; potential ceiling effects on assessments.

Implementation Barriers

Resource Limitations

Limited resources make it difficult to provide personalized instruction to all students.

Proposed Solutions: Using reinforcement learning to automate and optimize the development of educational software to lower costs.

Student Variability

Natural variation in student backgrounds and skills may lead to inconsistent performance of the RL policy.

Proposed Solutions: Stratification of student demographics to ensure stable distributions in initial performance for better policy convergence.

Assessment Limitations

Ceiling effects in assessments can limit the ability to gauge learning gains accurately.

Proposed Solutions: Improving assessment diversity and possibly employing normalized learning gains for better evaluation.

Project Team

Sherry Ruan

Researcher

Allen Nie

Researcher

William Steenbergen

Researcher

Jiayu He

Researcher

JQ Zhang

Researcher

Meng Guo

Researcher

Yao Liu

Researcher

Kyle Dang Nguyen

Researcher

Catherine Y Wang

Researcher

Rui Ying

Researcher

James A Landay

Researcher

Emma Brunskill

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Sherry Ruan, Allen Nie, William Steenbergen, Jiayu He, JQ Zhang, Meng Guo, Yao Liu, Kyle Dang Nguyen, Catherine Y Wang, Rui Ying, James A Landay, Emma Brunskill

Source Publication: View Original PaperLink opens in a new window

Project Contact: Dr. Jianhua Yang

LLM Model Version: gpt-4o-mini-2024-07-18

Analysis Provider: Openai

Let us know you agree to cookies