Skip to main content Skip to navigation

Raising Student Completion Rates with Adaptive Curriculum and Contextual Bandits

Project Overview

The document explores the innovative use of generative AI in education through an adaptive learning Intelligent Tutoring System (ITS) named Korbit, which leverages model-based reinforcement learning (RL) to tailor learning experiences for students. Utilizing contextual bandits, Korbit effectively assigns exercises designed to optimize completion rates and boost student engagement. A randomized controlled trial highlighted the superior performance of the LinUCB algorithm compared to traditional educational methods, leading to increased completion rates and decreased skip rates, which signifies enhanced student engagement and overall learning outcomes. The system's capacity to autonomously learn and adapt from historical data underscores its potential for scalability and implementation in future educational contexts, paving the way for more personalized and effective learning environments.

Key Applications

Korbit - Adaptive Learning ITS

Context: Online learning platform for students in various educational contexts, including free users and corporate training programs.

Implementation: The system uses reinforcement learning algorithms (specifically LinUCB) to assign learning activities based on student trajectories and preferences, adapting in real-time.

Outcomes: Higher completion rates (87.4% success rate under LinUCB) and reduced skip rates (7.8% under LinUCB) compared to heuristic approaches, along with increased study time.

Challenges: Requires sufficient data on available exercises and the need for all exercises to pertain to the same topic for optimal functioning.

Implementation Barriers

Data Availability

The requirement for sufficient historical data on student interactions and exercise performance to optimize the bandit model.

Proposed Solutions: Plan to explore more sophisticated bandit algorithms and gather larger datasets for better performance.

Topic Restriction

The limitation that all available exercises must be related to the same topic, which can restrict the adaptability of the system.

Proposed Solutions: Future exploration of collaborative filtering methods to enhance the system's adaptability across different topics.

Project Team

Robert Belfer

Researcher

Ekaterina Kochmar

Researcher

Iulian Vlad Serban

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Robert Belfer, Ekaterina Kochmar, Iulian Vlad Serban

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