Leveraging Reinforcement Learning for Enhancing Educational Environments
Event details
| Date | 12.01.2024 |
| Hour | 15:00 › 17:00 |
| Speaker | Bahar Radmehr |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Antoine Bosselut
Thesis advisor: Prof. Tanja Käser
Co-examiner: Prof. Caglar Gulcehre
Abstract
Digital learning environments have increasingly become
a cornerstone of modern education. This shift has spurred
extensive research on applied Reinforcement Learning (RL) in
education, focusing on enhancing various aspects such as personalized
curriculum design, providing customized hints, modeling
student learning behaviors, and auto-grading. Despite the growth
in research, RL methods are not yet widely adopted in a variety
of learning platforms, primarily due to the challenges in adapting
agents for reasoning tasks in complex educational environments
with usually large, natural language-based search spaces, misalignment
between agent and human behaviors, and limitations
on agent policies by constrained instructional opportunities.
This doctoral candidacy proposal integrates insights from three
papers to address these challenges in RL application in education.
The first paper introduces a language-enabled agent framework
for complex interactive reasoning tasks based on Behavior
Cloning and Prompting Large Language Models. The second
paper proposes a behavior cloning approach to align AI behavior
with human strategies, using next-move and error prediction
models that replicate humans’ behavior at different skill levels.
The third paper explores enhancing RL agents in constrained
instructional settings by identifying critical decisions for desired
outcomes. Our proposed research agenda, inspired by these
papers, aims to foster RL adoption in education by developing
language-enabled agents with reasoning and planning capabilities
tailored to complex educational environments, aligning agent
behavior with human behavior, and dentifying critical decisions
needed for constrained tutoring.
Background papers
1. Pick the Moment: Identifying Critical Pedagogical Decisions Using Long-Short Term Rewards
2. Aligning Superhuman AI with Human Behavior: Chess as a Model System
3. SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex Interactive Tasks
Exam president: Prof. Antoine Bosselut
Thesis advisor: Prof. Tanja Käser
Co-examiner: Prof. Caglar Gulcehre
Abstract
Digital learning environments have increasingly become
a cornerstone of modern education. This shift has spurred
extensive research on applied Reinforcement Learning (RL) in
education, focusing on enhancing various aspects such as personalized
curriculum design, providing customized hints, modeling
student learning behaviors, and auto-grading. Despite the growth
in research, RL methods are not yet widely adopted in a variety
of learning platforms, primarily due to the challenges in adapting
agents for reasoning tasks in complex educational environments
with usually large, natural language-based search spaces, misalignment
between agent and human behaviors, and limitations
on agent policies by constrained instructional opportunities.
This doctoral candidacy proposal integrates insights from three
papers to address these challenges in RL application in education.
The first paper introduces a language-enabled agent framework
for complex interactive reasoning tasks based on Behavior
Cloning and Prompting Large Language Models. The second
paper proposes a behavior cloning approach to align AI behavior
with human strategies, using next-move and error prediction
models that replicate humans’ behavior at different skill levels.
The third paper explores enhancing RL agents in constrained
instructional settings by identifying critical decisions for desired
outcomes. Our proposed research agenda, inspired by these
papers, aims to foster RL adoption in education by developing
language-enabled agents with reasoning and planning capabilities
tailored to complex educational environments, aligning agent
behavior with human behavior, and dentifying critical decisions
needed for constrained tutoring.
Background papers
1. Pick the Moment: Identifying Critical Pedagogical Decisions Using Long-Short Term Rewards
2. Aligning Superhuman AI with Human Behavior: Chess as a Model System
3. SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex Interactive Tasks
Practical information
- General public
- Free