From Machine to Human Learning: Towards Warm-Starting Teacher Algorithms with Reinforcement Learning Agents

Sidney Tio, Wenjun Li, Ramesha Karunasena, Ho Tian Sheng Jimmy, Pradeep Varakantham

Keywords: teacher algorithms, ai for human learning, automated curricula, cold-start problem

TL;DR: We use RL agents to collect data for warm-starting AI teacher algorithms, reducing the need for extensive human learning data while achieving effective adaptive curricula for humans in game-based training environments

Abstract: We present an investigation into using Reinforcement Learning (RL) agents to address the well-established cold-start problem in AI teacher algorithms that require extensive human learning data. While the challenge of bootstrapping personalized learning systems is recognized across domains, collecting comprehensive human learning data remains resource-intensive and often impractical. Our work explores a novel methodological approach: warm-starting data-hungry teacher algorithms using RL agents to provide an initial foundation that can be refined and augmented with human learning data. We emphasize that this approach is not intended to replace human data, but rather to provide a practical starting point when such data is scarce. Through extensive experiments in two game-based environments—a Super Mario-inspired platformer and an Overcooked-inspired medical training simulation—we conduct human subjects studies demonstrating that RL-initialized curricula can achieve comparable performance to expert-crafted sequences. Our preliminary analysis reveals that while human learning outcomes are positive, there remain notable gaps between RL agent behavior and human learning patterns, highlighting opportunities for improved alignment. This work establishes a promising potential for RL-initialized teaching systems, opening valuable research directions at the intersection of RL and human learning.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 10584