EVOL: Simulator-Guided Evolutionary Expert Synthesis for Deployment-Free Learning Path Recommendation

TL;DR

EVOL introduces a simulator-guided evolutionary approach to generate expert learning paths for reinforcement learning in education, overcoming the lack of expert data and sparse rewards. It uses a knowledge tracing simulator to evolve high-quality demonstrations and trains a policy via imitation learning without requiring deployment. The method outperforms multiple baselines across diverse datasets and path lengths.

Why It Matters

Many educational recommendation systems lack expert-curated learning paths, limiting the effectiveness of reinforcement learning. EVOL enables high-performing, deployment-free learning path recommendations by synthesizing expert behavior through simulation, making it practical for real-world adaptive learning platforms.

Key Findings

• EVOL generates expert learning paths via evolutionary search in a knowledge tracing simulator, addressing the absence of expert data in student logs.
• The asymmetric actor-critic architecture allows the actor to plan blindly (as in real deployment) while the critic uses privileged simulator states during training.
• Performance is primarily determined by the quality of evolved experts, not the choice of imitation learning objective (BC, AWR, DAPG).