Reinforcement Learning Models for Curriculum Personalization

Abstract

Curriculum personalisation -- the dynamic sequencing of learning activities matched to individual learner needs, readiness, and goals -- is one of the most impactful yet technically challenging applications of artificial intelligence in education. Reinforcement learning (RL) provides a natural framework for curriculum personalisation: the pedagogical agent selects instructional actions (content, difficulty, modality) based on learner state observations, receiving rewards from learning outcome signals, and optimising a cumulative pedagogical objective. This paper proposes the RL Curriculum Personalisation Framework (RLCPF), a systematic evaluation of five RL model families -- Q-learning, deep Q-network (DQN), proximal policy optimisation (PPO), actor-critic (A3C), and multi-armed bandit (MAB) -- for curriculum personalisation across three educational contexts: K-12 mathematics, undergraduate programming, and corporate compliance training. RLCPF evaluates 2,800 learner episodes across 12 curriculum personalisation benchmarks. Key results: PPO achieves the highest Curriculum Effectiveness Score (CES = 0.912), reducing learning time to criterion by 36.8% and improving knowledge retention by 24.2% versus fixed-order curriculum; MAB achieves competitive CES (0.874) with 94x lower computational cost, making it the recommended choice for resource-constrained deployments. The framework contributes RL model selection guidance, reward function design principles, and an open curriculum personalisation benchmark suite.