Rekomendasi Aktivitas Pembelajaran Anak Usia Dini Berbasis Q-Learning dan Profil Perkembangan

Abstract

This study designs a Q-Learning-based recommendation system for early childhood learning activities in a Raudhatul Athfal setting. The system supports teachers in selecting activities aligned with children's developmental profiles, especially physical-motor and cognitive aspects. The recommendation problem is formulated as a Markov Decision Process. The state consists of children's age in months, physical-motor level, cognitive level, previous activity, and previous participation score. The action space contains twelve learning activities, while the reward combines participation, developmental fit, and activity variation. Testing was conducted using scenario-based learning data with five experimental seeds. The results show that Q-Learning achieved an average evaluation reward of 24.667 with a standard deviation of 0.222 from a theoretical scenario bound of 30 points. Ranking evaluation produced Precision@1 of 0.645, Recall@5 of 0.448, and NDCG@5 of 0.641. These results support Q-Learning as a transparent tabular baseline, but they do not prove pedagogical impact. Q-Learning was selected because the state is discrete, the action space is limited, and Q-values are traceable. Since operational data have not been used, the claims are limited to system design, recommendation traceability, and computational evaluation.