Design and evaluation of personalised early childhood mathematics enlightenment games based on deep reinforcement learning

Early childhood mathematics learning benefits from adaptive technologies; however, effective real-time personalisation remains a challenge. However, prior models often rely on static difficulty or rule-based adaptation, with limited evidence in preschool-aged children. This study asks whether a multimodal deep reinforcement learning (DRL) model can outperform rule-based adaptation in personalising preschool math tasks. A total of 84 children aged 4–6 participated in a controlled trial comparing DRL personalisation to rule-based adaptation. The DRL system incorporated multimodal inputs (click latency, hesitation, facial affect) and was trained via a Deep Q-Network on a Dell server with GPU acceleration. Performance, engagement, and mastery outcomes were analysed using mixed-effects ANOVA and benchmarked against established adaptive models. DRL users improved EMAS scores by + 12.7 points versus + 6.8 in controls (p = 0.001), achieved 84.2% task accuracy compared to 76.3% (p < 0.001), and reduced response latency from 2.61 s to 1.91 s (p < 0.001). Level 3 mastery reached 69.0% versus 40.5% (p = 0.007). Hesitation frequency was lower (0.18 vs. 0.29 per task), joy expression higher (0.36 vs. 0.27 AU rate), and dropout flags fewer (3.1% vs. 8.9%), all p < 0.001. External comparisons confirmed superior accuracy, stability of engagement, and fidelity of execution. Multimodal DRL significantly enhances early math learning, supporting both technical personalisation and meaningful educational gains in preschool settings.