Reinforcement Learning System Plans Children’s Math Learning Paths With 93.8 Percent Accuracy

A newly developed Learning-Based Q Network designed to dynamically plan personalized mathematics learning paths for young children achieved a 93.8 percent accuracy rate in experimental evaluations. In comparative performance tests, the proposed framework surpassed multiple established models, specifically outperforming A-DRL, MLP-LSTM, and RL-DKT architectures across standard performance metrics. Alongside its primary accuracy score, the computational model recorded a recall of 90.6 percent, an F1-score of 91.7 percent, and achieved a user satisfaction rating of 4.9 out of 5.

The development addresses persistent structural challenges in adaptive educational technologies designed for early childhood mathematics instruction. Although digital platforms increasingly support early mathematics instruction, most existing systems depend heavily on static or rule-based personalization mechanisms. These conventional designs restrict an educational system’s ability to adapt effectively to ongoing changes in learner behavior during instruction. Furthermore, conventional adaptive learning approaches fail to integrate multimodal behavioral and emotional signals effectively, which researchers identify as a primary cause of reduced student engagement, slower knowledge acquisition, and constrained personalization for preschool learners.

To overcome these limitations, the researcher proposed a Learning-Based Q Network, designated as LBQN, engineered to formulate learning paths that respond dynamically to children’s cognitive and behavioral states. The architectural foundation of the LBQN model combines a Double Deep Q-Network, known as DDQN, with Teaching-Learning-Based Optimization, or TLBO. By integrating DDQN reinforcement learning with offline TLBO parameter refinement, the system establishes a sequential hybrid framework specifically structured to improve convergence stability, optimize learning path quality, and bolster adaptive, data-driven decision-making for young learners.

The experimental investigation utilized the Personalized Math Learning Behavior Dataset to train and evaluate the adaptive path planning algorithms. This empirical dataset comprises 4,000 individual data rows arranged across 14 distinct feature columns representing diverse learner interactions.

Prior to model training, the raw behavioral data underwent a systematic, multi-stage preprocessing workflow to ensure data reliability and consistency. The pipeline incorporated median filtering to eliminate noise from recorded signals, paired with min-max normalization to achieve uniform feature scaling across all variables. To address multimodal behavioral streams collected during learning tasks, the researchers applied temporal alignment using a sliding window technique, thereby synchronizing disparate behavioral inputs into a coherent chronological structure suitable for reinforcement learning inputs.

Following data preprocessing, feature extraction isolated both statistical measurements and explicit behavioral indicators from the synchronized records. The extracted statistical features included mean and variance calculations, while behavioral indicators focused on operational task metrics, specifically student response times and measured error rates. The LBQN framework utilized these extracted features within a Python-based implementation to guide action selection and learning path planning through reward-driven policy updates, allowing the system to adjust pedagogical trajectories based on observed performance and behavioral metrics.

The sequential combination of reinforcement learning and optimization parameters demonstrates that hybrid computational models can effectively enhance personalized mathematics instruction for early learners. By uniting DDQN policy learning with offline TLBO optimization, the framework establishes a reliable foundation for next-generation intelligent tutoring systems. The researchers conclude that this data-driven methodology provides practical capabilities for improving learning outcomes, delivering continuous personalization, and supporting adaptive instructional decision-making across early childhood education environments.

Subject of Research: Personalized mathematics learning path planning for children using reinforcement learning

Article Title: Deep reinforcement learning based personalized mathematics learning path planning algorithm for children

Article References: Qin, Q. (2026). Deep reinforcement learning based personalized mathematics learning path planning algorithm for children. Discover Applied Sciences. https://doi.org/10.1007/s42452-026-09464-6

Image Credits: AI Generated

DOI: 10.1007/s42452-026-09464-6

Keywords: reinforcement learning, adaptive learning, early childhood education, mathematics education, intelligent tutoring systems, Double Deep Q-Network, Teaching-Learning-Based Optimization, learning path planning, Deep, reinforcement, learning, based