An Instructional Optimization Method Based on Bidirectional Transformer and Reinforcement Learning

With the rapid development of information technology, personalized education has become a key direction for improving the quality of online learning and optimizing individualized learning paths. However, accurately recommending appropriate courses and exercises for diverse learners remains a significant challenge. Existing recommendation methods often struggle with effectively modeling learner interests, addressing the cold-start problem, and dynamically adapting recommendation strategies to meet personalized needs. To address these limitations, this paper proposes RL-TBTNet, a novel teaching optimization recommendation framework that integrates a bidirectional Transformer, BERT, and reinforcement learning (DQN). The model first vectorizes user behavior data, learning content, and knowledge base information. Transformer layers are employed for feature encoding, while BERT extracts deep semantic representations to form individualized feature vectors. These features are then fused via Transformer-based processing to predict optimal learning content. In addition, a DQN-based reinforcement learning module models dynamic shifts in user interests, enabling adaptive refinement of learning trajectories over time. Experimental evaluations on public datasets show that RL-TBTNet outperforms existing Transformer-based methods such as BST in terms of key metrics like HR and NDCG, particularly excelling in cold-start scenarios. Ablation studies further confirm the effectiveness of semantic enhancement through BERT and reinforcement-driven optimization. These results demonstrate the framework’s potential as a robust and adaptive solution for personalized educational content recommendation, offering both practical value and theoretical insights for the development of intelligent education systems.