Decision analysis based on reinforcement learning and transfer learning

Aiming at the static and one-size-fits-all limitations of traditional comprehensive student quality evaluation, this work proposes an intelligent evaluation and personalized development planning model that integrates reinforcement learning (RL) and transfer learning (TL). The students’ development process is modeled as a Markov decision process (MDP). A deep neural network (DNN) is adopted to represent the multi-dimensional comprehensive quality state. Personalized development suggestions serve as decision actions, and long-term ability improvement acts as a reward signal to drive the RL agent to learn an optimal guidance strategy. The TL mechanism is introduced to solve the problems of “cold start” and insufficient cross-scene generalization of the model; the universal network is pre-trained by large-scale data in the source domain and transferred to the target domain to realize rapid adaptation. The verification based on several public datasets, such as the Open University Learning Analytics Dataset, shows that the model achieves precision, recall, F1-score, and Area Under Curve (AUC) scores of 0.902, 0.873, 0.887, and 0.926; these are significantly better than the traditional algorithms. Personalized planning can improve the learning engagement by 27.6%, the course success rate reaches 84.5%, the convergence time of the target domain is only 41 min, and the stability coefficient is 0.019. This model provides practical support for the implementation of “teaching students in accordance with their aptitude” in intelligent education.