Dynamic data management in industry university integration platform: An embedded decision framework

Abstract

This article proposes an innovative embedded dynamic decision-making framework (KD-DRL) to address the key issues of insufficient interpretability and low training efficiency faced by intelligent decision-making systems in industry academia integration platforms. This framework achieves collaborative optimization of data-driven methods and domain knowledge through the organic integration of deep reinforcement learning and explicit knowledge reasoning, providing a new technological path for intelligent decision-making in complex educational scenarios. In terms of method design, KD-DRL innovatively constructs a dual module knowledge system: the heuristic acceleration knowledge module guides agents to explore efficiently in the early stages of training through intelligent intervention mechanisms, significantly reducing model convergence time; The evasive security knowledge module is based on formal rules to monitor the decision-making process in real time, effectively preventing potential risks. This hierarchical knowledge architecture not only retains the ability of deep reinforcement learning to handle high-dimensional state spaces, but also ensures the reliability and security of decisions through interpretable rule constraints. Experimental verification shows that in the multimodal data environment of real industry academia platforms, KD-DRL exhibits significant advantages: in terms of training efficiency, the model convergence speed is improved by more than 40% compared to traditional deep reinforcement learning methods; In terms of safety, the incidence of teaching accidents has been reduced to below 0.3%, while maintaining a reasonable intervention frequency; In terms of interpretability, the decision-making process supports rule tracing, and the average time for experts to verify a single decision logic is controlled within 5 s.