Prompt-to-policy: Leveraging large language models to guide deep reinforcement learning in public health emergencies

Executive Summary

The article discusses a hybrid agent that integrates D2QN-JDA with large language models (LLMs) for dynamic decision-making in public health emergencies. LLMs with retrieval-augmented generation (RAG) process multi-source heterogeneous data to calculate adaptive rewards. The D2QN-JDA method is validated on real-world COVID-19 data from Hong Kong, demonstrating superior cost efficiency and robustness compared to other algorithms. The approach automates the processing of heterogeneous, real-time data, reducing manual intervention and enhancing system robustness. LLMs outperform manual and regex methods in data extraction, improving accuracy, recall, F1 score, cost, and time.