TLDR: The H–J framework is a new AI-driven system designed to optimize urban flood emergency response. It uses large language models (LLMs) integrated with knowledge graphs and a multi-agent setup. Key innovations include combining structured and unstructured knowledge for better LLM prompts, using “entropy constraints” to ensure stable and diverse strategy generation, and a feedback loop that continuously refines policies based on real-time performance. Experiments show H–J significantly outperforms existing methods in managing traffic and tasks during floods, demonstrating improved robustness and adaptability across various rainfall scenarios.
Urban flooding, intensified by extreme rainfall events, presents significant challenges to city management, leading to traffic chaos and disruptions to essential services. Traditional emergency scheduling systems often struggle with balancing multiple objectives like traffic flow and task completion, adapting to rapidly changing conditions, and dealing with inconsistencies in AI-generated strategies.
A new research paper introduces an innovative solution called H–J, a hierarchical multi-agent framework designed to enhance urban flood response. This framework integrates large language models (LLMs) with knowledge graphs and advanced optimization techniques to create a robust and adaptive decision-making system.
Understanding the H–J Framework
The H–J framework operates on a closed-loop cycle, continuously perceiving the environment, generating strategies, executing them, evaluating outcomes, and refining policies. Its core strength lies in three key innovations:
First, it uses a “dual-channel knowledge retrieval” system. This means it combines structured information, like urban knowledge graphs (representing roads, regions, and flood spots), with unstructured data, such as emergency reports and historical logs. By fusing these two types of knowledge, H–J creates highly informed prompts for the LLM, ensuring that the generated strategies are relevant and grounded in real-world context.
Second, H–J employs “entropy-constrained hierarchical generation.” LLMs can sometimes produce unstable or inconsistent outputs. To counter this, H–J regulates the “information entropy” across different planning levels. High-level plans are kept semantically consistent, while lower-level execution maintains flexibility. This balance ensures that the multi-agent strategies are both robust and adaptable.
Third, the framework features “objective-driven feedback optimization.” It continuously evaluates the effectiveness of its policies using a weighted objective function that considers factors like flood severity, traffic congestion, task cancellations, and arrival rates. If there’s a significant deviation from expected outcomes, the system triggers a re-planning process, allowing for dynamic adaptation to changing flood conditions. This closed-loop feedback mechanism ensures that the system learns and improves over time.
Real-World Impact and Performance
The researchers evaluated H–J using real-world urban topology and rainfall data from a metropolitan area, testing it under various conditions: extreme rainfall, intermittent bursts, and daily light rain. The results were highly promising. H–J consistently outperformed traditional rule-based systems and reinforcement learning baselines in key areas such as traffic smoothness, task success rates, and overall system robustness. For instance, in scenarios with intermittent rainfall, H–J demonstrated superior stability compared to other methods, which often prioritized short-term gains over system-wide stability.
A detailed analysis showed that each of H–J’s core modules contributes significantly to its overall performance. The dual-channel knowledge indexing was crucial for semantic precision, especially during extreme rainfall. The entropy control helped balance strategy diversity with coherence, while the feedback mechanism ensured long-term adaptability and reduced output variance.
Also Read:
- Advancing AI’s Problem-Solving: A Dual Approach to Heuristic Design
- Improving LLM Graph Reasoning with a Human-Inspired Collaborative Framework
Looking Ahead
While H–J represents a significant leap forward in urban flood management, the researchers acknowledge practical limitations, such as high deployment costs and reliance on substantial inference resources. Future work aims to explore more lightweight coordination schemes and efficient knowledge fusion mechanisms to enhance its real-world applicability. The paper also suggests further systematizing this approach under a computational-experiment paradigm to improve explainability and transferability across various urban risks.
For more technical details, you can read the full research paper available at arXiv:2508.14654.
