Collaborative Research: Multi-Agent Adaptive Data Collection for Automated Post-Disaster Rapid Damage Assessment

在灾难发生后立即进行勘察工作，以确定建筑物损坏的严重程度和分布，对于搜救和其他时间敏感的决策至关重要。然而，现有的数据收集和分析进程对不可预见和意外情况的反应较差。因此，该项目开发了一种新的自适应数据收集框架，该框架不断分析最新的观测数据，以确定和更新数据收集器代理指向信息获取潜力最大的地区的轨迹。它使这些代理能够在严格的时间和资源限制下收集可靠的数据。该项目的成果为自动损害评估系统奠定了基础，以提高灾害多发地区建成环境和公民的复原力。为了广泛传播研究成果，并将其整合到本科生和研究生课程中，设想了一系列教育和推广工作。自适应数据收集系统建立在一个新的分层贝叶斯框架上，用于对灾害破坏程度进行建模，并对自适应目的地识别和轨迹规划进行贝叶斯优化。这一方法首先依赖于灾害前的初步概率模型，利用公众可获得的先验信息和空间属性，在人口普查区域的粒度级别上对不同类型的建筑物进行物理损坏程度的初步概率模型。然后，它通过与从走访地区收集的数据相结合，创建并不断更新人口普查区域的有形破坏程度分布。接下来，它动态地自适应地确定多个代理的轨迹，以在尽可能短的时间内最大化信息收益。贝叶斯概率模型可以转移到其他复杂的问题，如环境污染评估。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

In the immediate aftermath of a disaster, reconnaissance efforts to identify building damage severity and distribution are critical for search and rescue and other time-sensitive decisions. However, existing data collection and analytical processes are less responsive to unforeseen and unexpected circumstances. Therefore, this project develops a novel adaptive data collection framework that constantly analyzes the most recent observations to determine and update the trajectory of data collector agents toward areas with the greatest potential for information gain. It enables these agents to collect reliable data under severe time and resource constraints. The outcomes of this project set the stage for automated damage assessment systems to improve the resilience of built environments and citizens in hazard-prone regions. A set of educational and outreach efforts are envisioned for broadly disseminating the research findings and integrating them into undergraduate and graduate courses.The adaptive data collection system is built on a novel hierarchical Bayesian framework for modeling disaster damage levels and Bayesian optimization for adaptive destination identification and trajectory planning. This method first relies on a pre-disaster preliminary probabilistic model of physical damage levels for different types of structures at the census tract level of granularity using a priori information and spatial attributes available to the public. It then creates and constantly updates distributions of physical damage levels across census tracts by integrating with the collected data from visited zones. Next, it dynamically and adaptively determines the trajectories for multiple agents to maximize information gain in the shortest possible time. The Bayesian probabilistic models could be transferable to other complex problems such as environmental pollution assessment.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.