ATD: Surveillance Evasion and Threat Avoidance

行人运动数据的日益可用性允许人们的目标，偏好和对环境的感知的日益复杂的模型。 这样的模型不仅在交通工程中很重要（“我们如何构造这座建筑的门厅，以避免在疏散时发生踩踏事件？“），而且有助于提高我们发现城市环境中新出现的威胁的能力。 任何戏剧性的变化，在通常的“生活模式”可能会提供一个线索，不断演变的条件（“为什么这群行人采取这样一个不寻常的路径，他们的目标？ 这个路口怎么突然有这么多人？”），而对较长时间范围内聚集的数据进行分析也有助于改善我们的监测和建模（“城市的哪些部分通常被认为更危险？ 我们应该把有限的观察资源部署在哪里？"). 该项目侧重于两个具体应用：（a）危险环境中的平民规划其路径，以尽量减少威胁暴露，（B）了解现有监测措施的对手规划其路径，以逃避观察。 在这两种情况下，PI提出了模型和数值方法，用于（1）基于对环境的信念的路径规划和（2）对抗性/鲁棒性路径规划，其中环境可能会响应于人们的路由选择而变化。 自私/独立决策下的威胁规避也将在“平均场博弈”的框架下进行处理。 所提出的方法借鉴了博弈论，凸优化，最优控制和多目标动态规划的方法。 该模型内置的信息模式，使人们有可能利用效率的快速数值算法已经开发了广泛的确定性最优控制应用。 该项目继续了因果/非迭代数值方法和分布式最优控制的早期工作。

The growing availability of data on pedestrian movement allows for increasingly sophisticated models of people's goals, preferences, and perception of their environment. Such models are important not only in traffic engineering ("How do we structure the foyer in this building to avoid a stampede in case of evacuation?") but also in improving our ability to detect emerging threats in urban settings. Any dramatic change in the usual "patterns of life" might provide a clue about the evolving conditions ("Why did this group of pedestrians take such an unusual path to their target? Why is there suddenly a crowd at this intersection?"), while the analysis of data aggregated over a longer horizon can be also useful in improving our monitoring and modeling ("Which parts of the city are generally perceived as more dangerous? Where should we deploy our limited observation resources?"). This project focuses on two specific applications: (a) civilians in dangerous environments planning their paths to minimize threat exposure, and (b) adversaries aware of the existing monitoring measures planning their paths to evade the observation. In both contexts, the PIs propose models and numerical methods for (1) path planning based on one's beliefs about the environment and (2) adversarial/robust path planning, with the environment possibly changing in response to people's routing choices. The threat avoidance under selfish/independent decision making will be also treated in the framework of "mean field games". The proposed approach draws on methods from game theory, convex optimization, optimal control, and multi-objective dynamic programming. The information patterns built into this model make it possible to leverage the efficiency of fast numerical algorithms already developed for a broad range of deterministic optimal control applications. This project continues earlier work on causal/non-iterative numerical methods and distributed optimal control.