EAGER: Model-Free Classification of Collective Behavior Based on Automated Detection of Symmetry from Video Data

这个探索性研究（EAGER）项目的早期概念资助汇集了一位动物集体紧急动力学专家和一位自动化科学发现专家。自动发现涉及对观测数据中潜在模式的搜索，特别是对对称性的搜索。对称性是指在重新定位或重新定向的情况下保持不变的性质，许多基本物理定律都表现出某种形式的对称性。该项目将扩展在测量运动中寻找对称性的方法，以应用于原始视频数据。这个扩展将适用于动物单独移动的视频和羊群，学校，或群体。研究结果将用于理解集体运动与个体运动之间的联系和区别，以及不同动物中与集体运动相关的共同因素。该结果将对工程多智能体系统的安全有效性能具有价值，例如自动驾驶汽车排，与人类群体互动的机器人以及四旋翼无人机群。该项目旨在开发一种定量方法，使用基于“动态类型”的自动发现算法来区分和测量多智能体系统中的集体行为，“动态类型”是根据系统的动态对称性定义的，并已用于评估一对时间序列是否实现相同的控制动态。现有算法可以对耦合动力系统的时间序列进行分析。在这个项目中，该算法将应用于动物的数据，无论是个体的还是群体的。为了适应最广泛可用的数据形式，算法将被修改为直接将原始视频作为输入。该算法将用于在选定的数值模型中探索个体与集体行为，以及来自生物群体的视频数据。最后，该算法将用于跨物种数据集，以定量地关联动物群体及其跨物种和行为的模型。这里创建的方法将构成一个强大的模型验证协议，并将提供洞察动物群体在物种内部和物种之间表现出的集体行为之间的差异。

This EArly-concept Grant for Exploratory Research (EAGER) project brings together an expert in emergent dynamics in animal collectives with an expert in automated scientific discovery. Automated discovery concerns the search for underlying patterns -- specifically, the search for symmetry -- in observed data. Symmetry refers to the property of remaining unchanged under relocation or reorientation, and many fundamental physical laws exhibit some type of symmetry. This project will extend methods for finding symmetries in measured motion to apply to raw video data. This extension will then be applied to videos of animals moving singly and in flocks, schools, or swarms. The results will be used to understand how collective motion is related to, and different from, individual motion, and which common factors relate collective motion in different animals. The results will have value for safe and effective performance of engineered multi-agent systems, such as platoons of self-driving cars, robots interacting with human groups, and swarms of quadrotor drones.This project seeks to develop a quantitative method to discriminate and measure collective behavior in multi-agent systems using an automated discovery algorithm, based on "dynamical kinds," which are defined in terms of the dynamic symmetries of a system, and have been used to assess whether a pair of time series are realizations of the same governing dynamics. The existing algorithm can analyze time series from coupled dynamical systems. For this project, the algorithm will be applied to data from animals, individually and in groups. To accommodate the most widely available form of data, the algorithm will be modified to take raw video directly as input. The algorithm will be used to explore individual versus collective behavior in selected numerical models, and in video data from biological swarms. Finally, the algorithm will be used across a multi-species dataset, to quantitatively relate animal groups and their models across species and behaviors. The method created here will constitute a powerful protocol for model validation, and will offer insight into differences between collective behaviors exhibited by animals groups, both within and between species.