Statistical calibration of agent-based models of transport from aggregate data

本文主要研究基于聚合数据的基于agent的交通运输模型的统计标定问题。ABMs旨在通过一套概率规则来理解相互作用的个体选择（如路线旅行选择）对紧急行为（如网络流量模式）的影响。ABMs最近在流行病建模中越来越受欢迎，这引发了关于其可识别性和预测能力的重要问题，特别是在政策情景测试中（例如引入步行街）。该研究旨在开发计算统计学和机器学习技术，通过从粗略观察的数据中理解离散的个体代理选择来解决这些问题。这带来了巨大的计算挑战，因为它需要探索智能体决策的组合空间。从种群综合和ABM模拟两方面解决了这一问题。前者是一个预处理步骤，创建人工智能体，其社会经济属性与人口平均所描述的具有相同的统计属性，而后者使用粗数据重建由离散智能体选择和/或连续决策参数产生的智能体时空轨迹。

This research focuses on statistical calibration of agent-based models (ABMs) of transport based on aggregate data. ABMs aim to understand the effect of interacting individual choices (e.g. route travel choices) on emergent behaviour (e.g. network traffic patterns) via a set of probabilistic rules. ABMs' recent rise in popularity in epidemic modelling raised important questions about their identifiability and predictive power especially in policy scenario testing (e.g. introduction of pedestrianised streets). The research aims develop computational statistics and machine learning techniques to address those questions by understanding discrete individual agent choices from coarsely observed data. This poses great computational challenges as it necessitates the exploration of a combinatorial space of agent decisions. The problem is tackled in two contexts: population synthesis and ABM simulation. The former is a pre-processing step that creates artificial agents whose socio-economic attributes have the same statistical properties as those described by population averages while the latter uses coarse data to reconstruct agent trajectories in space-time generated by discrete agent choices and/or continuous decision parameters.