Understanding the inner-workings of city-level agent-based models

基于主体的建模（ABM）正迅速成为社会科学家的标准工具。它代表个体的能力、个体独特的特征和行为为模拟人类行为提供了模板。通常情况下，ABMs用于复制的应用系统，例如城市，都充满了复杂性和非线性动态。然而，我们缺乏工具来理解这些动态的相互作用以及它们在最终模型输出中所起的作用。如果我们要创建可用于政策领域的健壮模型，这一点尤为重要。这也是及时的；新形式的个人微数据（大数据）的增长只会增加ABM的潜力，为个人如何使用和与系统交互提供新的见解。利用从城市层面ABM的几个场景（如健康或零售相关）的输出，该博士将评估和应用机器学习（ML）中的一系列技术来揭示ABM的内部工作原理。这将使我们深入了解模型是哪些过程，哪些过程复制得不好。具体来说，目标将是：1 .对机器学习的不同技术进行批判性评估——哪种技术为理解ABM规则的相互作用提供了最大的潜力？这将包括系统的可视化方法以及统计评估将这些方法应用于ABM输出3将ML方法应用于ABM输出和关键评估。这个博士学位的输出将是一个或多个ML启发方法的应用，这些方法将阐明ABM的黑箱性质。

Agent-based modelling (ABM) is rapidly becoming established as a standard tool for the socialscientist. Its ability to represent individuals, their unique characteristics and behaviour providea template for allowing simulations of human behaviour to be created. Typically, theapplication systems that ABMs are used to replicate, for example cities, are heavily imbuedwith complexity and non-linear dynamics. However, we lack the tools to understand theinterplay of these dynamics and the part they play in the final model output. This is particularlyimportant if we are to create robust models that can be used within the policy arena. This isalso timely; the growth in new forms of individual micro-data (big data) will only increase thepotential for ABM to give new insights into how individuals use and interact with systems.Using the outputs for several scenarios such as health or retail related, from a city levelABM, this PhD will evaluate and apply a range of techniques from Machine Learning (ML)for uncovering the inner-workings of ABM. This will give insight into which processes themodel is, and is not replicating well. Specifically, the aims will be:1 Critical evaluation of different techniques from ML - which offers the greatestpotential for understanding the interplay of ABM rules? This will include bothmethods for visualisation of the system as well as statistical appraisal.2 Adaption of these approaches for application to ABM outputs3 Application of ML approaches to ABM outputs and critical evaluation.The outputs of this PhD will be the application of one or more ML inspired approaches thatwill shed light on the black box nature of ABM.