New Methods for the Steady-State Analysis of Complex Agent-Based Models
New Methods for the Steady-State Analysis of Complex Agent-Based Models
复制标题
DOI：
10.3389/fphy.2020.00103
复制
发表时间：
2020-04-08
影响因子：
3.1
通讯作者：
Camargo, Chico Q.
中科院分区：
文献类型：
--
作者：
Camargo, Chico Q.
在用于理解人类集体行为的所有工具中，很少有工具像基于代理的模型(ABM)那样成功。这些模型在描述突发的社会行为方面尤其有效，比如社区的空间隔离或社交网络上的意见分化。ABMS在观点和信念动力学的研究中特别常见，被从人类学到统计物理学的各个领域使用。这些模型很像它们所描述的社会系统，往往没有唯一的输出变量、规模或明确的顺序参数。这种缺乏明确可衡量的紧急行为使得这种复杂的ABM很难研究，最终将它们的应用限制在具有经验意义的案例中。本文从信息论和聚类分析的角度出发，介绍了一系列分析复杂多维ABM的方法。我们使用这些方法来探索Banisch和Olbrisch提出的基于多层次主体的意识形态结盟模型，以扩展马斯和弗拉奇的双极化论点交流理论。我们使用这里介绍的工具来对小系统规模的模型进行彻底的分析，识别向稳态行为的收敛，并描述由该模型产生的稳态分布的全谱。最后，我们展示了我们引入的方法如何容易地适用于更大的实现，以及其他复杂的基于代理的社会行为模型。
Among all tools used to understand collective human behavior, few tools have been as successful as agent-based models (ABMs). These models have been particularly effective at describing emergent social behavior, such as spatial segregation in neighborhoods or opinion polarization on social networks. ABMs are particularly common in the study of opinion and belief dynamics, being used by fields ranging from anthropology to statistical physics. These models, much like the social systems they describe, often do not have unique output variables, scales, or clear order parameters. This lack of clearly measurable emergent behavior makes such complex ABMs difficult to study, ultimately limiting their application to cases of empirical interest. In this paper, we introduce a series of approaches to analyze complex multidimensional ABMs, drawing from information theory and cluster analysis. We use these approaches to explore a multi-level agent-based model of ideological alignment introduced by Banisch and Olbrisch to extend Mas and Flache's argument communication theory of bi-polarization. We use the tools introduced here to perform a thorough analysis of the model for small system sizes, identifying the convergence toward steady-state behavior, and describing the full spectrum of steady-state distributions produced by this model. Finally, we show how the approach we introduced can be easily adapted for larger implementations, as well as for other complex agent-based models of social behavior.