Towards Synthetic Activity-based Models for Large-scale Socio-technical Simulations

Gabriel Istrate
Los Alamos National Laboratory
D-2 Division

Activity-based approaches emerged in the 1970s and are becoming increasingly popular in Computational Social Theory, particularly in areas such as travel planning and forecasting, disease modeling and mitigation, as well as in other problems related to homeland security. Indeed, it is often the case (for instance in modeling disease transmission) that the social interaction network is the determining factor for the overall social dynamics.

Developing generic models of human behavior for urban infrastructure simulations has a number of potential benefits to simulation-supported
decision-making: (i) Such generic models could replace nonparametric, purely data-driven activity-generation in social agent simulations, enabling rapid response, quick turnaround studies. (ii) Parametric models are more readily transferred to situations when detailed data is incomplete or lacking, and can overcome the inconsistencies of commercial and publicly available data sets. (iii) Such models facilitate the computational assessment of the robustness of various general guidelines (normative properties or policies) with respect to variations in the quantitative properties of the simulations.

In a recent article (Nature, 2004), Eubank et al. undertook a study of the bipartite people-location networks arising from computational runs of EpiSims. EpiSims is a large-scale individual-based epidemiology simulation developed at Los Alamos National Laboratory, based on census, land-use, and population-mobility sample data. These networks can be used to represent the physical contact patterns that arise from movements of individuals between specific locations, and their structural properties (e.g., graph expansion) have implications for disease mitigation and control. Experimental studies have further shown that the Aiello-Chung-Lu model from random graph theory captures a number of structural characteristics of this activity-induced physical contact graph.

Our long-term research objective is to complete this approach to a full-fledged generative model of activity behavior, which incorporates temporal and activity-type information (components that are currently missing in the results of Eubank et al. In this talk, we present a number of concepts and experimental results supporting this goal.

Audio (MP3 File, Podcast Ready)

Back to Crime Hot Spots: Behavioral, Computational and Mathematical Models