Author(s):
Daniel G. Brown* - University of Michigan
Derek T. Robinson, M.S. - University of Michigan
Li An, PhD - San Diego State University
Abstract:
We briefly describe simulation-based experiments to explore the implications of two types of empirical observations in the construction of an agent-based model. First, we analyzed the results of a social survey of residential preferences and used these analysis results to test the effects of agent heterogeneity on model behavior. Five experimental settings of distribution of agent preferences within the model were constructed and compared on several measures of spatial pattern and distributions of agent utility. The experiments included completely random agents, uniform agents, and agents with combinations of variability and categorization introduced. The results indicate that the model is sensitive to variability, but not as sensitive to categorization of agents. Next, we describe experiments designed to test the sensitivity of three statistical methods (OLS regression, logistic regression, and survival analysis), applied to spatial model output, to model parameter settings. We analyzed the ability of the statistical methods to estimate model parameters correctly, based on an analysis of model output. The results indicate that survival analysis is the best of the methods tested, and that model (and system) characteristics affect the efficacy of statistical analysis on spatial patterns for revealing agent-level parameters. We conclude with a brief discussion of the challenges of creating empirically grounded agent-based models using top-down and bottom-up methods.
