Return to search

Statistical Inference Utilizing Agent Based Models

<p>Agent-based models (ABMs) are computational models used to simulate the behaviors, </p><p>actionsand interactions of agents within a system. The individual agents </p><p>each have their own set of assigned attributes and rules, which determine</p><p>their behavior within the ABM system. These rules can be</p><p>deterministic or probabilistic, allowing for a great deal of</p><p>flexibility. ABMs allow us to</p><p>observe how the behaviors of the individual agents affect the system</p><p>as a whole and if any emergent structure develops within the</p><p>system. Examining rule sets in conjunction with corresponding emergent</p><p>structure shows how small-scale changes can</p><p>affect large-scale outcomes within the system. Thus, we can better</p><p>understand and predict the development and evolution of systems of</p><p>interest. </p><p>ABMs have become ubiquitous---they used in business</p><p>(virtual auctions to select electronic ads for display), atomospheric</p><p>science (weather forecasting), and public health (to model epidemics).</p><p>But there is limited understanding of the statistical properties of</p><p>ABMs. Specifically, there are no formal procedures</p><p>for calculating confidence intervals on predictions, nor for</p><p>assessing goodness-of-fit, nor for testing whether a specific</p><p>parameter (rule) is needed in an ABM.</p><p>Motivated by important challenges of this sort, </p><p>this dissertation focuses on developing methodology for uncertainty</p><p>quantification and statistical inference in a likelihood-free context</p><p>for ABMs. </p><p>Chapter 2 of the thesis develops theory related to ABMs, </p><p>including procedures for model validation, assessing model </p><p>equivalence and measuring model complexity. </p><p>Chapters 3 and 4 of the thesis focuses on two approaches </p><p>for performing likelihood-free inference involving ABMs, </p><p>which is necessary because of the intractability of the </p><p>likelihood function due to the variety of input rules and </p><p>the complexity of outputs.</p><p>Chapter 3 explores the use of </p><p>Gaussian Process emulators in conjunction with ABMs to perform </p><p>statistical inference. This draws upon a wealth of research on emulators, </p><p>which find smooth functions on lower-dimensional Euclidean spaces that approximate</p><p>the ABM. Emulator methods combine observed data with output from ABM</p><p>simulations, using these</p><p>to fit and calibrate Gaussian-process approximations. </p><p>Chapter 4 discusses Approximate Bayesian Computation for ABM inference, </p><p>the goal of which is to obtain approximation of the posterior distribution </p><p>of some set of parameters given some observed data. </p><p>The final chapters of the thesis demonstrates the approaches </p><p>for inference in two applications. Chapter 5 presents application models the spread </p><p>of HIV based on detailed data on a social network of men who have sex with</p><p>men (MSM) in southern India. Use of an ABM</p><p>will allow us to determine which social/economic/policy </p><p>factors contribute to thetransmission of the disease. </p><p>We aim to estimate the effect that proposed medical interventions will</p><p>have on the spread of HIV in this community. </p><p>Chapter 6 examines the function of a heroin market </p><p>in the Denver, Colorado metropolitan area. Extending an ABM </p><p>developed from ethnographic research, we explore a procedure </p><p>for reducing the model, as well as estimating posterior </p><p>distributions of important quantities based on simulations.</p> / Dissertation

Identiferoai:union.ndltd.org:DUKE/oai:dukespace.lib.duke.edu:10161/8687
Date January 2014
CreatorsHeard, Daniel Philip
ContributorsBanks, David L
Source SetsDuke University
Detected LanguageEnglish
TypeDissertation

Page generated in 0.0017 seconds