252

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 36, NO. 2, MARCH 2006

BioWar: Scalable Agent-Based Model of Bioattacks
Kathleen M. Carley, Douglas B. Fridsma, Elizabeth Casman, Alex Yahja, Neal Altman,
Li-Chiou Chen, Boris Kaminsky, and Démian Nave

Abstract—While structured by social and institutional networks, disease outbreaks are modulated by physical, economical, technological, communication, health, and governmental
infrastructures. To systematically reason about the nature of
outbreaks, the potential outcomes of media, prophylaxis, and
vaccination campaigns, and the relative value of various early
warning devices, social context, and infrastructure, must be considered. Numerical models provide a cost-effective ethical system
for reasoning about such events. BioWar, a scalable citywide
multiagent network numerical model, is described in this paper.
BioWar simulates individuals as agents who are embedded in
social, health, and professional networks and tracks the incidence
of background and maliciously introduced diseases. In addition
to epidemiology, BioWar simulates health-care-seeking behaviors,
absenteeism patterns, and pharmaceutical purchases, information
useful for syndromic and behavioral surveillance algorithms.
Index Terms—Bioterrorism, multiagent network, social network, syndromic and behavioral surveillance.

I. I NTRODUCTION

T

HE capability to assess the impacts of large-scale biological attacks and the efficacy of response policies is
necessary from intelligence and planning perspectives and requires reasoning about social response and disease transmission
within a complex social system. The recent case of an atypical pneumonia [severe acute respiratory syndrome (SARS)]1

Manuscript received November 26, 2003. This work was supported in part
by the Defense Advanced Research Projects Agency (DARPA) for work on
Scalable Biosurveillance Systems, by the National Science Foundation (NSF)
IGERT9972762 in CASOS, by the MacArthur Foundation, and by the Carnegie
Mellon Center on Computational Analysis of Social and Organizational Systems. Computations were performed on the National Science Foundation Terascale Computing System at the Pittsburgh Supercomputer Center. Any opinions,
findings, and conclusions or recommendations expressed in this paper are those
of the authors and do not necessarily reflect the views of DARPA, the National
Science Foundation, or the U.S. Government. This paper was recommended by
Associate Editor T. Erickson.
K. M. Carley is with the Center for Computational Analysis of Social and
Organizational Systems, Carnegie Mellon University, Pittsburgh, PA 15213
USA (e-mail: kathleen.carley@cmu.edu).
D. B. Fridsma is with The Center for Biomedical Informatics, University of
Pittsburgh Medical Center, Pittsburgh, PA 15213-2582 USA.
E. Casman is with the Engineering and Public Policy Department, Carnegie
Mellon University, Pittsburgh, PA 15213 USA.
A. Yahja is with the Institute for Software Research International, Carnegie
Mellon University, Pittsburgh, PA 15213 USA.
N. Altman and B. Kaminsky are with the Carnegie Mellon University,
Pittsburgh, PA 15213 USA.
L.-C. Chen is with the Department of Information Systems, Pace University,
New York, NY 10176 USA.
D. Nave is with the Pittsburgh Supercomputing Center, Pittsburgh, PA 15213
USA.
Digital Object Identifier 10.1109/TSMCA.2005.851291
1 http://www.who.int/csr/sars/country/2003_07_11/en

illustrated the importance of and close linkage among social
networks [1], [2], disease transmission, and early detection.
Like natural epidemics, biological attacks will also unfold
within spatially defined complex social systems, and the societal response will have profound effects on their outcome.
It is not always clear how to best detect and respond to a
disease outbreak, either natural or malicious. The goal of our
research is to develop tools to simulate how diseases spread
through socially connected groups so that these tools may be
used to test the various detection and response options. This
paper focuses on bioterrorist attacks, but the model structure
has been applied to emergent and familiar diseases as well.
In trying to prepare for attacks, policy makers need to be able
to think through the consequences of their decisions in various situations. Role-playing “simulation” physical exercises
can provide valuable insights, but they are limited in number,
size, scope, and scenario due to cost, time, and cognitive constraints [3].
Designed and validated correctly, computer simulations can
be more cost effective, faster, and more comprehensive, allowing enormous numbers of complicated outside-the-box scenarios to be examined systematically.
Recently, geographic-information-system (GIS)-based epidemiological models that take into account the spatial and
geographical dimensions have been developed [4], [5]; however, this paper does not consider the social dimension nor the
interplay between different dimensions that would affect the
complex outcome of the interacting real-world systems. There
has also been relevant work in RoboCup rescue [6], swarm/ants
intelligence [7], and mathematical modeling of multiagent systems [8], [9]. While RoboCup rescue is useful for investigating
rescue strategies in a simulated earthquake, it is concerned with
designing smart algorithms, not with investigating a current
human social system as it exists and designing a public policy for it. The swarm/ants intelligence work focuses on the
emergence of smart group behaviors out of simple individual
routines (e.g., for those of insects such as ants), instead of
on knowledge-intensive and context-intensive human social
system and the effects of epidemics within it. The mathematical
modeling of multiagent systems is an idealized approach to
modeling social systems that is good at describing population
dynamics. It concerns itself with deriving equations to describe
macrobehaviors from microbehaviors, often needing to make
assumptions such as the use of a generalized Markov model,
which are not necessarily sufficient to describe social systems.
BioWar is a single integrated model of the impact of a bioterrorist attack on a city that combines state-of-the-art computational models of social networks, communication media,
and disease transmission with demographically resolved agent

1083-4427/$20.00 © 2006 IEEE

Transportation networks may constrain but do not define social networks. however. or individual response to symptoms and diseases. BioWar still requires expert users: we are in the final phase of developing a simplified graphical user interface implementation. stakeholder communication. and other publicly available information. diagnosis. financial security and liberty genes). and hospital visit rates. System dynamics models such as the Epi-Engine of the National Bioterrorism Civilian Medical Response Center (CiMeRC) abstract away the underlying social interactions with a system of mathematical equations. It models individual attack diseases alone without concurrently simulating normally occurring diseases and does not properly model disease progression. The complexity and nuance of social interactions.CARLEY et al. Cellular-automata models for disease spread such as the Brookings’ individual-based computational model of smallpox epidemics [12] improve upon the differential model of SIR. Nonbinary and non-Markovian input and output values are not modeled. Moreover. local. medical website hits. Purdue University’s Measured Response simulation model is based on Synthetic Environment for Analysis and Simulation (SEAS). assuming homogeneously mixing populations. and public policy analysis. Though providing useful insights. including the coevolution of cognition and structure—a factor in disease spread—are overlooked. among others. The Measured Response model has been used for annual training exercises of the same name with local. The geometry of cellular automata. which may lead to incorrect results. Its hypothetical simulations of a stylized city are inadequate in fidelity. Measured Response applies population-based mathematical models as determinants on agents. Army. Episims also models disease spread by disease load. The Measured Response model and its model assumptions and variables have not been validated. . recreational preferences. [16]. and national). medical phone calls. over-the-counter pharmacy purchases. Measured Response imposes these levels onto its agent model by using small numbers of agents to represent large populations but does not provide adequate scientific justifications as to what level of abstraction and how much aggregation and layering are feasible without reducing its agents into simple homogeneous sample populations. it did not model the social interactions and physical dimensions of disease spread and response. even while being constrained by them. This enables observation of the repercussions of various attacks and containment policies through natural measures such as absenteeism. response analysis. L IMITATIONS OF P REVIOUS E PIDEMIOLOGICAL M ODELS Epidemiologists have used the susceptible–infected– recovered (SIR) framing for modeling the course of epidemics [10]. The model represents agent inputs/sensors as binary “genes” and agent outputs/ 2 http://www. no social (and network) connections [11]. Inc. B IO W AR O VERVIEW BioWar is a city-scale spatial multiagent network model capable of simulating the effects of weaponized biological and chemical attacks [17]. and they largely ignore the symbolic aspects of a population such as knowledge about school districts. partially because it does not have proper spatiotemporal coordinates. its disease model has not been validated against empirical data of past outbreaks. Such models are typically implemented.g. transportation patterns. including Charles River Analytics. spatial regions. they are problematic for the subtleties of micro. symptoms. no spatial dimension. and a diagnostic error model. disease confounding. the calendar of events and holidays.: BIOWAR: SCALABLE AGENT-BASED MODEL OF BIOATTACKS models. land use. with no medical intervention. instead of allowing for emergence. II. It integrates principles of epidemiology.g. Episims is better than other methods in that it does not assume homogeneous population mixing. To meet the requirement to simulate at varying scales (e. Unlike traditional models that look at hypothetical cities.html 253 actions/decision factors as binary “genes” (e. detection algorithms evaluation. The model has a routine that determines the mapping between its input and output attributes/genes and updates the state of its genes. A discrete event simulation model of antibiotic distribution was used to examine postexposure prophylaxis [13]. oversimplifying disease propagation processes.and macrobehaviors of agents crucial to the effective response against a bioterrorism attack. state.and mesobehaviors. weather models. ignoring the major raison d’être of the multiagent simulations: emergence of high-level behaviors and entities from low-level primitives. Versions of BioWar have been made available to several outside groups.. and it has a good interactive interface. and federal officials as participants.org/content/projects. and traffic regulations. school district boundaries. allowing spatial operation and discontinuities. and the U. while the reverse is true in reality: Agent behaviors drive social networks and social networks drive. This is relevant to the ability to examine microand emergent meso. BioWar is configured to represent real cities by incorporating census data.. But it lets normal nonattack day transportation patterns drive social networks and agent behaviors. urban spatial models. rather than just providing information on the number of infections. BioWar is useful for clients concerned with preparedness training. and spatial operations such as proximity and adjacency. forgoing the benefits of agent-based simulations. The intended user base of BioWar includes policy analysts and developers representing such clients. and census data to estimate contact graphs that are assumed to be social networks [15]. a deoxyribonucleic acid (DNA)-like agent model for a synthetic economy [14]. in particular. Los Alamos National Laboratories’ Episims applies graphbased methods to a transportation networks simulation based on population mobility. as transportation networks and other infrastructures provide a large number of choices for agents. III. among others. does not match the spatial reality of the real world. As it presently stands. treatment. state. BioWar models the agents as they go about their lives—both the healthy and the infected. the model often uses abstract/high-level entities and behaviors as genes. However. and no symptom-based behavior. insurance claims.2 While system dynamics models capture the general trend of epidemics and feedback loops.cimerc.S.

and effective responses in general. The arrows between “privacy” and “output and system compliance” in Fig. school districts. Outputs are used to test early detection algorithms and hypothetical what-if scenarios are used to analyze policy. While BioWar incorporates elements from previous models. health science. 1. social. 4) locally accurate agent and social network models with good disease and infrastructure models that facilitate high-precision detailed investigation of first-responder plans for bioterrorism response. geography. A simulation is only as good as its fidelity to the aspects of the real world that it tries to emulate. disease. work patterns. organization science. real data from census.g. Agent.) without explicit aggregation but with emergence [24]. local government officials. dispersion. What-if scenarios and impact analysis modules have also been encoded and used to validate BioWar against the SIR. The agent model takes as input the social. concurrent partnerships affect disease transmission dynamics in networks [19].. general social surveys. bridge populations have influence on the spread of the human immunodeficiency virus (HIV) in Thailand [21]. and communication technology (e. detailed. It integrates various models of disease. geography. [22]. BioWar is socially (with social. The detection module uses detection algorithms created by outside groups attempting to detect a bioattack 1 or more days earlier. 36. etc. etc. knowledge evolution. Individual characteristics of the agents. [18]–[20]. and task networks) and spatiotemporally more realistic than previous bioterrorism attack models. health-related website visits and 911 calls). . VOL. diseases. weather. manifest symptoms. The Behavioral Surveillance Surveys (BSSs) and National Electronic Disease Surveillance System (NEDSS) compliant privacy module is being implemented in cooperation with the Carnegie Mellon University (CMU) Data Privacy Laboratory. and environmental data streams. and seek and receive treatment. and aerosol transport. Design of BioWar. attack. 1 denote the use of privacy algorithms to ensure the security and anonymity of individuals. The agent model itself is a finite state machine describing how agents interact. demographics. and communication technology models have been implemented and used extensively. behavioral science. and environmental conditions are used by the agent model to change behavior. discharge diagnosis. the complex social networks are not shown. are used as input. BioWar also generates simulated output that is validated against real data: drug purchases. and general information on communication technology.. 3) detailed modeling and simulation of individuals and their social networks. AND CYBERNETICS—PART A: SYSTEMS AND HUMANS. temporal. social networks. Major components of BioWar are illustrated in Fig. epidemiological curves. Major advantages of BioWar’s spatial multiagent network approach include: 1) heterogeneous population mixing. geographic. defined by agent and social network characteristics. attack. geography. 2. disease. concurrent partnerships affect syphilis persistence [18]. first responders. 2) simultaneous modeling at multiple levels (pathogen. sociology. disease. In BioWar. medical phone calls. emergency room (ER) visits. Note that the agent model diagram only represents a small part of the complexity of agent behavior. social network statistics. a conventional epidemiological model for smallpox [25] and against a variant of SIR [26] for noninfectious diseases (in this case anthrax). and networks intertwine with epidemiology [20]. etc. doctor visits. move. and demographic dimensions affect disease spread [4]. contract disease. are exposed to bioagents. NO. it also addresses many of their shortcomings. MARCH 2006 Fig. weather. For example. 1. Recent work has demonstrated that the failure to take the social network and the physical locations of people into account leads to incorrect estimates of disease spread and of response policies because spatial. and fast. sexual networks affect HIV spread [23]. knowledge. chief complaints.254 IEEE TRANSACTIONS ON SYSTEMS. school and work absenteeism. MAN.

its severity is initialized to the symptom’s evoking strength. Agents with symptoms self-diagnose. Then. however. Each disease has its own set of symptoms. and the phase durations are calculated differently [25]. Noncontagious diseases do not have this phase. 3 http://www. is generally determined using a uniform random distribution with a range provided by expert analysis for background diseases (better distributions may be added when the need arises). In the current version of BioWar. is a qualitative measure of how frequently a doctor will associate a particular symptom with a particular disease. risk factors are distributed a priori to individuals in the population according to demographic characteristics such as age. We use a symptombased general disease model. and treatment. This symptombased disease model permits the representation of outliers and stochastic flux (not everyone with the same disease presents the same symptoms).: BIOWAR: SCALABLE AGENT-BASED MODEL OF BIOATTACKS IV. In many diseases. visit their doctor or pharmacist.gov/ncidod/EID/vol8no10/contents_v8n10. This phase may overlap with the above phases. Validating a new disease is harder: It may take days or weeks depending on the disease and the quality of empirical data. occupation. how often does one see symptom y?” For example. as the disease progresses. Evoking strength is coded as a number between 0 and 5. when not affected by treatment. Many diseases have no known or identifiable early symptomatic period. the transmissibility of the disease. 4) Communicable: The period of time during which an infected agent may infect other agents. patients with the diagnosis of anthrax will have a fever frequency of 5—nearly all patients with anthrax will have fevers at some point in the course of their disease. the total severity is determined quasi-randomly based on evoking strengths [31]. is a qualitative measure of how frequently people with a particular disease will manifest a particular symptom.. Weaponized diseases. We have also drawn on the experience of other medical expert systems developed to assist in diagnosis to ground our disease model in well-founded medical knowledge representations [31]. sex. timing for disease phases. 2) Early symptomatic (prodromal): The period of time during which an infected agent may experience mild or nondescriptive symptoms. the evoking strength for this symptom is 5. gender. 5) Recovery/death: A period of time during which a disease resolves or causes death.htm 255 A. thus. Frequency is denoted by a number between 1 and 5 that answers the question: “In patients with disease x. and contagiousness. Agent Behavior–Symptom Relationship Although individuals get symptoms based on the symptom frequency for the disease that they have. emerging infectious diseases like SARS) is relatively easy: It takes approximately one programmer day.3 and disease knowledge bases [28]–[30]. In BioWar. are treated as a special case. documents from the October 2001 bioterrorism attack. Once an agent is infected. B. Each instance of a disease infecting an agent is individually represented and progresses through time as the agent goes about his or her daily business. The second. except recovery/death. a widened mediastinum is a more specific symptom of anthrax—in patients who have a widened mediastinum. fever symptoms are not specific to any one disease—in our disease profile of anthrax. race. the severity is reduced each tick randomly by an amount depending on the type of treatment. evoking strength. Thus. Symptoms Symptoms are important in BioWar on two levels. These risk factors increase a person’s susceptibility to diseases through either host factors or environmental factors to which that person is exposed. they will actually alter their behavior based on the evoking strength of that symptom. However. Adding a new disease (e. When a symptom is added to a disease. [26]. . based on probabilistically determined agent risk factors. In constructing our disease model. is determined by a Bernoulli process with p equal to the death rate of the disease among untreated victims (again. how often would a doctor think the cause is disease x?” For example. depending on the severity of symptoms. Diseases can propagate through a population.cdc.CARLEY et al. the length of each phase. 1) Incubation: The period of time before the agent begins presenting symptoms. Symptoms are assigned two different measures that influence which symptoms agents feel and how that changes their behavior [31]. Risk Factors Certain demographic groups are more likely to be susceptible to particular diseases than other groups. and the spatial and temporal proximity of uninfected agents to infected agents. we used historical accounts of known anthrax releases [27]. 3) Late symptomatic (manifestation): The period of time during which an infected agent may experience severe and/or disease-specific symptoms.g. and race. If an agent is in treatment. variations in presentation based on age. It answers the question: “When you see symptom y. and disease prevalence. Recovery and death of an agent. A GENT -L EVEL D ISEASE M ODEL The current version of BioWar simulates 62 diseases— 4 weaponized diseases and 58 naturally occurring diseases— simultaneously in a population. this phase may not be distinct from the early symptomatic phase. and change their patterns of interacting with other agents. the possibility of anthrax should be routinely considered. C. the severity of each symptom is increased randomly each 4-h tick by a user-defined value. time. the infection progresses through time via a simple state machine that can be interrupted or altered by external events such as successful treatment of an agent at a medical facility. stay home from work. The duration before death or recovery is likewise stochastically determined. frequency. determined by expert analysis). Each disease progresses through up to five phases. fever is given an evoking strength of 1. They motivate agent behavior and determine the initial diagnosis of agents entering the medical system. The first.

and onset-of-illness versus inhaled dose was recently published for anthrax.S. we will incorporate detailed rule-based systems for specific symptoms-based behavior changes. this triggers the intervention such as distribution of the antibiotic Cipro. Noncontagious diseases do not have a communicable phase. self-medicate (Table I). which can include laboratory tests of varying accuracy (types 1 and 2 errors are possible) and report time. information about human response modeling for weaponized diseases is scarce. Our goal was not to build an error-free diagnosis model.. For short-duration noncontagious diseases such as food poisoning. As past medical history affects these transitions. Contagious Disease Submodel A set of user-specified symptom severity thresholds guides an agent’s initial decision to visit a medical facility. For chronic noncontagious diseases such as angina and diabetes. outbreaks are randomly generated based on prevalence data. Blaser. which allows for the possibility of initial misdiagnosis and the revision of diagnoses with additional information (e. disease transmission occurs with some specified probability. NO. and the state can be changed. a medical intervention can occur. 4) extreme severity—go to the ER. and Venezuelan equine encephalitis [32]. anthrax spread by U. V. this is a non-Markovian model. VOL. To determine which disease a person has. which is consistent with observed hospital performance [36]. the total symptom severity is calculated and measured against each threshold in the following order: 1) low severity—no effect. Moreover.256 IEEE TRANSACTIONS ON SYSTEMS. Dose–Response Relationship for Weaponized Pathogens Currently. Hence. MAN. Agents self-diagnose on the basis of visible or palpable symptoms. 36. We based the model on the Internist1/Quick Medical Reference (QMR) diagnosis model [31]. illness duration.g. but it serves to control the simulator’s response to diseases in a simulated population. . our model is not a true computational diagnostic tool. laboratory results). [34]. Transmission can occur via contact or air dispersal. the initial agent population is afflicted based on known race. to doctor offices— and compared with the simulated visit statistics that BioWar produces. 3) high severity—go to the doctor.g. but we have augmented the results with probabilistic “switches. and age distributions according to prevalence information. pneumonic plague. another agent in the same demographic cohort chosen at random is afflicted. D. In the future versions of BioWar. Chief complaints are not necessarily the same as discharge diagnosis. and the highest one is chosen as the diagnosed disease. BioWar takes into account dose and age of agent for infection probability for inhalational anthrax following formulas developed by Webb. If an agent dies. botulism. E. the disease most strongly associated with the most severe set of symptoms is chosen. These behavior thresholds for going places are currently estimated from empirical visit statistics—say. in turn. Mail). If anthrax infection is suspected to be present.. Our disease model also allows the representation of contagious diseases. we use differential diagnosis. The state of the disease also affects the medical intervention. Noncontagious Disease Submodel Our disease model tracks noncontagious diseases. Subsequent diagnosis can update the primary diagnoses based on the appearance of new symptoms and on the results of diagnostic testing. as do medical doctors. Agents experiencing disease state transitions are modeled as nondeterministic automata. When a person comes into contact with the transmission medium. 2. while for the other weaponized diseases. Initial medical diagnosis is simulated based on the apparent symptoms and their evoking strengths. Intervention affects disease outcome. The relationship between the infection probability. and Buckeridge [33]. though some noncontagious diseases can be spread by contact (e. Before the decision model chooses an agent’s next care-seeking behavior. and the agent only responds to symptoms that can be monitored by the normal human senses. delaying institutional realization of a bioattack. the true nature of the agent’s health status is opaque to the decision model. doctors and ER personnel take time to file a disease report (up to 8. ameliorates the symptoms and possibly cures the disease. At any time within the duration of a state. and the visit outcome is probabilistically determined based upon an agent’s demographic profile. F. we use a symptom-based differential diagnosis model to obtain information on the diseases infecting an agent who visits a medical facility. At this stage. gender. AND CYBERNETICS—PART A: SYSTEMS AND HUMANS.” As such. it uses an exponential model for infection probability. The thresholds are limits of the sum of the severities of observable symptoms over all diseases infecting an agent. In other words. Medical personnel diagnose on the basis of visible symptoms and other information. rather. 2) mild severity—go to the pharmacy.74 days according to a claim delay data gathered by IBM [35]). the groups of evoking strengths of symptoms associated with potential diseases are compared. MARCH 2006 TABLE I SYMPTOM TRIGGERED MEDICATION PURCHASE Giving Cipro. D IAGNOSIS M ODEL As previously mentioned.

The principle of expertiseseeking states that the information poor are more likely to initiate communication with the information rich to fulfill their information needs. and the one for smallpox is in [25]. an agent has sociodemographic and economic status. infecting the partner. • voluntary (friend. institutions.4 The overall network size and distribution were drawn from Klovdahl’s study along with some target numbers for individual relationship counts [1]. S OCIAL N ETWORKS Epidemiologists have long recognized that groups. Treatment may not be immediately effective. Do the same check for the agent for each communicable disease. sex. social network. The propensity of an agent to seek treatment is affected by sociodemographic position (age. nurse. B. social class. economic status. The principle of homophily states that people are more likely to communicate with others who are similar to them. BioWar implements disease exchange as the result of agent interaction as follows: For each partner of an agent (computed during the interaction step). BioWar implements multiple agent interactions based on common knowledge and knowledge difference. If an agent reports directly to a hospital’s emergency department. VII. a population not covered by the GSS. Agent Interaction Agents interact with each other based on BioWar’s CONSTRUCT model [40] and on spatial and social network proximities. parent. the interaction is mediated by social. Note that even if an agent seeks treatment. an agent can exhibit symptoms and has mental model of diseases. child. agent’s customary locations. group member. While noncontagious bioagents such as anthrax do not spread through social networks. recovery after treatment is modeled as a lognormal distribution with a mean of 3 days with a range of 0–10 days for all diseases except anthrax and smallpox. going to school. and other networks. Social networks are also important in delimiting bioattacks using noncontagious microorganisms. recreation. or an admission to the hospital. Moreover. The construction of social networks begins by defining the ego net for each agent based on empirical data on the size and constitution of networks. belief. gender. Additionally. organizations. Relationship types currently implemented are as follows: • family (spouse. getting medical information. get infected. purchasing over-the-counter medicines and cold supplies. Agents may be linked unidirectionally or bidirectionally.).icpsr. The core of the CONSTRUCT model encapsulates the coevolution and emergence of communication and knowledge networks. Future versions will have more realistic treatment models for more diseases. schoolmate. prestige. neighbor). and moving to other places. The relationship types were drawn from the University of Chicago General Social Services (GSS) survey data with the addition of “schoolmate” for younger agents. driven by homophily and expertise-seeking. The significance of social networks to contagious disease transmission is obvious. and agent’s state of health. occupation. and attitudes.edu:8080/GSS/homepage. etc. Each link is a descriptor of a pair of agents and their relationship type. An ego net is a set of agents with whom an agent primarily interacts. C. educational level. It is only recently that we have had the tools for modeling these systems. These tools include multiagent computer models and the body of statistical tools and measures that have arisen in social networks [37].umich. infect). class. frequency of relationship type. An agent is located at specific spatiotemporal coordinates and exhibits behaviors.htm 257 TABLE II SOCIAL NETWORK SIZE AND RANGE Table II. Factors considered during the construction of social networks include target network size for an agent. teacher. [2]. tests. interests. going to work. The recovery distribution for anthrax is described in [26]. doctor. etc. and symptoms vary in visibility and type of testing required for their detection. sleeping 8 h a day). Representation of an Agent An agent is represented as a probabilistic finite state machine that has roles such as father. race. [39]. advisor.: BIOWAR: SCALABLE AGENT-BASED MODEL OF BIOATTACKS VI. In the current version of BioWar. Moreover. infect the partner with the disease. CONSTRUCT has been validated multiple times using social and organizational data [40]. A. schoolmate. and the societies in which they are embedded. seeking treatment. family. other). and random chance. T REATMENT M ODEL Diagnosis at a doctor’s office results in treatment or ordering of additional tests. other family). These behaviors include interaction (communicate. [41]. and vice versa. race. drawn from the results of “Challenge 3” version of BioWar. throw a random die against the transmissibility of each communicable disease affecting the agent. Representation of Social Networks in BioWar In BioWar. and severity of perceived symptoms. [38]. shows the size of constructed social networks for three simulated cities compared with empirical data. are complex systems that affect the propagation of diseases through a population. culture. diagnosis results in treatment. sibling. professional. agent demographics. each agent is linked to other agents in the simulated population by a set of links (relationships) modeling a social network. Similarity in CONSTRUCT is assessed by attributes such as age. social networks define the exposed subpopulation through colocation of agents at the time and place of an attack.CARLEY et al. If the die roll succeeds and the partner does not already have the same disease. 4 http://www. . • proximity based (coworker. Each agent has an ego net and natural biological rhythm (for example.

Representation of Weather treatment is not always available such as when doctor offices are closed.gov/oa/ncdc. these tables are set to preserve overall rates: Recreation that is deferred is taken later and recreation taken early reduces the chance of recreation later in the day. NO.5 Consistent with the data. VA.berkeley. Pittsburgh. σy and σz are the dispersion parameters that are functions of downwind distance x. and H is the height of the release in meters. 36.258 IEEE TRANSACTIONS ON SYSTEMS. D. u is the wind speed in meters per second. For the current BioWar implementation. 2) Bioagent Delivered Dose: The dosage inhaled by the agent is calculated using [43]. or in hospitals). The adjustment table allows recreational rates to be adjusted to match the normal day cycle: Little recreation occurs during hours of sleep and work. the wind speed is constant with height.edu:7502/cgi-bin/hsda?harcUMD+time The weather model determines atmospheric temperature.html . etc. Paired tables were employed: one holding the overall recreational probability and the second determining recreational location. Recreational Activities The recreation code simulates some of the additional activities people engage in beyond their normal routine (time spent at home. in the evenings during the week). The wind model also includes methods for determining whether the agent is located in the downwind zone and how far the agent is from the point of the release. PA. that is. VOL. which is considered to be a reasonable approximation [42]. at pharmacies. our wind model calculates the dependence of wind speed on height. 2. B = 5 ∗ 10−4 m3 /s).ncdc. but it does not place the agent at a recreational location. an adjustment table was also introduced. work or school) and medical concerns (time spent with doctors. Wind is important at and after the moment of the attack. The EPA Time Use data set is sufficiently large to allow some seasonal. The recreation code adds several additional types of recreation venues where agents gather and potentially interact on a more random basis (Table III). for light work. there are tables for minor and major holidays. 6 http://www. E NVIRONMENTAL M ODELS A. the information was derived from annual averages for Saturday. When multiple ground or airborne releases are simulated. MARCH 2006 TABLE III RECREATION VENUES VIII.).. 2) The lateral and vertical variations of the material concentration can both be described by Gaussian distributions. Buildings provide some measures of protection to inhabitants from outside spores. 5 http://sda. The assumptions of the model are as follows. Recreation rates were derived from the 1994 Environmental Protection Agency (EPA) Time Use Survey by grouping activity categories to determine the percentage of the day normally spent in recreation and using the time-at-location data to make a determination as to how much time is spent in each recreation location. and demographic resolution. CA. Generated temperature. and precipitation for the period of simulation. Each seasonal table contains separate entries for each day of the week. B is the breathing rate (usually. which are functions of downwind distance only. pressure. we assign a Pasquill atmosphere stability category based on the wind speed and time of the attack but not the sky condition. especially when the attack occurs outdoors and the biomaterial is dispersed though wind puff movement. weekly. Since the Time Use database excluded holidays. Norfolk. for young versus old agents and for male versus female. so BioWar has a dose reduction factor to describe the building protection for different kinds of building based on [45].noaa. Because the simulator divides each day into smaller time units (4-h “ticks”). These tables were derived for stereotypical patterns and are not empirically based. and more occurs when the agent would logically have free time (for example. recreation tables were constructed for the four seasons. An essential function of the wind model is to assess the Pasquill atmosphere stability category for the period of an attack. In order to reflect the empirical data. much recreation is assumed to occur at the agent’s home. the total effect on the agent (the summary dosage) is calculated as a sum of dosages from individual releases.6 1) Wind Representation: The wind model generates wind speed and direction for the period of simulation.g. pressure. [44] Dose = [QB][πuσy σz ]           2 2 1 1 y H exp − − 1 exp − 2 σy 2 σz where Q is the source strength (e. AND CYBERNETICS—PART A: SYSTEMS AND HUMANS. number of anthrax spores). recreation is subservient to other more important activities. An agent’s routine activity is mapped to a number of set locations. 1) The dispersed biomaterial is chemically stable and is not deposited to the ground. and precipitation yearly distributions closely match the historical data published by the National Oceanic and Atmospheric Administration for the simulated regions (San Diego. Building layouts could be used in future versions of BioWar to improve the estimation of delivered dose of both outside and inside attacks by applying more complex calculations. In the absence of detailed meteorological data. We use a modified Gaussian Puff model of wind dispersion. 3) Although in the simplest Gaussian model. MAN. In addition. The recreation code’s role is advisory: It indicates that an agent “wants” to recreate.

CARLEY et al. Note that the specification here is only for the attack or weaponized diseases. distributing seven bombs along an attack line of 1. and others). The generated wind speed and direction distributions closely match the empirical data for the simulated regions published by the EPA and averaged over 3 years. waterborne. means of attack (land or airborne).1 300 m 1.7) processors.usgs. containment (inside or outside of a building).html 11 http://www. A module conceptually corresponds to a BioWar simulation capability.gov/pubsearch 14 http://www. and Tru64 UNIX.ed. The BioWar disease model has different specifications for the naturally occurring background or outbreak diseases. A module in the system interacts with other modules via a published interface of methods. C URRENT I MPLEMENTATION BioWar is designed as a modular system. The relative difference between simulated and average historical frequency distribution values is less than 30%.gov/nchs 7 http://www.html 15 http://www. which comprises 64 four-way Alpha SMP processing elements (PEs). which generates an outside. 2. BioWar’s run time scales linearly with the number of agents. X.cdc. and single-point or multiple-point attack. at 22:00 using 25 kg of material for the attack at 10% efficiency. facilitates the rapid development of reliable and more realistic simulation models.html 10 http://www. An example of the attack specification in BioWar is out large anthrax_inhalational 2002/7/4 22 : 00 25 kg .gov/publications.census.ed.5 km 7.gov 16 http://www. IX.noaa. The agent model. A TTACK M ODEL BioWar has a flexible attack model for both contagious and noncontagious pathogen release. release height and efficiency. Comparison between simulated wind direction frequency distribution and average historical 1990–1992 data for San Diego. consists of both an agent behavior module and an agent social network module.7 A comparison between the simulated wind direction data for San Diego and the historical 1990–1992 average data is shown in Fig.5 km. Windows 2000 and XP. A simulation with about 500 000 agents can take 5 or more hours to complete. agent carrier (airborne. 2002. inhalational anthrax attack on July 4.4. with airborne delivery at an altitude of 300 m.cdc. time of day. BioWar is designed to be reasonably portable. This flexibility. The model lets attacks be varied by location.htm .ncdc.epa. like disease progression and diagnosis. Table IV shows some of the cartographic and other input sources used [31]. Higher fidelity data variants of BioWar are available to health and government authorities with appropriate access privileges. 8 http://geonames.census. This correspondence is not necessarily one to one. large. CA. each with 4 GB of RAM and four 667 MHz Alpha 21264A (EV6. biomaterial mass.gov/geo/www/cob 12 http://nces.gov 9 http://factfinder.census. and it currently runs under Linux.gov/oa/ncdc. 2. pathogen.gov/home/en/sfl.gov/econ/www/econ_cen.: BIOWAR: SCALABLE AGENT-BASED MODEL OF BIOATTACKS 259 TABLE IV DATA SOURCES FOR INPUT Fig. Most experiments were performed using the Pittsburgh Supercomputing Center (PSC) TCS1 system. food borne.gov/ccd 13 http://nces. date. which is not possible with a nonmodular design. 8−16 We currently have access to additional data streams that we could not describe here due to confidentiality agreements. delivery type (spray or explosion).noaa. for example.

making automation helpful. It extends the response surface methodology [47] by performing knowledge-intensive datadriven search steps via an inference engine constrained by simulation. BSS is a set of codified surveillance guidelines created by the World Health Organization (WHO) and the Joint United Nations Programme on HIV/AIDS (UNAIDS) for tracking agent behaviors relevant to the transmission of HIV. processing. we have validated BioWar against the SIR model. MAN. 3. An automated tool—named What-If Analyzer (WIZER)— has been designed and partially implemented [46]. for smallpox [25] and against a variant of SIR [26] for anthrax. 3. 2) as a data layer for some outputs (such as ER) and the full data source for other variables (such as drug purchases). 2. and continuous model augmentation. 4. and drug purchase at pharmacies data. XII. 4) as a complete and full-scale city simulation (for all data not just bioattacks). WIZER dataflow diagram. the significant number of input parameters. model parameters. WIZER is an integrated inference and simulation engine to do validation and provide explanations. and output variables. BioWar simulator utilizes a tick as the smallest time unit. In the current implementation. 36. pharmacy visits. Comprehensive validation of a model such as BioWar is. refinement. instead of just doing statistical and mathematical calculations. The state machines used during a tick of the BioWar simulator are shown in Fig. In addition. AND CYBERNETICS—PART A: SYSTEMS AND HUMANS. work absenteeism data. Integration of multiple models is greatly facilitated by the nature of multiagent systems. M ODEL V ALIDATION We have validated BioWar outputs against empirical data on school absenteeism. As shown in Fig. XI. school absenteeism data. BioWar state machines. doctor visits. difficult to Fig. BioWar produces output streams including insurance claim data. VOL. transmission. work absenteeism. In these . and analysis of disease surveillance data developed by the Centers for Disease Control (CDC). management. however. which is a set of standards to facilitate the exchange of data for the collection. 3) as a scaled city simulation. WIZER checks the outputs of simulation and adjusts the simulation parameters and metamodels based on empirical data to validate the simulator. ER registration data. BioWar output is generated to conform to NEDSS. drug purchases. MARCH 2006 Fig. BioWar also generates BSSs compliant outputs. 4. and development. NO. Features and parameters of these agents and modules are defined based on empirical data and knowledge. and ER visits [17]. a conventional epidemiological box model.260 IEEE TRANSACTIONS ON SYSTEMS. do manually or semimanually due to model complexity. B IO W AR O UTPUT S TREAMS BioWar can be used to present results in the following ways: 1) as an additional data layer added to existing background data (injection). one tick equals 4 h.

This finding implies that. Fig. Figs.. M . for the “base (no quarantine or vaccination)” scenario. small (150 g spores). Empirical data were used to tune model parameters.05. Each simulation ran for a year of simulation time starting at September 1. XIII. secondary infections. when comparing an agent-based model and the SIR model. simulated at 100% scale. Pittsburgh MSA. Aligning R0 in SIR with R0 in an agent-based model will provide a misleading comparison. 7. VA. and large (3000 g spores). The simulated population of the city was 148 000. determines the social contact(s) most likely to spread the disease. 6 shows that BioWar generates outcomes comparable to IPF and the Sverdlovsk empirical data for the overtime mortality rate among an anthraxinfected population. These are mostly people at the stadium (during a stadium recreational activity) in large measure because anthrax is noncontagious. 5. a variant of the SIR model used for noninfectious diseases like anthrax [26] shows that BioWar produces outcomes comparable to IPF. and simulation results. the mass of the release. The first window of opportunity (W1 ) is defined as the time from the occurrence of the attack to the time when the first person shows first nonspecific symptoms. we tried an anthrax attack at a stadium with three different mass of release scenarios: 261 Fig. validation studies. 7 and 8 contain graphic displays of model output of doctor visits and the number of deaths for an anthrax attack within flu season and a smallpox attack outside flu season in the city of Hampton.5% of the agents in a locality. 2002. San Francisco MSA.CARLEY et al. Our scenarios simulated biological attacks over the city of Hampton. available treatment. S AMPLE R UNS We have run simulations for several geographical areas (San Diego MSA. The nature of the disease. The second window of opportunity (W2 ) is defined as the time from the attack to the time that the first person is confirmed by diagnosis as having the attack disease. Norfolk MSA. 6. R0 changes for each run. BioWar aligns with SIR for cumulative number of infections. and the city of Hampton. However. VA). The simulated attack occurred on October 4. which person gets infected critically depends on the release location. Comparison of BioWar against the incubation–prodromal– fulminant (IPF) model. so secondary infections do not occur. it increases the doctor visit rate significantly for a short period of time after the attack. The smallpox attack left a much longer footprint compared with the anthrax attack due to the contagious nature of smallpox. but we describe just a few examples in this section. For example. may occur through the social contacts. The height of release was 5 m with an efficiency of 0. particularly in first responders such as doctors and nurses. In another experiment. no distinction between R and R0 is made in the SIR model and R is constant for each run and at each simulation step. 5. The population was about 144 000. Comparison of BioWar against SIR for smallpox outbreaks [25] shows that BioWar and the SIR model produce similar outcomes when model parameters are aligned to the same scenario. Mortality comparison between IPF. Many other scenarios can be simulated by BioWar. etc. 2002. We conducted 20 runs for each scenario and calculated the mode. 8 illustrates a simulated aerosolized smallpox attack on Hampton. Cumulative number of infections comparison between BioWar and SIR. and the wind direction.: BIOWAR: SCALABLE AGENT-BASED MODEL OF BIOATTACKS Fig. The number of deaths by anthrax is also tightly clustered. and standard deviation of W1 and W2 . Fig. For anthrax. we aligned BioWar and the corresponding SIR model based on their assumptions. The fraction of infected people who died of smallpox is also larger. a value commonly used in the SIR model. Weapons-grade smallpox was released infecting 2–2. Both the exposed population and infected population vary. in a stadium with the wind blowing at 4. as shown in Table V. For other diseases such as SARS. VA. When an anthrax attack occurs during a flu season. as shown in Fig. demographics. mean. The attack happens at 4 P. VA. modelers should align R0 (or R) in the SIR with R in the agent-based model since only the average cases are comparable. BioWar. R (the reproduction rate over the entire simulation) is different from R0 and is calculated as an average reproduction rate over all relevant time steps in a simulation. is not comparable to R0 in BioWar. and the Sverdlovsk empirical data. as shown in Table VII. shown in Fig. Washington DC. BioWar can estimate windows of opportunity for attack detection. medium (300 g spores). In BioWar. We found that R0 (the average number of secondary cases in a totally susceptible population infected by one primary case). The reproductive rate for smallpox was estimated using yet another BioWar experiment as shown in Table VI.617 m/s. model parameters. The scenario name denotes the .

the earliest time that an infected agent is diagnosed also decreased when the size of the sampling pool increased. For anthrax simulations. In both anthrax and smallpox scenarios. 36. of January 25. AND CYBERNETICS—PART A: SYSTEMS AND HUMANS. 7. in the city of Hampton with a 2–2. Smallpox50 means the smallpox attack scenario with 50 agents infected. on May 26. 2. W1 decreases when the number of initial infections increases. This attack happened during the flu season. attack disease followed by the number of initial infected agents.9 for the smallpox cases. M .05. infecting 2122 people. Number of doctor visits and deaths occurring after a smallpox attack at 4 P. 2003. Fig. and so does W2 . as shown in Table VIII. Similarly. the average number of doctor visits per year is around the midpoint of the empirical bounds and the average number of ER visits is slightly higher than .5% initial infection rate. Since the incubation period of anthrax is proportional to a lognormal distribution. The empirical values are estimated from CDC statistics when no known attacks occur. 2. MAN. This result is intuitively comprehensible. M . 2003. NO. We compared the average doctor/ER visits per person per year with the estimates from empirical data. VOL.262 IEEE TRANSACTIONS ON SYSTEMS. This attack occurred outside the flu season. MARCH 2006 Fig. The average reproduction rate is around 3. the earliest time for an agent to show the first symptoms would decrease on the average when the sample size (the number of initial infections) increases. For example. 8. Number of doctor visits and deaths occurring after an anthrax attack during a sports event at a stadium in the city of Hampton at 4 P.5 kg of spores was released with an efficiency of 0.

and D. [9] K. the average ER visit per person does not exceed the upper bound by a large amount. Graviss. Eds. It is reasonable that the doctor/ER visits increase more than they do for the anthrax cases. 52.gov/abs/cs. 1997. MA: MIT Press. Singa.. Wanger. to be published. Rojas. and J. Anderson and R. M. Cambridge. Cambridge. since the number of infected is relatively small compared with the total population.” in Santa Fe Studies in the Sciences of Complexity. pp. I. and T. Social Network Analysis: Methods and Applications. 1991. U. Conf. Musser. Int.” in Educational Research. vol.K. the average doctor visits are close to the upper bound and the average ER visits are much higher than the empirical range. [Online]. M. Theraulaz. and G. Cummings. 1999.” Soc. no. Models of Bounded Rationality. R EFERENCES TABLE VIII DOCTOR AND EMERGENCY ROOM VISITS the upper bound. Oxford. Integrating Geographic Information Systems and AgentBased Modeling Techniques. Sci. 2004. vol.: Cambridge Univ. 22. vol. Ed. For smallpox simulations. Lerman and A. Methodology and Measurement: An International Handbook. [1] A. and R. Since smallpox is infectious. Epstein. B. A. and unexpected outcomes based on local interactions and ability to be configured to represent actual cities. . Decis. Mushlin. A. no. E. “Agent memory and adaptation in multiagent systems. D. S. U. Mak. R. 2002.” in Proc. 2001. 2752. Germany: Springer-Verlag. “Modeling the public health response to bioterrorism: Using discrete event simulation to design antibiotic distribution centers. Washington. 2001. This result is reasonable because most agents with severe symptoms would rush to the ER. emergent properties. G. M. P. Press. cs. multiple levels.: BIOWAR: SCALABLE AGENT-BASED MODEL OF BIOATTACKS 263 TABLE V ANTHRAX EFFECTS TABLE VI REPRODUCTIVE RATE FOR THREE SCENARIOS OF SMALLPOX ON THREE TYPES OF POPULATION TABLE VII WINDOWS OF OPPORTUNITY XIV. S. Lawson. Berlin. Klovdahl. [11] S. However. Autonomous Agents and Multi-Agent Systems (AAMAS). 1 & 2. the current version of BioWar represents a significant advance over other numerical disease models.RO/0404002 [10] R. [8] K. Galstyan.K. May. Available: http://xxx. [3] H. Med. Press.CARLEY et al. [2] A. M. vol. Oxford. [7] E. 2002. 5. 2) adding organizational response and cost models. Wasserman and K. G. D ISCUSSION Several upgrades for BioWar are planned. New York: Pergamon. [13] N. U. This includes BioWar’s ability to model socially defined mixing and spatiotemporal effects. A. A. Ross. 2003. 1–9. [5] H. A. M. Keeves.K. “Networks and tuberculosis: An undetected community outbreak involving public places. Toward a Containment Strategy for Smallpox Bioterror: An Individual-Based Computational Approach. U. Simon. and M.. S.” Med. M. pp. Australia.lanl. [6] RoboCup-2002: Robot Soccer World-Cup VI. Statistical Methods in Spatial Epidemiology. Press. Hupert. diverse outputs. 1994. Lerman. W. 797–803. Galstyan. New York: Wiley. A. J. A. A. Melbourne. Computer Science abstract. Oxford Univ. Gimblett. Lima. Press. 2003. While there is much to enhance in the future versions. Infectious Diseases of Humans: Dynamics and Control. including: 1) adding physical infrastructure such as road networks and air traffic networks. [12] J. S. T. 2003. Chakravarty. DC: Brookings Institution Press. not only those initially infected need to go to doctors/ER but also agents who are infected over time. M. Faust.RO/0404002. R.. Kaminka. [4] A. 681–694. Mathematical Analysis of Multi-Agent Systems. 3) connecting with real time data streams. “Social network analysis. Bonabeau. Callahan. Hogg. Adams. Yaganehdoost. “Swarm intelligence: From natural to artificial systems. P. New York: Oxford Univ. Burke. Klovdahl. 1982.: Oxford Univ. 22.: Santa Fe Inst. pp. Dorigo.

New York: Wiley. M.” Amer. Carley is a member of the Academy of Management. 2. “Networks and epidemiology: Modeling structured diffusion.. J. Jones.” Nature. Response Surface Methodology: Process and Product Optimization Using Designed Experiments. Infectious Disease Handbook for Emergency Care Personnel. group. Assoc. University of Pittsburgh. [24] S. 1999. 10–11. 2. no. Her specific research areas are computational social and organization theory.” presented at the North American Association Computational Social and Organizational Science (NAACSOS) Conf.K. D. 3. Fridsma.html [18] M. and G. Carley. ACM. K. pp.. 1995. 17. private communication. Amer.D. no. 299–318. 2nd ed. “A theory of group stability. PA. Morris. V. Glob. Carley. She received two S. Nave. A. 429. CMU-ISRI-04-101.ccs. She is a Professor of Computers. “Computational organization theory. 7–9. A. Eubank. no. 2665. Chen. Netw. P. Gasser. B.. and D. 6988. Newsl. 1996. Blaser.: SpringerVerlag. Morris. pp. 2001. dynamic network analysis. 2004 (forthcoming). pp. vol. Pittsburgh. and M. Altman. vol. O. His research interests include the development of computational tools to study patient safety. Barrett. [21] M. no. pp. pp. D. L. in 1978.. [27] T. K. 1997. Guest Editor. an experimental computer-based diagnostic consultant for general internal medicine. [Online]. [15] S. Oper. Assoc. Inst. Pople. Podhisita. Yahja. Carley. M. C. one in political science and one in economics from the Massachusetts Institute of Technology. AZ. no. 307. Dis. vol. Kretzschmar. Morris and M. 1995. vol. 1995. Infectious Disease Handbook. [34] D. Isada. D. Infect. 87(01-6a). H. Secur.asanltr. pp. [Online]. no. Yahja and K. Cambridge.” J. [23] M.D. MA: MIT Press. 2002. “Social networks in contemporary societies. M. vol. Madrid.” AIDS. Washington. pp.lanl. 281. V. L. no. in 1984. Med. for Infectious Diseases. Fort Detrick. vol.. Engl. Prof. 1982. and the impact of telecommunication technologies on communication and information diffusion within and among groups. M. and the Ph. Spain. J. G. cs. Res. Cambridge. pp. Yuan. implement. B. 287–313. Available: http://www. degree in sociology from Harvard University. Weiss. and J. 2nd Symp. Myers and D.264 IEEE TRANSACTIONS ON SYSTEMS. PA. 2000. 10.B. Z. FL: Lewis Publishers. [35] M. 2001. M. He is currently the Faculty Lead in the caBIG project through the NCI to develop. Casani. Tech. London. Edwards. 36. Goldman. Cincinnati. [44] M. J.htm [45] L. L. Sterling. 22. Cambridge. Support Syst. [26] L.edu/isri2004. L. [20] M. pp. and Construct—a model of the coevolution of social and knowledge networks and personal/organizational identity and capability. Nov. L. Carnegie Mellon Univ.. MD. no. McClellan. Kathleen M. A. 1265–1271. 11 (suppl. and A. DC. Also to appear in Springer-Verlag Lecture Notes in Computer Science as vol. 10–11. pp. B.adm. Mehta. Introduction. Inglesby et al. Sci. Rickmeier. 1999.” Proc. Miller. Special Issue. [39] ——.D. 17.” in Proc.” Decis. Carley. Popul. A. 180–184. M. [Online]. vol. 1. Netw. 331–354.. US Army Med. “Mailborne transmission of anthrax: Modeling and implications. Carley received the Ph. Toroczkal.. Available: http://www. [42] R. [38] A. 3–4. Stanford University. Comput. MAN. 7027–7032. M. American Sociological Society. Informs. “Modeling disease outbreaks in realistic urban social networks. M. Carley. Myers.gov/ccs5/documents/pdf/2001/LA-UR-01-5513. no. unpublished. Blythe. [36] E. “Going beyond the data: Empirical validation leading to grounded theory. Pittsburgh. OH: American Conference of Governmental Industrial Hygienists (ACGIH). VA: Earthscan. Pittsburgh. pp. (2001).” Commun. and collaboration between healthcare providers. G. Turner.” in ECAL 95. Methods Res. Wang. 9. [37] K. [22] M. MA. “Concurrent partnerships and syphilis persistence: New thoughts on an old puzzle. organizational and social adaptation and evolution. Mar. D. 3. Res.” in Proc.Yahja. and A.. “The design of automated validation and explanation of large-scale social agent systems. H. 393–396. scale model of weaponized biological attacks.. H.cmu. 2003. D. 56. and N. 35– 47. 99–126. Guclu. 11–14. 28. 10.-C. D. M. Acad.-Amer. Branch. 11. BioWar: A city-scale multi-agent network model of weaponized biological attacks. [30] C. 2000. E. vol. 3. . vol. H. Her computer simulation models meld multiagent technology with network dynamics and are in areas such as BioWar—a city. to be published. A). Yahja “Aligning simulation models of smallpox outbreaks. NO. N. pp. [47] R. Washington. Pittsburgh. and computer science. Stud. PA. Rev. Lecture Notes in Computer Science. Mar.com/ newsletter/01-6/articles/016a. pp. pp. Rep. vol. USA. Douglas B. [28] USAMRIID’s Medical Management of Biological Casualties Handbook. Transm. 2002. Med.” presented at the Dynamic Social Network Modeling and Analysis Workshop. Sockwell. pp. [40] K. 109–134. 1994. 2004. Note regarding source strength. 3. “Interist-I. E. 504–507. “Elements of a theory of simulation. AND CYBERNETICS—PART A: SYSTEMS AND HUMANS. 2001. Rasmussen and C. pp. [31] R.” Soc. and did his internship and residence at Stanford University Hospitals.” in Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence. computational text analysis. 2002. Kaminsky.” AIDS 97: Year Rev. T. 468–476. Marathe. S. Handcock. vol. PA. Sociol. Begier. 299–330. MARCH 2006 [14] A. Stanford. Kaminsky. 2003.pdf [16] S. C. 8. M. Gray. J. (2004). M. 8.” N. Stud. clinical work processes. 3–4. Atmospheric Dispersion Modeling: An Introduction to Practical Applications. Chaturvedi and S. [32] G. Jun. M. Chen. no. Casman. A. S209–S216. “Scalable. May 1999. S. 42. Wawer. V. “Bridge populations in the spread of HIV/AIDS in Thailand. Schreiber and K. “Sheltering effects of buildings from biological weapons.” Sci. Kaminsky. 2003. 6. “A micro-simulation study of the effect of concurrent partnerships on HIV spread in Uganda. Montgomery. 18. “Model alignment of anthrax attack simulations. 2001. “Anthrax as a biological weapon..-C.” Math. “Simulations in economics and management: Using the SEAS simulation environment. Ann Arbor. the American Association for the Advancement of Science and Sigma XI. A. Eubank. and J. B. Morris. 2003. DC: American Association for Clinical Chemistry (AACC).. Campbell. [29] K. He is an Assistant Professor at the School of Medicine. L. West. and evaluate standards-based interoperable clinical trial applications for managing clinical trials. vol. the International Network for Social Networks Analysis. Tucson. vol. [43] D. OrgAhead—a model of strategic and natural organizational adaptation. M. 2002. MA. Pittsburgh. R. pp. 1993. Klovdahl.” presented at the North American Association Computational Social and Organizational Science (NAACSOS) Conf. Boca Raton. R. Dis.” Mil. Davies-Cole. Morris. pp. Available: http://reports-archive. Morris and M. [17] K. private communication. degree from the University of Michigan. Buckeridge.. she received the Lifetime Achievement Award from the Sociology and Computers Section of the American Sociological Association. “Social networks and infectious disease: HIV/AIDS. 8. Webb and M. She is the author or coauthor of five books and over 100 articles in the area of computational social and organizational science and dynamic network analysis. 163–165. and Society at the Institute for Software Research International in the School of Computer Science at Carnegie Mellon University. [19] M. Kasten. Intelligence and Security Informatics. H. vol. “Biological warfare human response modeling. Kretzschmar.” Sociol. vol. no. degree from Stanford Medical Informatics. Kumar. “Concurrent partnerships and transmission dynamics in networks. [41] C. no. May 2004. Srinivasan. Aberg. L. ACM Appl. CA. efficient epidemiological simulation. 1735–1745.. Fridsma. Carley and L. no.. Tummino. Her research combines cognitive science. 3.” Soc. “The national capitol region’s emergency department syndromic surveillance system: Do chief complaints and discharge diagnosis yield different results?” Emerg. Meselson. in CASOS. “Sexual networks and HIV. F. In 2001. degrees. Fridsma received the M. 60–61. 2002. Anno. 99. [46] A. Organizations. Workbook of Atmospheric Dispersion Estimates: An Introduction to Dispersion Modelling. VOL. Carley. Ed. no. and A. 1991. [25] L. vol. Barrat. with specific interest in clinical trials and oncology care. [33] G. 2001. B. no. Natl. PA... L.” invited editorial for Sex. U. social networks. E. 9.

risk communication. His research primarily focuses on theoretically good parallel finite-element mesh generation. policy analysis. management and organization science. Pittsburgh. and also in the television industry. He is currently the Team Leader of the BioWar biological weapons simulator development effort. Her current research concerns the interactions between social. Organization. Pittsburgh. and on new developing systems. degree in geography and environmental engineering from the Johns Hopkins University. and Carnegie Mellon University. degree at the Computation. and technical factors in the transmission of infectious disease.D. She is also interested in water resource issues.D. PA. meshing from implicit surfaces. knowledge-based systems. IN. PA.S. Pittsburgh. Pittsburgh. and modeling tools for various scientific applications. in 1997 and 2002.: BIOWAR: SCALABLE AGENT-BASED MODEL OF BIOATTACKS Elizabeth Casman received the M. Pittsburgh. and has co-edited a book on the contextual determinants of malaria. Notre Dame. Démian Nave received the B. computational modeling for defenses against distributed denial of service attacks. He worked in the fields of solid state physics. degree in computer science from the University of Notre Dame. Flagstaff. degree in engineering and public policy from Carnegie Mellon University. both on traditional systems. and economic variables into contagion models. Her research interests are focused on combining artificial intelligence and agent-based modeling to conduct technological and policy analysis in the area of information security. like clusters of workstations. Lviv. social networks. and agent-based simulations on policies to counter the spread of epidemics. He has published over 40 papers in journals and conference proceedings and obtained three patents. He is currently an Applications Programmer at the Biomedical Supercomputing Initiative. degree in physics from Lviv State University. simulation. New York. environmental. His other work includes improved algorithms for surface reconstruction from volumetric image data sets.S. Since coming to Carnegie Mellon University. technological. She is an Assistant Professor at the Department of Information Systems. PA.D. in 1987. institutional. optics. where he designed acoustooptical multispectral cameras and microscopes. He is currently with the Pittsburgh Supercomputing Center. . PA. degree in microbiology from the Northern Arizona University.S.CARLEY et al. Carnegie Melon University. and Society program in ISRI. Baltimore. degrees in anthropology and human-computer interaction from the University of Pittsburgh. Specific areas include countermeasures against the propagation of computer viruses. he joined Carnegie Mellon University. 265 Li-Chiou Chen received the Ph. degrees in computer science and aerospace engineering and the M. Pittsburgh. she has specialized in integrated assessment modeling of infectious disease epidemiology. Alex Yahja received the M. Ukraine. like the National Science Foundation’s TeraGrid. respectively. and parallelization of highperformance scientific applications.S. genetically modified organism policy. both in the context of bioterrorism surveillance/response and the epidemiology of diseases linked to global change. and developed software. School of Computer Science and Information Systems. PA. PA. and the Ph. He is a Senior Programmer for the CASOS Program. SCS. degree in engineering and public policy and in robotics from Carnegie Mellon University. PA. MD. building numerical models that incorporate social. Pittsburgh. and risk analysis. in 2003. Pittsburgh Supercomputing Center. The main focus of her bioterrorism work is the early detection of covert attacks. Pittsburgh. In 1996. His work includes the first “guaranteed-quality” parallel tetrahedral mesh generation algorithm that has been proven in theory (and practice) to produce meshes no worse than a comparable sequential meshing algorithm. Boris Kaminsky received the Ph. model validation and improvement.D. His research interests include simulation. Neal Altman received the M. and is currently working toward the Ph. respectively. PA. developing stochastic simulation applications in the field of computational microphysiology.Sc. and machine learning. Pace University. NY.