Socio-Cognitive Agent-Based Simulation of Evacuation Behavior

Collaborative Research: Agent-Based Modeling of Hurricane
Evacuation Decision-Making
I.
Objectives
This research will develop an agent-based model (ABM) of hurricane evacuation decisionmaking founded on the latest theoretical and empirical research into the socio-cognitive aspects
of evacuation behavior. In doing so, three major hypotheses will be examined: (i) synthesizing
existing statistical models of evacuation in an ABM framework allows them to be applied in new
contexts; (ii) households’ evacuation decisions will trend together based on identity and social
network, and these trend groups will correspond to clusters of model agents with similar
evacuation narratives; and (iii) social network connectivity is a significant influence on the
homogeneity of evacuation response within a subpopulation.
Evacuation is a major component of hurricane response and is the most effective strategy in
mitigating human mortality related to hurricanes. Many of the policy options that can have a
significant effect on evacuation decisions and timing are influenced by traffic curves (functions
that describe the number of evacuees on the road at any given time). While sophisticated models
of evacuation traffic exist, the assumptions they make about the demand curves generating that
traffic are frequently far less sophisticated.
Those demand models that can accurately reproduce observed evacuation behavior are
empirical in nature and difficult to translate to other locales and events, while models of the
individual decision processes that generate traffic demand tend either to assume homogeneity of
agency and response across a region, or to be formulated in terms that are tied to the
demographics of the region. However, as was learned with Hurricanes Rita and Katrina, many
groups that are similar in terms of geographic vulnerability will not only respond differently to
risk, but also have different ability to respond. Further, that evacuation choice is typically
decided not at the individual level, but at the level of the household or social group. Existing
models do not capture the dependency of differential response on the details of local
demographics. Agent-based modeling provides a robust methodology for addressing this
problem of generalization.
The proposed model represents decision-makers, at the household level, as independent
software entities. This approach scales well to different population sizes and distributions of
subpopulation attributes; can capture emergent phenomena such as the cumulative, collective
influence of independent decisions within a social network; and is a natural way of mapping the
problem from the physical world to the realm of simulation, representing correlations between
demographics, geography, behavior, and decision-making in a natural and realistic way. In
addition, this approach supports simulation of alternative evacuation policies such as public
information strategies and organization of community-based evacuation networks.
Building on the results of Gladwin, Gladwin, and Peacock’s 2001 survey-based ethnographic
model [1] of the evacuation decision process, focus groups will be used to further elucidate the
decision-making algorithms used in evacuation. The agents will enact these decisions within a
temporal event structure compatible with the state-of-the-art empirical dynamic travel demand
model of Fu and Wilmot 2006 [2], which generates an evolving statistical description of
evacuation decisions. Using an agent-based framework to meld and generalize these two
approaches will result in a model that is consistent with the empirical models, but also provides
enhanced insight into the factors and interactions of factors affecting evacuation decisions. The
ABM will also make explicit some of the implicit factors in each of these models, in particular
the social networks that link individuals and households for communication and provide them
1
with access to opinion leaders. This opens the door for broader study and more intuitive
understanding of evacuation behavior, and may allow emergency planners to more effectively
anticipate and manage the demand-side aspects of hurricane evacuations and to more
appropriately target risk communication.
The proposed research will develop the computer simulation; conduct focus group studies on
evacuation decision-making to inform, refine, and validate the computer model; use modern
experimental designs to analyze the sensitivity of the model to different demographic and
circumstantial factors, developing understanding of their importance in policy-relevant ways;
apply the model to simulate the effects of alternative policies; and convene a cross-disciplinary
student seminar to communicate and explore the results from multiple perspectives.
II.
Background & Significance
Transportation planning/engineering models of evacuation, which are becoming increasingly
sophisticated, are hampered by relatively primitive demand models. This mismatch in the
quality and resolution of models limits opportunities for improving crisis management in
evacuations. Two key points of leverage are (i) understanding how to constructively influence
demand, and (ii) providing more efficient egress from impacted areas. Transportation
planning/engineering models address the latter point, and demand models of commensurate
resolution and responsiveness are needed to drive these transportation models.
Existing evacuation demand models (i) are not explanatory of the factors affecting evacuation
decisions, (ii) are not responsive to policy inputs, (iii) fail to explicitly account for crucial sociodemographic differences and social structures (e.g., social networks), (iv) ignore information
dissemination and flow, (v) cannot be transferred geographically or temporally to different
evacuations, and (vi) fail to quantify uncertainty. However, the best of them can successfully
reproduce individual evacuation decisions as elicited by post-storm interviews [1] and
approximate the collective time-dependent observed evacuation-departure times at a coarse
level [2].
Agent-based methods have proven useful for understanding the collective dynamics of large
groups of individuals engaged in similar behaviors or sharing a similar identity [3]. They are
especially useful for the study of emergence, which is when a complex group behavior results
from the interaction of simple individual rules. Examples of emergent behavior include traffic
jams [4], flocking in fish and birds [5], and foraging in ants [6]. Agent-based modeling has also
been successfully applied to the study of human movement in building evacuations [7], yielding
operational knowledge that can be used to enhance survivability in an emergency.
Ideally, evacuation-demand models should be informed by the latest research in social and
cognitive theory. Moreover, with a sound research foundation of this kind, evacuation models
can more reliably predict responses to policy inputs. Responses to evacuation policies are
formed through social practices and cognitive processes. Demand-based models of evacuation
policy have not been sufficiently sensitive to social and behavioral variation, and in many cases
have failed to capture critical characteristics of population groups facing natural disasters. A
unified approach, which integrates social and behavioral research, agent-based demand models
and policy simulations, will help decision-makers in the public realm create more effective
evacuation policies.
A.
Determinants of Evacuation Behavior
Central to appropriate transportation planning for evacuation is accurately gauging the demand
for roadway usage—that is, accurately accounting for the number of vehicles that will be in
transit on the roadways at any given time in the evacuation process. A number of evacuation
2
studies conducted under the aegis of the Army Corps of Engineers and Federal Emergency
Management Agency (FEMA) have consistently called for greater flexibility in modeling
differential behavior of subpopulations according to different contexts and storm scenarios [812]. This demand is based upon the decisions of households to evacuate. Research on the
determinants of hurricane evacuation has broadly fallen into three categories: research that
approaches evacuation decision-making in terms of risk factors for evacuation failure, research
that explores the cognitive processes that are tied to specific action or to inaction (i.e. risk
perception and risk communication), and research on social vulnerability to disaster.
Differential responses to hurricane risk are well known. In his 1959 study on response to
Hurricane Audrey, Fogelman [13] found prior experience with storms, perception of risk,
communal and historic knowledge of hurricane response, social networks, and gender were
important factors in human behavior prior to and during the storm. However, despite increasing
understanding of the determinants of evacuation decision-making, models of hurricane
evacuation behavior tend to capture only a single dimension of this complex decision process,
often handling these variables linearly and in isolation.
Although explanatory frameworks and mechanisms differ across approaches, a number of
factors are consistently found to impact evacuation actions. These include socio-economic and
demographic features of the households, risk perception, and risk communication. While it is
understood that within a household, different individuals may evacuate separately or split
evacuation decisions (e.g. some members may chose to remain) [14], research focused on those
factors that correlate to the behavior undertaken by the majority of household members shows
that evacuation decisions are influenced by size and type of household [15, 16], the presence of
children or elderly people [15, 17], pet ownership [16-18], gender of household head [16, 19],
ethnicity [15, 16, 20], socio-economic status [15-17, 21-24], and fear of inability to return home
following evacuation [25, 26]. These factors may serve as proxy measures for access to
transportation and other material resources that influence evacuation behaviors [20, 27].
Additional fault lines for vulnerability include, age, poverty, education, and other markers of
socioeconomic status and reinforcement of warnings within the familial or local social network.
[13, 28-30].
Social factors that impact evacuation decisions include the importance of family networks,
sources of warning, and availability of non-governmental refuge [8, 31]. Other determinants
outside the household include risk perception [20, 26, 32-34], event-related factors such as
storm severity and trajectory [18, 35, 36], and various aspects of risk communication [8-12, 16,
20, 26, 32].
No one factor appears preeminent in evacuation decision-making; in fact, research has revealed
myriad complexities. Geographic and storm aspects place households at risk. Material,
demographic, and experiential aspects of households influence both the desire and ability of the
household to evacuate. The decision-making process is influenced by fixed factors and by factors
that evolve along with the hurricane threat. These factors are further complicated in that they
speak to context- and event-specific characteristics of the household, the community, and the
hurricane. The factors that remain fixed over the course of a single evacuation event include
socio-economic and demographic characteristics of the household, geographic situation, and
previous storm and evacuation experience. Those that change as a storm approaches include
risk communication, risk perception, hazard magnitude, community and social network actions
and resources, and response options. The confluence of the static and dynamic features of
household and community inform evacuation decisions and thereby impact the evacuation
demand curves.
3
B.
Critique of Evacuation Modeling
The landscape of hazard has repeatedly been shown to be structured by society and culture, but
there remains a disconnection between evacuation planning, risk assessment and social reality.
Despite our understanding of the complicated and dynamic processes involved in decisionmaking, hurricane evacuation modeling has suffered from a methodological inability to
reproduce interactions of heterogeneous systems and evolving system dynamics and feedbacks.
Traditional hurricane-evacuation models are based on biophysical properties of the
environment (e.g. whether or not an area is in a flood plain) and assume a homogeneous social
character. The reasons for this are two-fold.
First, evacuation models rely upon empirical studies of evacuation behavior. Many published
studies reflect publication biases that lean toward statistical models based on quantifiable data
derived from surveys. Qualitative work, by contrast, has shown that cultural context, social
networks, social resources, empowerment, and agency influence evacuation behavior through
pathways that are not often reflected in quantitative models. As a result, warnings, evacuation
planning and timing, and other policy measures for hurricane mitigation are targeted based on a
false assumption of homogeneity. The fallacy and costs of such assumptions were laid bare with
the evacuation response to Hurricane Katrina, in which sub-groups such as the elderly, disabled,
and poor were unable to get out of harm’s way.
A second factor contributing to the weak performance of evacuation modeling has been the
failure of evacuation planners and modelers to incorporate behavioral aspects of the human
population into their models. Sims and Baumann [37] denounced this practice twenty years ago,
faulting studies of evacuation for failing to understand factors that underlie the decision-making
process. It is striking that this disparity continues despite a much wider base of information and
more flexible methodologies to redress this shortcoming.
The differing analytic frameworks used to explore evacuation decision-making (i.e. risk factors
for failure, cognitive process, and vulnerability) yield varying interpretations of governing
factors and thereby suggest different mechanisms for changing evacuation outcomes. For
example, socioeconomic status or the material culture is considered a key determinant within
the context of vulnerability. Appropriate mechanisms for intervention in this dimension of
vulnerability therefore need to address a broad framework of social and economic development.
A different set of policy intervention points are suggested if the primary determinants of
evacuation are related to risk communication (i.e. timing of information, who delivers
information, and risk perception). The impact that social networks have on filtering
communication and moderating the influence of socio-demographic and material aspects
suggests still different intervention points.
C.
Strengths of Agent-Based Modeling
Agent-based modeling is an approach that allows for straightforward examination of a system’s
workings due to the very direct correspondence between real-world entities and elements of the
model. An ABM represents a real-world system as a collection of dynamically interacting rulebased entities (the agents) [3, 6]. Agents are generally situated in space and time within an
environment that allows them to act. These kinds of models are a very natural way of
representing social or ecological systems where living things interact. An ABM is built by
enumerating the components of the system, creating “agents” whose attributes reflect those
components, and assigning each agent a set of behavioral rules that mimic the behavior of the
corresponding entity. To the degree that rules of behavior can be accurately articulated, an ABM
provides a very simple way of exploring the interactions and implications of those rules.
ABMs are well suited for studying problems of emergence, where complicated collective
behaviors result from the interaction of many agents obeying simple rules [38]. The foraging
4
behavior of ants is one example of this kind of emergence, as are flocking behaviors in fish and
birds [39]. For example, realistic flocking can be produced by providing a group of moving
agents that follow three simple steering rules: (1) head towards the middle of the group, (2) head
in the same direction as your neighbors, and (3) avoid collisions. A collection of agents following
these rules will exhibit flocking as seen in the real world, without any need for centralized
control or high-level awareness or intentionality [5].
Agent-based modeling has been used to study the flow of foot traffic evacuating a building in an
emergency. This research demonstrated that a pillar placed a short distance off-center in front
of the main exit will significantly reduce jamming and increase the flow through secondary exits,
a result that can be used by architects to improve the fire safety of buildings [3, 7, 40]. Along
with biological systems like ant colonies [41], ABM techniques have also been used to study
many social systems and phenomena, such as vehicular traffic [42], civil violence [43], finance
[44], prehistoric settlement patterns [45], and transportation [46], to name only a few.
Another strength of agent-based modeling is that it is highly amenable to extension for the
purpose of including new elements in the model or elaborating on processes of interest. With
this approach, it will be simple for the research team to refine and adapt the model in response
to the observations of the focus groups, capturing any significant features that may have been
omitted and correcting any faulty assumptions in the initial formulation. An ABM is also very
well-suited for interfacing with a broad spectrum of transportation planning models. Many of
these models are themselves agent-based in formulation, and can be driven by exporting
relevant data about the agents in this model, while suitable driving data for those that are
formulated in statistical terms can be created by simple aggregation of data from the agent
population.
D.
Public Policy Requirements
Agent-based models have been used to test policy alternatives in a variety of planning
environments. For example, members of the research team have successfully applied an agentbased model of social characteristics of resort-community in-migrants to an evaluation of resort
housing policy [47], and an agent-based model of urban location decisions to an evaluation of
land use policies [48]. Agent-based policy models typically involve construction of scenarios as
alternative rule sets within the model. These scenarios are designed by distilling assessments of
current policy discussion and best practice alternatives into model elements. The outputs of
these scenarios are evaluated according to criteria defined through external research, such as
interviews with policymakers. This approach is a powerful method for exploring policy
alternatives because it directly maps diverse social characteristics to policy response [49] and
provides a framework for the simulation of policy constructs that illuminates key strategies and
dimensions of current debate [50-53].
III. Hypotheses
This project will test three hypotheses: (i) synthesizing existing statistical models of evacuation
in an ABM framework allows them to be applied in new contexts; (ii) households’ evacuation
decisions will trend together based on identity and social network, and these trend groups will
correspond to clusters of model agents with similar evacuation narratives; and (iii) social
network connectivity is a significant influence on the homogeneity of evacuation response
within a subpopulation.
A.
Transferability of Agent-based Evacuation Models
This project will build on two existing statistical models: a sequential logit model and an
ethnographic decision tree. These models are based on extensive survey and observational data
5
for particular events and locales. Approaches such as sequential logit or survival analysis find
associations between time-dependent variables like time remaining before landfall and total
number of evacuations during a given period; any static factors are resolved at a very coarse
scale and aggregated across the population. As a result, the correlations between evacuation and
demography are calibrated only for the given input data, and do not apply in other locales where
the static factors may be different. The decision tree model, on the other hand, disaggregates the
decision factors into independent elements that can be applied in a region with differing
demographics, but is formulated in a time-invariant way that cannot be used to generate an
evacuation demand curve. This project will synthesize the two models by recasting the decision
tree from a population-sorting mechanism to the behavioral algorithm for agents, and using the
temporal structure of the sequential logit model as the framework within which the agents act. It
is hypothesized that, absent network effects, the resulting agent-based model will be consistent
with the decision tree if only demographics are considered, and consistent with the sequential
logit model if only time-evolution of demand is considered. If so, then the ABM will be able both
to generate demand curves and to do so for different contexts, because it evolves the
populations’ decisions in a time-dependent way while also explicitly accounting for the static
and dynamic factors that influence those decisions.
B.
Narratives for Clusters & Outliers
Since an ABM simulates the interactions of discrete decision-making entities (households, in
this case), it provides a great deal of information regarding the evolution of simulated decisionmaking. In fact, agent-based models intrinsically produce narratives of individual decisionmakers’ behavior. Usually these decisions and their outcomes are aggregated into output
variables that can be observed in surveys, traffic counts, censuses, etc., obscuring heterogeneity
of response. Because of the close integration of the focus groups with the modeling effort, this
project has the opportunity to qualitatively verify the ABM-generated narratives. The focus
groups will be stratified to represent groups that the literature suggests use different decision
strategies based on identity and social relationships. These groupings are expected to exhibit
similar intra-group evacuation decisions, and their input will be used both to inform the initial
model and evaluate the degree to which the model narratives resonate with both experienced
and potential evacuees in the field. Specifically, clusters of simulated households with similar
decision histories will be identified, clearly defined narratives describing those histories will be
developed, and feedback will be solicited from focus groups regarding the realism and
completeness of these narratives. In addition to “closing the loop” between the modeling effort
and the agents being modeled, this will provide better understanding of the decision-making
process, supply qualitative data on the validity of the model, identify potential deficiencies, and
suggest data collection strategies. Closely allied with this effort will be the identification of
influential “outliers” —decision-makers that exert disproportionate influence on the simulated
outcomes due to social connectivity or other factors. Careful design of sensitivity studies will
enable the analysis of outliers, who may be key targets for policy initiatives.
C.
Influence of Social Network Connectivity
Existing literature on evacuation behavior offers several explanatory models that assume social
networks play a role in both mitigating lack of material resources for evacuation (i.e. social
vulnerability) and influencing behaviors through risk communication [20, 27, 34]. The ABM will
examine the effect of these assumptions by including a social network that makes agents’
evacuation decisions perceptible to other agents in the network. The connectivity structure of
the network will be related to diverse features of the modeled households, including
demographic traits (e.g., social class, ethnicity) and non-traditional identifiers of membership
(e.g., neighborhood of residence/place attachment), reflecting the social vulnerability of
6
different groups. Increasing or decreasing the effect of signals transmitted through the social
network and/or changing the structure of the social network will allow testing of the hypothesis
that social connectivity is a significant factor in the homogeneity of evacuation response: if true,
it is expected that a subpopulation that is poorly-connected will respond in a more
individualistic and piecemeal fashion than one that is well-connected. Connectivity in this
context can be measured in a number of different ways, including clustering coefficient, link
density, and degree distribution of groups and actors in an affiliation network [54].
IV. Proposed Work
A.
Work in Progress and Model Development
Prior to this proposal, Dr. McGinnis and Dr. Bush developed a conceptual prototype ABM to
explore this idea in a preliminary way. The prototype contains statistically identical agents that
respond to generic time-dependent signals by increasing an evacuation readiness variable.
When this variable reaches a certain threshold, the agent decides to evacuate. Agents are
connected in a simple network that provides them with awareness of other agents’ evacuation
decisions, which is treated as another time-dependent evacuation signal. This model exhibits
dynamics that are sensible and realistic in a rough and qualitative fashion. This project will use
the structural skeleton of the prototype as a starting point to speed the development of the
agent-based model of a population’s collective hurricane evacuation decisions described below.
Prototype ABM Behavior
10000
Agents
8000
Home
6000
Preparing
Traveling
4000
Safe
2000
0
20
15
10
5
0
Hours before Landfall
Figure 1. Agents in the simple prototype inhabiting different states over time.
The household is the typical unit of decision-making and behavior in a hurricane evacuation;
therefore, in this model, each agent will represent a household. The household is embedded
within a social network that constrains and influences its evacuation decision, and, as stated in
Section A, it is acknowledged that individuals within a household may arrive at a split decision
as to whether or not to evacuate. There is a rich set of constraints and interactions operating at
the household level, and modeling the system at this level allows them to be captured. Because
evacuation research focuses on the decision that impacts the greater part of the household
members, it is reasonable to focus on households as representative of household majority.
7
Each agent will have various socio-demographic and geographic attributes associated with it
that are relevant to its behavior, like income, pet ownership, elderly members of the household,
whether the house is in an evacuation zone, and so on. When the model is populated with
agents, randomly chosen (in a statistically precise manner) attributes that fit the demographic
profile of the modeled population will be assigned to the agents. For modeling large and
heterogeneous populations, this research will use multiple subpopulations, each of which will
have its own distinct demographic character.
These agents will use a decision-making algorithm that is similar to the ethnographic decision
tree model of Gladwin, Gladwin, & Peacock [1]. This algorithm is a tree of Boolean yes/no
decisions that are evaluated in sequence to determine what behavior the agent will exhibit:
whether it will evacuate or shelter in place, whether to leave now or prepare to leave later, and
so on. The decision steps depend on both the agent’s attributes and the current situation, so
demographic and circumstantial differences will affect the outcome. For example, each agent
considers, in turn, questions like: “Do I live in an evacuation zone? Has there been an
evacuation order? Do I have a place to go? Do I have money available for a hotel?” that lead up
to a go/no-go decision; the answers to these questions will differ between agents and may
change over time. The decision tree of Gladwin, Gladwin, & Peacock was constructed based on
interview data and tested against survey data, with an 87% success rate in predicting the
evacuation decision of the respondent. The ABM decision tree will be structurally similar, but
formulated at the household level rather than the individual level.
Historical
Preferences
Evacuee Focus
Groups
Social
Network
Peer
Communication
Ethnographic
Decision Tree
Models
Infrastructure
Sociocognitive
decisions
Decision
Tree
Demographic
Profiles
Hurricane
Public Information
Agents
Policy Analysis
Agent-Based Evacuation Model
Sequential
Decision
Models
External Data
Transferability of
Agent-based
Evacuation Models
Narratives for
Clusters & Outliers
Influence of Social
Network Connectivity
Hypotheses
Figure 2. Connections and influences in the proposed agent-based model of evacuation behavior.
The agents will make their decisions in a temporal framework similar to that of Fu & Wilmot’s
dynamic travel demand models [2, 55]. In these models, all the households that have not yet
evacuated are examined at each time step and assigned a probability of evacuating based on
intrinsic and time-dependent variables such as time of day, hurricane speed, and whether the
residence is likely to flood. A fraction of the households are then evacuated according to this
probability. The agent model will operate in a similar fashion, but instead of calculating a
probability of evacuation, each household’s decision to evacuate will be determined by
evaluation of the ethnographic decision tree.
8
In addition, this project will include the effects of interpersonal communication on the decisionmaking process by synthesizing a social network that allows the agents to communicate with one
another. Studies have shown that interpersonal communication, especially with regard to the
valued judgment of opinion leaders, is a significant factor in the decision to evacuate [56]. The
model will include a bipartite affiliation network representing social connections between
agents: agents will be randomly assigned to a number of groups, all of whose members know
one another. This structure captures the observed character of real-world social networks based
on shared contexts of interaction (e.g., church, work, extended family, etc.). Agents will be
allowed to include the decisions of other agents in their immediate network neighborhood in
their decision-making algorithm [54, 57-61]. The effects of other agents’ evacuation decisions
will thus depend not only on the importance they are given in the algorithm, but also the
connectivity of the social network, which determines what agents’ decisions are visible to one
another.
Illustrative Agent-Based Evacuation Model
We use the index i  1,  , N to label each of the N agents in the simulation. Each agent
must be in one of five possible states:
si (t )  S  " HOME " , " PREPARING" , "TRAVELING" , " SAFE" , " DIVERTED",
where t indexes time steps in the simulation. All agents start the simulation at home:
i.e., si (0) " HOME " . The agents also have a quantitative measure of their “urgency” to
evacuation, ui (t ) , the threshold at which their urgency spurs them to action,  i , the
expected travel time to the safe zone,  i (t ) , and a temporal safety margin, mi ; for our
simple prototype model these are Gaussian IID random variables chosen at the start of the
simulation, t  0 . Additionally, each agent is embedded in a social network nij which we
represent in the following manner:
1 if agents i and j communicat e informatio n
.
nij  n ji  
0 if agents i and j do not communicat e informatio n
Finally, each agent evaluates or weighs information according to a function
f : I  
where
I  " landfall forecast" , " evacuation order" , " peer pressure" , etc.
are types of information an agent might receive.
The dynamics of agents is determined by the receipt of new information and by peer
pressure.
ui (t  1)  ui (t )   w  f ( g )  N (0,1)
( w, g )Gi ( t )
where Gi (t ) is the set of signals received by each agent at time t and N (0,1) is a
standard normal IID random variable; each signal ( w, g ) consists of a numerical strength,
w , and information type, g . The strength of most signals is given externally, but peer
pressure is computed from the average urgency of the agent’s peers in the social network:
g ipeer (t )   niju j (t )  nij .
j i
j i
The first time that ui (t ) exceeds i (t ) , the agent makes the transition from
si (t ) " HOME " to si (t  1) " PREPARING" ; additional rules handle other state
transitions and consider the expected time to landfall, the travel time to the safe zone, and
the temporal safety margin. This simple model also allows for the possibility of roadway
congestion forcing traveling evacuees to be diverted to unplanned shelter locations.
Melding different models in
this manner makes explicit a
number of factors that are
handled in an empirical,
implicit, or statistical way in
existing models. For example,
to evaluate perceived risk,
agents will aggregate
information about the
environment (e.g., type of
dwelling, whether it is in a flood
plain), the event (e.g., storm
track, speed, and intensity), and
other relevant signals (e.g.,
hurricane warnings, traffic
reports, behavior of other
agents) and compare the result
to an alarm threshold to answer
the decision-tree question, “do
you feel endangered by the
storm?” This approach makes it
possible to model response to
hypothetical future storms
based on presently available
data, rather than relying on
retrospective data like risk
perception that can only be
collected after a storm.
Model development will take
place in parallel to the work
Figure 3. Simplified description of an evacuation-behavior ABM.
with focus groups. It is expected
that this approach will create valuable synergy, with questions and insights arising in one arena
providing guidance in another. To this end, development will proceed iteratively, starting with a
very simple, but functional, model and adding features one at a time, so that results from
intermediate modeling stages can inform research with focus groups and vice versa.
9
B.
Calibration & Validation
The agent-based evacuation-demand model will have only a few “free” tunable parameters—
most of its input data will be based on the characteristics of the synthetic population, resource
constraints (i.e. vehicle availability, road network status, etc.), and exogenous variables such as
the hurricane track and intensity. Because sufficiently detailed evacuation time-series data for
hurricanes is scarce, this model will use standard cross-validation techniques [62] to calibrate
the ABM’s free parameters and quantify its prediction errors: the set of available
calibration/validation data will be separated into “training” and “testing” partitions, and the
“training” sets will be used for calibration while the “testing” sets are used for validation. As the
training-testing partition is varied, the accumulated statistics will indicate the prediction error.
The actual calibration will use the “downhill simplex” method [63] or possibly a genetic
algorithm [64] to minimize the difference between the predicted and observed evacuationdemand curves for multiple demographic groups and geographic zones as the ABM’s tunable
parameters are varied. Stratifying the comparison ensures that the calibrated ABM parameters
are applicable to the full spectrum of households in the simulation. The calibrated ABM will be
compared with published calibrations of sequential logit and survival analysis models [2, 55].
Additional metrics indicating the quality of the model will be developed to the extent that
empirical data supports their computation.
C.
Qualitative Behavior & Sensitivity Analysis
The first step in analyzing the ABM is to thoroughly map out the various behavioral regimes
exhibited by the simulation and examine their qualitative plausibility in terms of the feedback
loops present in the system. A significant product of the analysis will be a phase diagram
delimiting various possible “decision regimes” and identifying the possible phase transitions
between them. Statistical methods for experimental design provide a structured methodology
for performing this exploration. This research will use Latin hypercube sampling (LHS) and
orthogonal arrays (OA) as the basis of these designs [65, 66]. Because of the speed of the ABM
simulations, it will be possible to run large designs with numerous replications (necessary
because of the stochastic nature of the ABM). The research team believes it will be practical to
implement strength three (and higher) orthogonal array designs in this study—this will enable
the study of three-way correlations in input parameters.
Initially, the researchers will examine the sensitivities of parameters and processes internal to
the ABM model. Key candidate variables for sensitivity analysis are those related to awareness of
risk, expected travel times, destination constraints, and evacuation preparation factors. Variance
decomposition methods [67] will be used to identify the dependencies of the model output on
individual factors. As part of this analysis, input factors that are highly correlated (e.g., not fully
independent) will be identified, as will factors to which the output is insensitive. A thorough
understanding of sensitivities will have benefits for model refinement, data collection and
survey design, and development of mitigation strategies.
Model sensitivity to factors left outside the models and to random events will be evaluated. It
can be argued that a particular crisis scenario can have multiple materially different outcomes
due to factors outside anyone’s control (such as an accident clogging a major evacuation route,
wrong information about the feasibility of a particular evacuation path, or the specific structure
of social interactions in the affected area, all of which will alter how people heed evacuation
orders)—a social systems “butterfly effect”. In such situations, prediction may not be feasible,
but a simulation may allow investigation of the set of possible outcomes. Given constraints like
those imposed by the transportation system, the number of qualitatively different outcomes may
not be large. Furthermore, those outcomes may serve as guidance for developing flexible
10
emergency response plans usable in a variety of circumstances, rather than attempting to design
a single “optimal plan”.
The research will move beyond sensitivity analysis to quantify uncertainties present in the
model to the extent prior probabilities can be assigned to key model parameters. Uncertainties
are conditioned on external factors such as hurricane track and intensity. Latin hypercube
sampling on these prior probability distributions is the most appropriate method for quantifying
ABM output uncertainty. Because the model is stochastic, several replications may be needed in
the experimental design. The impact of varying the model structure will be examined in order to
elicit information about structural and systemic uncertainties in the ABM. The value of such
rigorous quantification of uncertainty lies in its providing guidance about the plausible range of
outcomes, rather than just a single estimate (or unquantified group of estimates) for outcomes
simulated by the agent-based model.
D.
Focus Groups
During the early stages of the project, focus groups will be formed to elaborate the factors,
interactions, and cognitive pathways that influence evacuation decision-making. The results
from these studies will feed into the model development process. As the ABM is refined and
elaborated, a second phase of focus groups will be held to validate the model. This critical
second phase reflects a small degree of community-based participatory research. The
communities that inform the model development will be given the opportunity to vet it and
discuss its possible utility. Through this process, the researchers hope to build a synergistic
interaction between the human narrative, the knowledge gained through qualitative exploration
of the lived experience, and analytic power of mathematical modeling. Unlike large surveys,
which tend towards reduced variety in response, the focus group methodology allows for a large
degree of elaboration upon and departure from theorized models and feedbacks. As qualitative
storm narratives show [14], factors both endogenous and exogenous to the household or wider
social network, including extended family, interact in non-linear ways across time.
Understanding the relative weighting of variables that influence evacuation behavior requires
details best elucidated in small focus groups.
The focus groups will draw from individuals currently residing in the greater New Orleans and
greater Galveston areas. In order to obtain information about how groups that share a similar
sense of identity respond to cues from the social and physical environment, the focus groups will
be stratified around parameters believed to influence diverse evacuation responses to include
membership based on place and membership within social networks to incorporate the
influence of factors such as neighborhood, extended family, social class, and ethnicity.
Focus groups to provide variables and interactions upon which the ABM will be based will occur
in year one of the grant. In year two of the grant, the “narratives” generated from the ABM will
be presented to and critiqued by a second round of focus groups. Ideally the same cohort will be
used for both types of focus groups, but this may prove infeasible, especially in the rapidly
changing social landscape of New Orleans. A total of ten focus groups, five in each of the study
sites, will be realized for each phase for a total of twenty groups over the two-year course of the
grant. Each focus group will consist of 8-12 individuals.
Local coordinators from both New Orleans and Galveston will be employed to assemble focus
groups that feature known modifiers of evacuation behavior, reflecting different genders,
ethnicities, social classes, and household sizes and compositions. (See Letters of Support.)
While the financial scope of this proposal does not allow for the creation of ethnographic records
of the focus group participants, the researchers hope to obtain supplemental funding to conduct
semi-structured interviews with each participant to obtain a more robust ethnographic record of
11
evacuation timing and decision-making. The local coordinators are well-versed in assembling
focus groups based on specific community and social network characteristics.
New Orleans and Galveston share a risk profile in that they lie within an active hurricane zone
and have experienced destruction from storms of historic proportion. The population of both
areas has recent experience with major evacuation efforts related to hurricanes Katrina and
Rita.
In addition, the research team has several years of combined experience working with local
communities in these two regions. In addition to receiving her Ph.D. from Tulane University in
New Orleans, Dr. Johnson worked on a number of community-based projects in New Orleans
prior to Hurricane Katrina and was part of a research project on post-Katrina resource loss and
coping behaviors. In 2006, Dr. Muller and Dr. McGinnis worked with Dr. Robert Harriss,
Director of the Houston Advanced Research Center, on the Galveston Futures project. This
project was a community design charette that brought together representatives from many
sectors of the Galveston community, public and private, to develop plans for the redevelopment
of public lands focusing on issues of sustainability and hazard mitigation.
E.
Synergies between Agent-Based Modeling and the Use of Focus Groups
The ABM will create pathways through which different agents navigate the evacuation event. In
essence, each agent represents a household and will generate a narrative of its journey and
interactions through the system. Researchers will extrapolate from the agents’ decisions to
create a human narrative of information and interaction in the changing landscape of the
community evacuation. Narratives of typical or highly influential agents will be presented to
community members for them to assess how realistically the interactions, imputed judgments,
and ensuing behavior mirror their lived experiences. In turn, the narratives generated by
community members in the focus groups will be compared to agent behavior within the model
to verify that the decision-making algorithms accurately reflect the human decision process. The
degree of correspondence between the two types of narrative and ease with which the decision
processes can be translated from the domain of explanation to the domain of simulation will
give insight into the validity of the hypothesis that the agent-based model generates narratives
that resonate with potential evacuees.
F.
Education and Dissemination
The major outcomes of the modeling will be disseminated on three tiers: local community,
(inter)national research community, and academic institutions. The focus group mechanism
provides a valuable tool for researchers to share and validate the results of the modeling,
particularly in the second year of the research plan when the local community will evaluate the
ABM-generated narratives. In addition, once the final project is complete, results will be
disseminated via the local coordinators embedded within the communities being studied. The
results of this study will further be submitted to peer-reviewed publications.
One of the more valuable educational components of this research is the involvement of
students in graduate level programs at the University of Denver (DU) and the University of
Colorado, Denver (CU). DU and CU will employ one graduate research assistant each for the
duration of the project, thus training a small portion of the next generation of
researchers/planners. The DU graduate research assistant will actively participate in field
research in Louisiana and Texas, and the CU student will actively take part in policy analysis. As
the project nears completion, a cross-institutional colloquium targeting graduate and postgraduate students will be held. The intent is to bring together students from the DU Graduate
School of International Studies who are focusing on issues of disaster preparedness and
homeland security (mostly master’s degree students, some Ph.D. candidates), and urban
12
planning students from CU (mostly master’s level students), with evacuation modelers and
transportation planners to discuss the utility of the ABM and issues surrounding evacuation
planning.
V.
Broader Impact & Significance
This project may have significant impact on the way that evacuation planning is undertaken, as
it will account for heterogeneity in evacuation decision responses and context-specific variables,
both of which have not been elaborated upon in prior evacuation planning models. By providing
a model of evacuation that can explain the effectiveness of various policies and contextual
factors, this research will enable evacuation planners to make more informed and effective
decisions. This project speaks to the NSF concern for the integration of science research and
education through: (1) a collaborative design that integrates feedback mechanisms between
researchers and lay community members as noted in Section E above, (2) a day-long,
multidisciplinary seminar on evacuation decision-making that will unite graduate students from
urban planning, public policy, and disaster planning, and (3) the participation of graduate
students in the research. Working with the lay public to validate the cognitive choices that occur
in the evacuation decision-making process is the first step in what will become an on-going
dialogue with urban planners to introduce it into practice.
The proposed work nicely complements several ongoing initiatives and addresses
recommendations of national and international organizations: The National Science Board’s
Task Force on Hurricane Science and Engineering has identified “human behavior and risk
planning” and “evacuation planning” [68] as high priority research investments. Furthermore,
the Transportation Research Board’s Emergency Evacuation Subcommittee identified, in its
January 2006 meeting, [69] the incorporation of human behavior into evacuation modeling as
critical to effective evacuation prior to events like hurricanes. Finally, the World Meteorological
Organization’s THORPEX program includes a Societal and Economic Applications subprogram
[70] that has identified the societal impacts of high impact weather as a key research topic.
VI. Management Plan
All senior members of the project team are based in the greater Denver area. This geographic
proximity allows for ease in meeting on a monthly basis or more frequently, as necessary.
Further, the team will keep in contact through phone and e-mail.
Each institution will be responsible for managing its portion of the grant. Dr. McGinnis, the P.I.,
located at the National Center for Atmospheric Research (NCAR), will oversee proposal
performance under the guidance of Dr. Bush.
A.
Investigator Roles and Experience
Seth McGinnis, principal investigator, will develop the agent-based model, perform the model
runs, and assist with the analysis of the results. He has more than a decade of experience in the
development and analysis of computer models and simulations, and performed most of the work
on the prototype evacuation model.
Brian W. Bush, consulting co-investigator and senior mentor, will assist in developing the agentbased model, preparing input data sets, and analyzing the results. He has twenty years of
modeling, simulation, and analysis experience, and was a lead contributor (architect, developer,
and researcher) to TRANSIMS (the Transportation Analysis Simulation System [42]), a highly
detailed agent-based transportation planning model. He will mentor the team in advanced
agent-based modeling techniques and simulation analysis.
13
Sandy A. Johnson, principal co-investigator, will lead the qualitative fieldwork in New Orleans,
Louisiana, and Galveston, Texas. She will coordinate educational outreach. Her work on social
vulnerability and negative health outcomes included several community-based projects in New
Orleans. Most recently, she was involved with a study on post traumatic stress disorder,
resource loss and hurricane evacuation being conducted in New Orleans.
Brian Muller, principal co-investigator, will lead the policy and utility analysis aspects of the
study and critique the modeling effort. He has more than two decades of experience in policy
and planning analysis. His recent research focuses on use of agent-based models in land-userelated policy simulations and regional transportation planning.
B.
Time Line
Month
1-6
7-9
Key Activities
 Preliminary model development
 Construction of focus group questionnaire
 Calibration and validation data collection
 Supplemental review of literature
 Phase I focus groups
 Incorporation of focus group results into
model
 Model refinement
 Preliminary analysis of model
Milestones
 Preliminary model
 Focus group questionnaire


Focus group report
Revised model





Final version of model
21-24



Phase II focus groups
Incorporation of focus group results into
model
Finalize model
Further analysis of model
Document and disseminate results
Refined model
Draft model analysis section
of project report
Focus group report
Revised model

Project report
23-24

Multidisciplinary seminar

Seminar report
9-13
14-17
18-20


14