ArticlePDF Available

A-RESCUE: An Agent based Regional Evacuation Simulator Coupled with User Enriched Behavior

Authors:

Abstract and Figures

Household behavior and dynamic traffic flows are the two most important aspects of hurricane evacuations. However, current evacuation models largely overlook the complexity of household behavior leading to oversimplified traffic assignments and, as a result, inaccurate evacuation clearance times in the network. In this paper, we present a high fidelity multi-agent simulation model called A-RESCUE (Agent-based Regional Evacuation Simulator Coupled with User Enriched behavior) that integrates the rich activity behavior of the evacuating households with the network level assignment to predict and evaluate evacuation clearance times. The simulator can generate evacuation demand on the fly, truly capturing the dynamic nature of a hurricane evacuation. The simulator consists of two major components: household decision-making module and traffic flow module. In the simulation, each household is an agent making various evacuation related decisions based on advanced behavioral models. From household decisions, a number of vehicles are generated and entered in the evacuation transportation network at different time intervals. An adaptive routing strategy that can achieve efficient network-wide traffic measurements is proposed. Computational results are presented based on simulations over the Miami-Dade network with detailed representation of the road network geometry. The simulation results demonstrate the evolution of traffic congestion as a function of the household decision-making, the variance of the congestion across different areas relative to the storm path and the most congested O-D pairs in the network. The simulation tool can be used as a planning tool to make decisions related to how traffic information should be communicated and in the design of traffic management policies such as contra-flow strategies during evacuations.
This content is subject to copyright. Terms and conditions apply.
A-RESCUE: An Agent based Regional Evacuation
Simulator Coupled with User Enriched Behavior
Satish V. Ukkusuri
1
&Samiul Hasan
2
&Binh Luong
1
&
Kien Doan
3
&Xianyuan Zhan
1
&
Pamela Murray-Tuite
4
&Weihao Yin
4
#Springer Science+Business Media New York 2016
Abstract Household behavior and dynamic traffic flows are the two most important
aspects of hurricane evacuations. However, current evacuation models largely overlook
the complexity of household behavior leading to oversimplified traffic assignments and, as
a result, inaccurate evacuation clearance times in the network. In this paper, we present a
high fidelity multi-agent simulation model called A-RESCUE (Agent-based Regional
Evacuation Simulator Coupled with User Enriched behavior) that integrates the rich
activity behavior of the evacuating households with the network level assignment to
predict and evaluate evacuation clearance times. The simulator can generate evacuation
demand on the fly, truly capturing the dynamic nature of a hurricane evacuation. The
simulator consists of two major components: household decision-making module and
traffic flow module. In the simulation, each household is an agent making various
evacuation related decisions based on advanced behavioral models. From household
decisions, a number of vehicles are generated and entered in the evacuation transportation
network at different time intervals. An adaptive routing strategy that can achieve efficient
network-wide traffic measurements is proposed. Computational results are presented
based on simulations over the Miami-Dade network with detailed representation of the
road network geometry. The simulation results demonstrate the evolution of traffic
congestion as a function of the household decision-making, the variance of the congestion
Netw Spat Econ
DOI 10.1007/s11067-016-9323-0
*Satish V. Ukkusuri
sukkusur@purdue.edu
1
Lyles School of Civil Engineering, Purdue University, 550 Stadium Mall Drive, West Lafayette,
IN 47907, USA
2
CSIRO, Cities Program, Land & Water Flagship, Bayview Ave, Clayton, VIC 3168, Australia
3
Urban-Civil Works Construction Investment Management Authority of Ho Chi Minh City, Ho Chi
Minh City, Vietnam
4
Department of Civil and Environmental Engineering, Virginia Tech, 7054 Haycock Road, Falls
Church, VA 22043, USA
across different areas relative to the storm pathandthemostcongestedO-Dpairsinthe
network. The simulation tool can be used as a planning tool to make decisions related to
how traffic information should be communicated and in the design of traffic management
policies such as contra-flow strategies during evacuations.
Keywords Agent based modeling .Hurricane evacuation .Dynamic routing .Discrete
choice model .Traffic simulation
1 Introduction and Motivation
Integrated models of household-level behavior and dynamic traffic flows are important
in determining appropriate strategies in hurricane evacuations. However, current hur-
ricane evacuation models largely overlook the complexity of household-level decision
making behavior and as a result, lead to oversimplified traffic assignments and
evacuation clearance times in the network. To fully capture the complexity of hurricane
evacuation, one has to integrate the process of household decision-making and the
resulting traffic interactions that follow after a decision to evacuate has been made. In
this paper, we present a high fidelity multi-agent simulation model called A-RESCUE
that integrates rich activity level behavior including pre-evacuation and shadow evac-
uation with the network level assignment to predict and evaluate clearance times. This
captures the complexity of household-level decision-making and the adaptive dynamic
routing behavior of the evacuees during an evacuation process.
In spite of the considerable advances in dynamic network modeling, most network
models for hurricane evacuations (Yao et al. 2009;HeandPeeta2014;Tarhini
and Bish 2015; Zhang et al. 2015) still suffer from an overly simplified
representation of behavior, which limit their usefulness. At a conceptual level,
in tandem with a traditional transportation planning process, the evacuation
clearance time determination can be divided into four phases trip generation,
modal split, trip distribution and traffic assignment.
In the trip generation phase, households engage in a variety of preparatory
activities prior to evacuation, which Lindell et al. (2007, p. 54) classified as
Bpsychological preparation (seeking and processing additional information until
they are certain evacuation is necessary) and logistical preparation (uniting house-
hold members, protecting property, and collecting clothes and other materials
needed while away from home)^. Previous studies do not fully understand the
activities in which households participate pre-evacuation and it is unclear which,
if any, of these activities can be simultaneous and which must be sequential. In this
paper, A-RESCUE incorporates the logistical preparation of households using a
rich behavioral model (Yin et al. 2014).
The modal split phase addresses the relative proportions of evacuees using private
vehicles and public transportation, which accounts for a relatively small portion of the
population outside urban areas. This has largely been an understudied problem by
evacuation modelers until recentlysee Bish (2011); Sadri et al. (2014a) and Naghawi
and Wolshon (2012). A-RESCUE can incorporate a modal split framework but the data
does not include a modal split. Additional details on the modal split by the authors can
be found in Sadri et al. (2014a).
S. V. Ukkusuri et al.
The trip distribution phase determines where most people evacuate - to the homes of
friends and relatives or stay in commercial facilities rather than public shelters (Lindell
et al. 2011; Mileti et al. 1992;Wuetal.2012). These assessments of aggregate facility
utilization have recently been extended by Mesa-Arango et al. (2013), where a nested
logit model was used to predict evacueeschoice of accommodations in Hurricane
Ivan. Results from this model are incorporated in A-RESCUE to determine the
selection of the safe locations for evacuation.
The above three phases represent the demand side of the problem. On the supply
side, the traffic assignment phase includes the route choice, which has a major impact
on evacuation times (Tarhini and Bish 2015; Yu-Ting and Peeta 2015), but the basis for
evacueesroute choices is not well understood. As Lindell and Prater (2007)noted,
evacuation modelers have often assumed a computationally convenient reason for route
choices such as a Bmyopic view^in which evacuees choose the least congested path in
the network. However, this assumption conflicts with the limited behavioral data that
indicate route choice is based primarily on familiarity. Though Bfamiliarity^provides
some explanatory power, it is useless for modeling unless we can identify generalizable
rules for identifying which routes are most Bfamiliar^to evacuees for specific origins
and destinations. In addition, these route choices vary across households as shown in
Sadri et al. (2014b). A-RESCUE has the capability to include both adaptive and
familiarity based routing using empirical data obtained in this research.
The selected path can be compared with options indicated by the transportation
network to derive behavioral based routing rules and learning mechanisms rather than
relying on the traditional Wardropian principles, which have been criticized for lack of
applicability to evacuation conditions (Lindell and Prater 2007). The data will show
locations where travelers switch routes due to delays. Switching routes will indicate
some comfort, and possibly familiarity, with the network. Familiarity with the network
for the evacuation trip will also be indicated to some degree by the pre-evacuation tours.
The objective of this study is to develop a high fidelity integrated agent-based
simulation model based on a high resolution demand model capturing the complexity of
evacuation choices and the dynamics of routing behaviors during an evacuation process.
There is a considerable amount of research efforts in modeling traffic during evacua-
tions. A few of the evacuation models developed earlier include NETVAC (Sheffi et al.
1980), MASSVAC (Hobeika and Jamei 1985;HobeikaandKim1998), TEDSS (Sherali
et al. 1991), IMDAS (Franzese and Han 2001), OREMS (Rathi and Solanki 1993), and
CEMPS (Pidd et al. 1993). Typically these models are developed for specific types of
emergency situations. For example, MASSVAC is developed for traffic simulation in
hurricane evacuation; it requires evacuation routes as model input from all origins to all
destinations, where the origins and destinations depend on the projected path of the
hurricane. Very recently, the advancement in the development of dynamic traffic simula-
tion models encourages researchers to apply them in evacuation scenarios. For instance,
microscopic traffic simulation models, such as PARAMICS (Cova and Johnson
2003), CORSIM (Williams et al. 2007), VISSIM (Han and Yuan 2005),
MITSIMLab (Jha et al. 2004) and INTEGRATION (Mitchell and Radwan
2006), and meso-scopic or macroscopic models, such as DYNASMART
(Murray-Tuite 2007), DynaMIT (Balakrishna et al. 2008), DynusT (Noh et al.
2009), TransCAD (Wang et al. 2010), and INDY (Klunder et al. 2009)have
been applied to study evacuation problems.
A-RESCUE: An Agent based Regional Evacuation Simulator
Agent-based simulation is a computational methodology to model systems comprised of
interacting autonomous agents situated in an artificial environment (Macal and North 2005;
Flötteröd et al. 2012). These autonomous agents are self-directed objects with the capabil-
ities of making decisions and reacting to the environment. They also have the ability to
learn and adapt based on their goals and interactions with each other. Such a framework is
particularly suitable for simulating individual behaviors and exploring emergent collective
phenomena in evacuation. To capture some of the phenomena and complexities during
hurricane evacuations, we develop an agent-based model to simulate household-level
evacuation decisions and the resulting interactions among the evacuee drivers.
The overall framework of the agent-based simulation model consists of two major
components: Household Decision Making Module and Traffic Flow Module. The
household decision making module sets the rules for household agentsevacuation
behaviors obtained from advanced behavioral models. The behavioral model captures
the complexity of different dimensions related to hurricane evacuations such as deci-
sion to evacuate (yes or no), evacuation timing, evacuation destination choice, mode
choice, route choice and pre-evacuation and en-route evacuation activities as well as the
number of household vehicles used to evacuate. These behavioral models are devel-
oped, estimated and validated using the data from Hurricane Ivan including the
households from Alabama, Louisiana, Florida and Mississippi (evacuate/stay decision,
number of household vehicles used), and data from a behavioral-intention survey for
Miami (accommodate type, evacuation destination, mode assignment for non-personal
vehicles, departure time for the ultimate evacuation trip, activity participation and
scheduling). Behavioral models also address the issue of household level heterogeneity
by incorporating random parameters in the models.
Behavioral models are integrated with a microscopic traffic simulation module. This
module uses the output from the household decision-making model to design its input
(i.e. traffic demands). In the simulation, each household is an agent, which makes
various evacuation related decisions. From household decisions, a number of vehicles
are generated and entered in the evacuation transportation network at different time
intervals. The traffic flow in the network is based on a car-following model (Yang 1997)
that calculates a vehicles acceleration rate based on its relationship with the leading
vehicle.
To mi m ic the dynamic routing behavior of evacuees an en-route route choice model is
integrated with traffic simulation module. In this dynamic routing behavior, at each step,
agents can update information from the network and take the shortest route to the
destination. Finally, an adaptive routing strategy that can achieve efficient network-wide
traffic operations during evacuations is proposed. Computational results are presented
based on simulations over a medium-scale traffic network. The simulation results dem-
onstrate the improvement of traffic operations against driversroute switching behavior
based on the prevailing traffic conditions. The specific contributions of this study include:
1. Develop A-RESCUE, a high fidelity agent-based evacuation simulation model
integrating household evacuation behaviors and traffic flow behaviors.
2. Integrate a rich household level evacuation demand model governing the various
evacuation decisions.
3. Develop an adaptive routing strategy that can achieve efficient traffic operations
during evacuations.
S. V. Ukkusuri et al.
4. Provide computational results using the simulation of a medium size transportation
network for the evaluation of the proposed adaptive strategy against the instanta-
neous route switching behavior.
The remainder of the paper is organized as follows. The next section presents the
framework for the agent-based simulation model. The subsequent two sections describe
the household decision making module and the traffic flow module. The next section
presents the proposed adaptive routing strategy algorithm. The next section gives a
brief description of the developed simulation tool. The last two sections present a case
study with results and some concluding remarks, respectively.
2 Framework of the Agent-Based Simulation Model
In this section, a framework for the agent-based simulation model is presented. The
framework is developed based on the integration of household-level behavioral models
with the traffic flow models to fully capture the complexity of hurricane evacuations.
Based on the input household characteristics, the behavioral models set different decisions
related to the evacuation including the decisions to evacuate or not, timing of the
evacuation and destination of the evacuation for a household. Having made these deci-
sions, the household creates vehicle agents to enter the network. The vehicles make en-
route activity decisions and route choice decisions. Traffic interactions among the vehicles
in the network are created based on a car-following model. Figure 1presents a flow chart
of the framework. The framework is built upon two major contexts: household decision
making context and traffic simulation context.
Data containing spatial population distribution and socio-demographic characteris-
tics from the census are used in the Household Agent creation module to assign
geospatial locations and individuals attributed to each household. Household agents
then belong to the Household Decision Making context where different decisions
regarding evacuations are made by the household agent.
Household Decision Making context is one of the central components of the
proposed agent-based simulation model. In this context, the inputs include the socio-
demographic characteristics of the households and the characteristics of the ap-
proaching hurricane; and the outputs are the decisions moving the household agent
during the evacuation. The most representative decisions are whether to evacuate or
not, evacuation departure time and destination of the evacuation. First, a random-
parameter binary logit model for evacuation choice is used to make the decision
whether to evacuate or not. If the household decides to evacuate, the accommodation
type choice is then modeled with a multinomial logit model. Third, a binary logit model
and simple frequency based model are used for the destination. Fourth, the travel mode
is assigned based on whether the household owns vehicles. If not, they are assigned as a
passenger to be picked up by a family member, friend etc., or to transit. If the household
owns vehicles, the number of vehicles used is determined from a truncated Poisson
model. After the travel modes are assigned, the departure time is determined from a
hazard-based duration model. Based on these main evacuation decisions, activity
participation is then determined. First, a binary logit model determines whether a
household participates in out-of-home activities. If so, then the number of tours is
A-RESCUE: An Agent based Regional Evacuation Simulator
modeled with a frequency model and the specific activities are assigned based on
frequency from the Miami behavioral intention survey data. Then based on the type of
pre-evacuation activities and the households identified as passengers, pickups are
assigned. The details of these models can be found in Yin et al. (2014). Finally, a
multinomial routing strategy choice model is used to select a route from a set of routing
strategies including using their familiar routes, following the suggested evacuation
route, and switching routes based on the prevailing traffic conditions. For each house-
hold, this household decision making context determines the decisions using suitable
models that were previously calibrated. The context then generates vehicle agents that
interact with the transportation network in the Traffic Simulation context.
The Traffic Simulation context establishes the behavioral rules that guide the
vehicles from their origins to their destinations. If a household agent decides to
evacuate, a vehicle agent is generated and enters the transportation network at a specific
evacuation time. A pre-evacuation activity choice model is implemented to determine
the set of pre-evacuation activities of the household. An En-Route Route Choice Model
is integrated in the Traffic Simulation context to mimic the routing behavior of the
evacuees. This context uses a routing algorithm based on the k-shortest path adaptive
routing algorithm to find optimal routes guiding vehicle agents to their destinations.
The Traffic Simulation Module requires a GIS map of the transportation network as
an input to move the vehicles. Using a pre-processing step called data modeling, the
GIS map is converted to a detailed lane-based network that is required for the
simulation. Vehicles are moved in a lane based on a car-following model and can
switch lanes based on lane-changing decision model. More details about the data
modeling step and the traffic flow models are discussed in subsequent sections. Finally,
an output module is designed to analyze vehiclestravel times and total clearance time.
Fig. 1 Integration of the Household Decision Making Context with the Simulation Context
S. V. Ukkusuri et al.
The major distinction of the proposed decision-making context with the existing
multi-agent simulation models is as follows. Traditional simulation approaches usually
generate an evacuation demand for a region for a given population; and based on the
demand, they generate the traffic demand for the transportation network. These two
processes are done separately; specifically the vehicles moving in the network do not
retain the characteristics of the households they belong to. As a result, different aspects
of the decisions are not reflected in the subsequent decisions of the agents traversing the
network. For instance, for these approaches, the routing decisions made in the trans-
portation network cannot introduce the characteristics of the households used to
generate the evacuation demand. On the other hand, the proposed integrated agent-
based simulation approach can use the household-level characteristics to make the pre-
evacuation and intermediate activity and routing decisions. This integrated approach
can further be attributed to the feature of a truly dynamic simulation. The decisions
made in the transportation network can easily be incorporated to the household decision
making context. This will generate the evacuation demand on the fly based on the
prevailing traffic condition whereas the traditional simulation-based approaches gener-
ate the time-dependent evacuation demand beforehand. This truly dynamic nature is of
greater importance for an evacuation context where demand decisions change rapidly
as a function of traffic conditions and other social conditions thereby impacting the
supply side performance.
3 Household Decision Making Module
This section presents the household decision-making module that integrates household
evacuation behavior into the simulation. Households located in an area under hurricane
threat make multiple decisions for their evacuation. Researchers have developed models
to understand how characteristics of the households and the evacuation-related factors
such as hurricane trajectory and evacuation warning influence this decision.
The evacuation decision sub-module considers five decisions, as identified by Hasan
et al. (2011); Mesa-Arango et al. (2013); Sadri et al. (2014a); Murray-Tuite and Wolshon
(2013); Hasan et al. (2013) in a sequential manner. These decisions include (1) whether
to evacuate, (2) accommodation type selection (e.g., friends/relatives homes, public
shelters or hotels), (3) evacuation destination (the specific destination), (4) evacuation
mode (auto-based evacuation or carpooling), and (5) departure time. If the household
uses personal vehicles to evacuate this sub-module also determines the number of
vehicles used. All these decisions are modeled via probabilistic (frequency) or econo-
metric models relating relevant explanatory variables to the choice in question. The
departure time decision dictates the time window available during which pre-evacuation
activities such as purchasing gas, food/water, and medicine or withdrawing cash can
take place. The pre-evacuation activity sub-module captures these pre-evacuation
activities.
The pre-evacuation activity module adopts an activity-based approach, which views
the pre-evacuation activity-travel from a tour-stop perspective. This sub-module has
three components: (1) activity generation, which simulates the number of tours and
activities performed by a household, (2) passenger household assignment, where
households are assigned to pick up their friends and relatives with transportation needs,
A-RESCUE: An Agent based Regional Evacuation Simulator
and (3) activity scheduling, that assigns days, times and locations to these activities via
simulation.
For each household, the household decision making module determines a schedule
of pre-evacuation trips plus a final evacuation trip. Each of these trips is described by an
origin, destination, departure time, and activity duration. These trips are inputs into the
simulator, which are then assigned into the traffic network as illustrated in Fig. 1.
4 Traffic Flow Module
In A-RESCUE, the traffic flow model is an important component since it represents the
movement of vehicles in the traffic network and allows the modeler to capture various
performance measures of interest. The vehicle is the basic unit of the traffic flow model.
The traffic flow is implemented in three steps:
a) Vehicle loading
b) Computing acceleration or deceleration
c) Vehicle movement
4.1 Vehicle Loading
&At first, vehicles are loaded in pre-trip FIFO queue based on its evacuation time. To
enter into the network, at their corresponding evacuation times, vehicles look for
available leading spaces in the entrance link. If adequate leading spaces are not
available on the entrance link, vehicles are stored in the FIFO queue and wait to
enter the network during subsequent simulation steps.
&If necessary leading spaces are found, vehicles enter into the network through the
entrance links. Its initial position and speed are determined by the simulation step size,
the drivers desired speed, and the traffic conditions on the entrance link. If a vehicle
enters into the network successfully, it is removed from the pre-trip queue of the link.
&When a vehicle arrives at an activity location, the new departure time of this agent
will be computed based on its arrival time and the time duration it spends at the
location. The agent then is loaded to the queue and sorted based on its new
departure time computed. A vehicle is removed totally from the network only if it
reaches the final destination.
4.2 Computing Acceleration or Deceleration
To set the acceleration or deceleration behavior rule for the vehicle agents, we
implement a car-following model (Yang 1997) in the simulation. The car-
following model computes a vehicles acceleration rate based on its relationship
with the leading vehicle. Depending on the magnitude of its headway with
the front vehicle, a vehicle implements one of three rules: free flowing, car
following, and emergency decelerating (Herman et al. 1959;Hermanand
Rothery 1963;Wicks1977).
S. V. Ukkusuri et al.
Free Flowing Rule If the time headway is larger than a pre-determined threshold
h
upper
, the vehicle does not interact with the leading vehicle. The vehicle acceleration
rate is calculated based on the following relationships (Yang 1997)
an¼
aþ
nif vn<vtarget
n
0if vn¼vtarget
n
a
nif vn>vtarget
n
8
<
:
where:
a
n
acceleration rate;
a
n
+
maximum acceleration rate;
a
n
normal deceleration rate;
v
n
current speed;
v
n
targ et
target speed in current link
Emergency Decelerating Rule If a vehicle has a time-headway smaller than a pre-
determined threshold h
lower
, it is in the emergency stage. In this case, the vehicle uses an
appropriate deceleration rate to avoid collision and extend its headway (Yang 1997).
an¼min a
n;an10:5vnvn1
ðÞ
2
gn
()
if vn>vn1
min a
n;an1þ0:25a
n

if vnvn1
8
>
<
>
:
Car-Following Rule Finally, if a vehicle has a time headway between h
lower
and h
upper
,
it is in the car following stage. In this case the acceleration rate is calculated based on
Hermans general car-following model (Herman et al. 1959)
an¼αvβ
n
gγ
n
vn1vn
ðÞ
Where α
±
,β
±
and γ
±
are model parameters; α
+
,β
+
,γ
+
are used for accelerating (v
n
<v
n-
1
), and α
,β
,γ
for decelerating (v
n
>v
n-1
) cases, default values for these parameter
are based on Subramanian (1996).
4.3 Vehicle Movement
Once the acceleration or deceleration rate of a vehicle is computed for a simulation step
based on the car-following model described above, the vehiclesspeedandpositionare
updated. The movement function of the simulation tool is then used to move the vehicle
using its current speed. In addition, while moving on the network, a vehicle needs to
stay in a specific lane. Since road lanes of two consecutive links connect to each other
using some rules, the vehicle must be on a proper lane in order to travel from an
upstream link to a downstream link. The movement that allows a vehicle to shift from
A-RESCUE: An Agent based Regional Evacuation Simulator
an improper to a proper lane is referred as lane changing. In a more advanced model,
vehicles also change lanes to improve the driving condition such as desired speed,
visibility, etc. The model is embedded in the simulation and is discussed in the next
subsections.
4.4 Lane Changing Model
Lane changing is one of the fundamental components in the A-RESCUE traffic
simulator. It significantly impacts the traffic flow and thus affects the congestion level
of the transportation network. Lane changing, in most of microscopic simulators, is
classified into two types; mandatory and discretionary, (Gipps (1986), Yang (1997)and
Ahmed (1999)). Mandatory Lane Changing (MLC) happens when a driver must
change lanes in order to get on the lane that connects to a downstream link. Discre-
tionary Lane Changing (DLC) happens when a driver changes the lane to have a better
perceived traffic condition. Since MLC has higher hierarchy, it overrides all other
decisions in consideration. Figure 2illustrates the two types of lane changes. Vehicle
numbered 2 is going to make a left turn at the downstream junction, thus it needs to
change to the left most lane, the mandatory lane change needs to be made. Meanwhile,
the vehicle numbered 1 has no problem connecting to the downstream link in a through
movement and therefore no lane changing is required. However, the driver of this
vehicle might perceive a better traffic condition on the right lane and thus has an
incentive to change to that lane under the discretionary lane changing state.
To execute lane changing, the driver first seeks the target lane, the lane which either
connects to the downstream link in MLC or the lane with better perceived traffic
condition in DLC. In the MLC, the driver might need to change the lane several times
before getting to the target lane; however, the target lane in the DLC is most likely the
lane adjacent to the current lane of the vehicle. After the target lane is determined, either
in MLC or in DLC, the driver will seek acceptable gaps, the distances from his/her
vehicle to the leading vehicle, referred as lead gap, and to the lagging vehicle, referred
as lag gap, in the lane that the driver will change to. If these gaps are greater than
critical gaps (minimum thresholds) then the driver will change lanes, otherwise the
driver will stick to the current lane until the acceptable gap is available. Figure 3
21
Fig. 2 Illustration of different types of lane changes
S. V. Ukkusuri et al.
illustrates different gaps in lane changing. The framework and parameters of the lane-
changing model embedded in the simulation are based on the work from Ahmed
(1999).
5 Adaptive Routing Strategy for Hurricane Evacuation
In this section, an adaptive dynamic routing algorithm is presented that can achieve an
efficient network-level traffic operation during evacuations. The algorithm will yield
better travel times for the evacuees over switching routes based on the prevailing traffic
conditions. The proposed algorithm is based on the notion of an adaptive and
decentralized routing of agents (Mahmassani 2001;Farver2005). An adaptive route
choice model based on a simple logit-type splitting function is adopted. A good
adaptive based dynamic routing can yield good system states. Such an adaptive model,
correctly reacting to current traffic conditions, has a significant value especially when
real-time information is lagged/delayed. This type of routing approach is particularly
useful to model situations when traffic networks are congested due to emergency
situations such as hurricane evacuations. Such evacuations will have rapid fluctuations
of demand and capacity forcing travelers to take travel decisions based on the current
traffic situation (i.e., instantaneous travel time).
In this approach, vehicles have information of link travel times at the intermediate
nodes and, using this information, update their routes from the current node to the
destination. Consider a vehicle igoing from origin node rto a destination node s.A
route pdenotes the path from the decision node jto the destination node s. The problem
is to assign vehicle ito an outgoing link aB(j), where B(j) is the set of links incident
to node j; such decisions are made repeatedly upon reaching the next decision node,
until vehicle ireaches destination node s.
Traffic assignment based on reactive dynamic user equilibrium requires each agent
to follow the shortest route at every intermediate decision node. However, an iterative
approach is necessary to anticipate the impact of following the shortest route at the
intermediate node based on the instantaneous travel time. This will be computationally
burdensome and may be difficult for real-time deployment. Thus different heuristic
rules for route assignment can be developed based on the criteria for evaluating routes
at intermediate nodes. Here we propose a heuristic where agents, instead of following
the shortest route, adopt mixed routing strategies based on a logit-based splitting
function.
lag vehicle lead vehicle
Subject
vehicle Front vehicle
lag gap Lead gap
Fig. 3 Gap acceptance in lane change
A-RESCUE: An Agent based Regional Evacuation Simulator
Denote by ψ
i,p
js
(t) that the instantaneous travel time of route p, for vehicle i,from
node jto node sat time t. The probability of any feasible route pis inversely
proportional to the instantaneous travel time, ψ
i,p
js
(t) and is given by:
βjs
i;ptðÞ¼
exp θψjs
i;ptðÞ
hi
Xpexp θψjs
i;ptðÞ
hi
The above equation allocates a vehicle to a feasible route pbased on ψ
i,p
js
(t). The θ
parameter affects the probability of using each route by assigning higher probability to
a route with less instantaneous travel time.
The pseudo code of the algorithm, as applied by a vehicle i, is described below. The
algorithm describes the routing of a vehicle from its current node jto its destination s.
We denote by A(j)
p
the node after jon route p. The vehicles current route is
denoted by p'.
6DataModeling
An important module in A-RESCUE is the representation of various lanes in the traffic
network that will allow realistic movement and routing of vehicles. The purpose
of the data modeling module is to support the traffic simulation module in
handling the complex network connectivity and improving the simulation visu-
alization. Without this module, its impossible to simulate the traffic movement
on different lanes within Repast.
In addition, our data modeling is more advanced than the GIS modules from the
traffic simulation software packages such as SYNCHRO or VISSIM because they
automatically convert the large-scale link-based GIS data into the lane-based network
with full connectivity.
S. V. Ukkusuri et al.
The data modeling module is developed based on the work by some of the authors and
is published in Su et al. (2015). The idea is summarized as follows. Before running the
simulation in Repast, we prepare the supply side (road networks) in such a way that Repast
can read and assign traffic demand directly. Specifically, the data modeling offers:
(1) Lane visualization
By using Repast directly without the data modeling module, the vehicles can
only move on a center line and we cannot differentiate vehicles in opposite
directions (Fig. 4).
With the data modeling module, we can separate vehicles by lanes. As the
results show, we will never observe vehicles Bcrashing^when they are passing
each other or moving on opposite directions (Fig. 5).
(2) The link and lane connectivity
Link-lane connectivity at intersections is extremely difficult to handle in Repast
coding. On the other hand, the data modeling consists of a procedure to trim the
lanes at intersections so that vehicles do not move in an unrealistic
trajectory (Fig. 6).
In our data modeling module, we not only offset each link to obtain
different lanes but also automatically trim the beginning and ending seg-
ment of the lanes. Without this step, vehicles will move along the entire
lane before shifting to the next lane, which causes the impractical turns.
By trimming the lane ends, vehicles can approach the next lanes allowing
for appropriate movements. An example is shown in Fig. 6.
(3) Link-link connectivity and lane-link connectivity.
This feature keeps track of the connection among all segments in the network
especially at the complex at-grade intersections and interchanges.
Thedatamodelingmoduleoutputsatableforeachlink(Table1) and another table for
each lane (Table 2).
Those data allow us to determine all connectivity in the road networks which is very
important for the traffic simulation. At a certain time when a vehicle is moving in a lane
of a link, we know exactly which link and lane will be used next, given the routing for
that vehicle has been defined.
A key difference from other simulation based tools such as SYNCHRO, VISSIM,
etc., which require one to specify the link/lane connectivity manually, the proposed data
modeling can automatically generate not only the connectivity but also the network
visualization for very big networks.
Figure 7shows the results of data modeling in the Miami-Dade county network. All
roads have been offset and trimmed. The road and link connectivity is output in a data
file and this can be used directly within Repast simulation. In Fig. 8, we magnify the
view of one interchange. We can see how links and lanes connect together. In Repast
simulation, vehicles will move along the purple lines which represent the center
of the lanes.
Fig. 4 Visualization of vehicles in a single lane without data modeling
A-RESCUE: An Agent based Regional Evacuation Simulator
In addition to the three main features above, the data modeling module can also
perform the following tasks:
1. Find the intersections with certain number of legs. This function enables us to
enumerate the ID of intersections with different legs. For example, the five-leg
intersections must be pre-processed before running the data modeling.
2. Prepare the signalized properties for intersections. This function is designed for the
very detailed level of simulation with signal control. For each intersection (node), it
enumerates all of the links that are going into it. The result will be displayed in the
node layer.
3. Use the hybrid model with background traffic. It enables the data modeling to add
additional columns to the road shape file that corresponds to hourly speeds (e.g.
24 h) in the link file extracted from TRANSCAD.
4. Find zigzags in the shape file and help avoid erroneous trajectories of vehicles in
the Repast simulation.
The data modeling module is coded in Visual Basic as an add-on in ArcMap. It is
called from ArcMap to run specific tasks such as building connectivity, trimming lanes,
finding potential error (five-leg intersections), etc. The data modeling inputs a link-
based map within a standard GIS shape file. The output of this procedure is the lane-
based shape file and new link-based shape file, which can be input and used directly in
Repast simulation module.
Fig. 5 Lanes are separated from link in data modeling
Fig. 6 Trimming lanes in data modeling to avoid unrealistic vehicle trajectory
S. V. Ukkusuri et al.
7 Simulation Tool Development
The simulation model is developed using the Repast Simphony (Macal and North
2005;North et al. 2005) agent-based toolkit.
The model includes four high-level components that play an important role in the
simulation.
a) Network Constructor: The purpose of this module is to build transportation
networks from shape files generated by the data modeling. The shape files here
include three types: Point, Poly Line (which generates roads and lanes), and
Polygon (which generates traffic analysis zone (TAZ) centroids). The network
includes roads and lanes on which evacuee agents travel, intersections, and TAZ
centroids which are the origins, activity locations, and destinations of evacuee
agents.
Repast Simphony has an extensive library to incorporate shape files into the
simulation. We have used data modeling (see Section 6) to modify the shape files
of the networks to be used in the simulation. We then use the GIS library of Repast
Simphony to read the network files.
Tabl e 1 Link data from the data
modeling Link ID To node
Length Left number of lanes
Direction Right number of lanes
Name Lane #1
Class Lane #2
Number of lanes Lane #3
Speed limit Lane #4
Left Lane #5
Through Lane #6
Right Twin lane #1
Twin link ID Twin lane #2
Twin link left Twin lane #3
Twin link through Twin lane #4
Twin link right Twin lane #5
From node Twin lane #6
Tabl e 2 Lane data from data
modeling Lane ID
Link ID
Left lane
Through lane
Right lane
Length
A-RESCUE: An Agent based Regional Evacuation Simulator
b) Agent Constructor: This module generates agents on the basis of action rules that
are pre defined. The evacuee agents first build the travel plan, i.e., the list and
visiting sequence of the agents including duration the agents spend at intermediate
locations. The agents plan is based on the output of the behavioral models. The
way evacuee agents are loaded to and travel on the network is discussed in Section
4.
Fig. 7 Miami-Dade county road network after running the data modeling
Fig. 8 Visualization in a specific interchange
S. V. Ukkusuri et al.
The creation of agents in Repast starts by building a context for the agents. A
context is a named set of agents. In simple words, a context is a bucket full of
agents. A context is named and the agents are added to it. The current simulation
implements various contexts to create agents including the network features. For
example, we have the following contexts for different agents in our simulation:
road, lane, junction, house and vehicle. The simulation structure in flexible enough
to include any other agents. For instance, to include signals we just need to
implement a signal context in the simulation.
A related concept is the geography. If a context is a virtual bucket to an agent,
the geography is a physical bucket to the agent. If an agent is required to be
visualized in the display, it must have geography. When creating contexts, we have
to specify the schedule of the simulation events. A schedule is an instruction for the
execution of the events. The simulation procedure followed here is time-depen-
dent, hence the start time, end time and recurrence interval has to be mentioned in a
schedule.
c) Displayer: This module mainly displays the animation of simulation dynamics and
results. The display module shows vividly the interaction among evacuee agents
and the level of congestion on the network. Through the display module, a specific
vehicle can also be tracked and its trajectory can be visualized. It can also simul-
taneously show additional information (e.g., speed, origin, destination, next road
and so on) about the vehicle in a separate window.
d) Simulation Context: This module is the global environment. In the environment,
each autonomous agent proceeds to its destination subject to certain imposed con-
straints and interacts with the environment and other agents. It also provides real-time
information about other household agents as well as the transportation network during
the simulation.
The proposed simulation model is a decentralized multi-agent system. That is, there
is no central control mechanism. The evacuee agent acts independently based on local
information and the network conditions. Each agent has the ability to update
the route information and change its route if necessary. From the simulation,
statistical information of evacuation can be obtained. The information includes
total evacuation time for each agent, network clearance time, maximum and
minimum travel time, and more. Detail information of each evacuee agent,
departure time, route, speed, and total evacuation time can also be obtained
easily based on its trajectory.
The simulator can be run in both Windows and Unix operating systems. It is
convenient to visualize the simulation in Windows as a desktop application. However,
all the large-scale simulation runs are conducted in a Unix operating system in a cluster
computing environment with 64 bit, dual 12-core AMD Opteron 6172 processors (24
cores per node) and 96 GB of memory for each node.
8 Data Collection
This section presents the data and scenarios designed for the numerical exper-
iments. The data are extracted from the outputs of the behavioral models
A-RESCUE: An Agent based Regional Evacuation Simulator
presented in Yin et al. (2014). The outputs from Yin et al. (2014)present
overall data of evacueesbehavior and include all modes of transportation used
in the evacuation. A-RESCUE uses the data from households that only use
vehicles to evacuate. Thus, all households that do not use vehicles to evacuate
are removed from the input data of the A-RESCUE model.
A-RESCUE was tested with a scenario which included pre-evacuation activities, en-
route evacuation activities and direct evacuation. The network data and the demand
data have the following characteristics:
&Network size: number of links: 8578, number of zones: 364
&Total number of evacuation vehicles: 89,379
&Total number of trips (including pre-evacuation trips): 162,002
&Vehicle classification according to trips:
Vehicles make pre-evacuation activities (35168, 39.35 %)
Make multiple trips and return to home (587, 0.66 %)
Directly evacuate (53624, 60 %)
&Scenarios are based on different routing strategy, rather than demand level. Seven-
teen routing scenarios were tested:
Single shortest path routing (1 case)
K-shortest path routing (total 16 cases)
k=2and 3
θ= 0.25, 0.5, 0.75, 1, 1.25, 1.5, 1.75, 2
9 Computational Results and Analysis
9.1 Demand Patterns
Figure 9shows the departure time distribution across the 6 days of evacuation
period. It is found that the peak demands occur in the middle of a day. This
result is consistent with previous literature, which suggests that hurricane
evacuees prefer to depart earlier in the day so that they can reach their
destinations by nightfall (Lindell et al. 2005). A trip is made either to evacuate
to a safe place directly or to participate in preparatory activities. About 40 %
(n= 35,168) of these trips are made to participate in preparatory activities while
about 60 % (n= 53,624) of these trips are directly related to evacuation to a
final destination.
Figure 10 shows the spatial distributions of the evacuation trips aggregated
over the entire period of evacuation of 136 h. Figure 10a shows the total
number of trips (including evacuation and pre-evacuation trips) departing from
the zones; Fig. 10b shows the total number of starting trips of the trips chains
for each vehicle departing from the zones; and Fig. 10c shows the total number
of final evacuation trips departing from the zones. In general, we find that
S. V. Ukkusuri et al.
southern zones generated most of the trips. This is due to the imminent threat
of the hurricane to these zones.
(a) (b) (c)
Fig. 10 Spatial distributions of evacuation trips
Fig. 9 Departure time distribution
A-RESCUE: An Agent based Regional Evacuation Simulator
9.2 Evacuation Travel Time Analysis
In a given demand scenario, we run the simulation for different routing parameters.
Figure 11 shows the distribution of travel times. As a base case, we run the simulation
for single shortest path routing. Then we run the simulation for k-shortest path routing
for k= 2 and 3 and θ= 0.25, 0.5, 0.75, 1, 1.25, 1.5, 1.75, 2. The higher the value of kthe
more options an evacuating vehicle will consider for choosing routes. The higher the
value of θthe more sensitive the agent will be to the differences in travel times across
different routes. We find that the improvement due to higher values of kis very modest.
The improvements in reducing travel times become greater for higher value of θwhen
its value is below or equal to 1. It is observed that the different values of kand θdo not
significantly impact the evacuation time.
9.3 Analysis of Pre-Evacuation Activities
We found that about 40 % (n= 35,168) of the total trips are made to participate in
preparation related activities before evacuation. In this section, we investigate if the
length of preparation time impacts the evacuation travel time. Figure 12 shows the
scatter plots of evacuation travel times against pre-evacuation travel times over different
routing parameters. High pre-evacuation trip travel times result in lower evacuation
time, perhaps due to the selection of nearby final destinations reducing the overall
travel time. It is found that routing parameter has an impact on local congestion
reflected though the high pre-evacuation travel times. For instance, for k= 2 and θ=
0.25 the maximum of pre-evacuation trip travel time is lower than that of for k=3,and
θ= 0.25. This indicates that increasing the routing options has increased the pre-
evacuation trip times for some of the travelers. However, for a higher value of θ,
opposite results are found; consider the case k= 2 and θ= 0.75 against the case k=3
and θ=0.75.
Fig. 11 Distribution of trip travel time under different routing parameters settings
S. V. Ukkusuri et al.
9.4 Spatial Level Analysis
In this section, we present the spatial distributions of various aspects of evacuation trips
(Fig. 13). Note that these trips do not include the pre-evacuation trips.
Figure 13a, b, and hshow the average evacuation trip travel time, average evacu-
ation trip distance for each departing zone, and average evacuation trip departure time
for each departing zone, respectively. We find that southern zones near the coast have
higher evacuation travel times, and longer evacuation distance since they are directly in
Fig. 12 Scatter plots of evacuation travel times vs. travel time of pre-evacuation trips
A-RESCUE: An Agent based Regional Evacuation Simulator
the path of the hurricane and choose destinations farther away. However, the departure
times are evenly distributed across all the departing zones. This is due to impacts of
various socio-economic characteristics of the households affecting the departure
time choice.
Figure 13c shows the average final evacuation trip speed for each departing zone.
Most of the zones have similar speeds.
Fig. 12 (continued)
S. V. Ukkusuri et al.
Figure 13d and eshow the maximum evacuation trip travel time and maximum
evacuation trip distance for each departing zone, respectively. A few of the southern
zones experience very high maximumtravel timesbecause of the long distances traveled.
Figure 13g shows the clearance time for evacuation trips for each departing zone.
Most of the southern zones have very high clearance times because of long distance and
higher number of evacuation trips.
Fig. 12 (continued)
A-RESCUE: An Agent based Regional Evacuation Simulator
10 Summary and Conclusions
In this paper, we present A-RESCUE, a high fidelity multi-agent simulation model that
integrates household-level activity behavior with a network-level traffic assignment to
evaluate a broad range of evacuation strategies. The novelty of the proposed multi-
agent simulation model lies as follows:
&The proposed integrated approach can use the household-level characteristics to
make evacuation related decisions featuring a truly dynamic simulation. The
situations observed in the transportation network can easily be incorporated to the
household decisions generating evacuation demand on the fly based on the prevail-
ing traffic condition. This feature is of greater importance for an evacuation context
avg travel time (min)
(a) Average travel time (min)
avg distance (km)
(b) Average trip distance (km)
avg speed (m/s)
(c) Average trip speed (m/s)
max travel time (min)
(d) Maximum travel time (min)
max distance (km)
(e) Maximum trip distance (km)
max speed (m/s)
(f) Maximum trip speed (m/s)
clearance time (hour)
(g) Clearance time (hour)
avg departure time (hour)
(h) Average departure time (hour)
Fig. 13 Spatial distributions of evacuation trips (a) Average travel time (min) (b) Average trip distance (km)
(c) Average trip speed (m/s) (d) Maximum travel time (min) (e) Maximum trip distance (km) (f)Maximumtrip
speed (m/s) (g) Clearance time (hour) (h) Average departure time (hour)
S. V. Ukkusuri et al.
where demand changes rapidly as a function of traffic conditions, hazard informa-
tion and other social phenomena. Thus, through an integrated modeling approach,
A-RESCUE captures the complexity of household-level decision-making and the
dynamic traffic flows during an evacuation process.
&The proposed simulation approach assigns a rich set of behavior for household
agents including whether to evacuate, the pre-evacuation and intermediate activi-
ties, accommodation type selection, final evacuation destination, evacuation mode,
and departure time. All these decisions are modeled via probabilistic (frequency) or
econometric models relating relevant explanatory variables to the choice in ques-
tion. One of the most important capabilities of the simulation is the pre-evacuation
activity module, which adopts an activity-based approach and views the pre-
evacuation activity-travel from a tour-stop perspective. Most evacuation models
are likely to underestimate the demand because of ignoring these pre-evacuation
activities.
&A-RESCUE implements an adaptive route choice model based on a simple logit-
type splitting function. Such an adaptive model, correctly reacting to current traffic
conditions, has a significant value especially when real-time information is lagged/
delayed. Typically, evacuations have rapid fluctuations of demand and capacity
forcing travelers to take travel decisions based on the current traffic situation where
an adaptive routing approach is particularly useful.
&A-RESCUE represents the embedded traffic network in the most detailed way
possible for a simulation allowing realistic movement and routing of vehicles. It
uses data modeling approaches to support the traffic simulation module handling
the complex network connectivity and improving the visualization. The data
modeling module of A-RESCUE is more advanced than the GIS modules from
the microscopic traffic simulation software packages as it converts a large-scale
link-based GIS data into a lane-based network with full connectivity.
This paper presents the basic features of the multi-agent simulation model and the
initial findings based on the simulation runs on a medium-scale regional traffic
network. The various future directions that the opportunities of the simulation model
can be utilized:
&The behavioral decisions modeled in A-RESCUE are made in a sequential manner.
The model can be improved significantly if a dynamic joint decision-making
context is simulated.
&Currently the simulation system does not model how the warning information is
propagated and processed and how it influences the evacuation decisions. Consid-
ering an underlying social network among the agents and modeling a dynamic
decision-making context, such phenomena can be simulated.
&We have not simulated the process of disseminating real-time information, the value
of such information and the decisions resulting from it for managing evacuation
traffic.
&On the traffic network side, there can be various traffic flow restrictions (e.g.
contra-flow) and signal priorities for managing evacuation traffic efficiently. It
will be particularly useful to test such traffic options in an integrated simulation
system.
A-RESCUE: An Agent based Regional Evacuation Simulator
&We have not compared simulation run-times under various scenarios. A very
important direction will be to test the scalability of our approach and to
implement high-performance distributed computing algorithms for the simu-
lation system without losing the fidelity of the social and engineering
system characteristics.
Acknowledgments This research presented in this paper was supported by National Science Foundation
Awards SES-0826874 and CMMI 1520338 for which the authors are grateful. However, the authors are solely
responsible for the findings of the research work.
References
Ahmed KI (1999) Modeling driversacceleration and lane changing behavior. Doctoral dissertation,
Massachus etts Institute of Technology
Balakrishna R, Wen Y, Ben-Akiva M, Antoniou C (2008) Simulation-based framework for transportation
network management for emergencies. Trans Res Record, J Transportation Research Board 2041:8088
Bish DR (2011) Planning for a bus-based evacuation. OR Spectr 33(3):629654
Cova TJ, Johnson JP (2003) A network flow model for lane-based evacuation routing. Transp Res
A 37:579604
Farver J (2005) Hybrid vehicle-centric route guidance. Ph.D. Dissertation, Massachusetts Institute Technol
Flötteröd G, Chen Y, Nagel K (2012) Behavioral calibration and analysis of a large-scale travel
microsimulation. Networks Spatial Econ 12(4):481502
Franzese O, Han LD (2001) Traffic modeling framework for hurricane evacuation. Internal report, Oak Ridge
National Laboratory, Oak Ridge
Gipps PG (1986) A model for the structure of lane-changing decisions. Transp Res B Methodol 20(5):403
414
Han LD, Yuan F (2005) Evacuation modeling and operations using dynamic traffic assignment and most
desirable destination approaches. Proceedings of the 84th Annual Meeting Transportation Research
Board, Washington
Hasan S, Ukkusuri S, Gladwin H, Murray-Tuite P (2011) A behavioral model to understand household level
hurricane evacuation decision making. J Transp Eng 137(5):341348
Hasan S, Mesa-Arango R, Ukkusuri S (2013) A random-parameter hazard-based model to understand
household evacuation timing behavior. Trans Res Part C: Emerging Technol 27:108116
He X, Peeta S (2014) Dynamic resource allocation problem for transportation network evacuation. Networks
Spatial Econ 14(34):505530
Herman R, Rothery RW (1963) Car-following and steady state flow. Theory Traffic Flow Symp Proc, 113
Herman R, Montroll EW, Potts R, Rothery RW (1959) Traffic dynamics: analysis of stability in car-following.
Operation Res 1(7):86106
Hobeika AG, Jamei B (1985) MASSVAC: a model for calculating evacuation times under natural disaster.
Proceedings of the Computer Simulation in Emergency Planning Conference, La Jolla
Hobeika AG, Kim C (1998) Comparison of traffic assignments in evacuation modeling. IEEE Trans Eng
Manag 45(2):192198
Jha M, Moore K, Pashaie B (2004) Emergency evacuation planning with microscopic traffic simulation. Trans
Res Record, J Trans Res Board 1886:4048
Klunder G, Terbruggen S, Mak J, Immers B (2009) Large-scale evacuation of the Randstand: evacuation
simulations with the dynamic traffic assignment model Indy. Proceedings of the 1st International
Conference on Evacuation Modeling and Management, The Hague
Lindell MK, Prater CS (2007) Critical behavioral assumptions in evacuation time estimate analysis for private
vehicles: examples from hurricane research and planning. J Urban Planning Dev 133(1):1829
Lindell MK, Lu J-C, Prater CS (2005) Household decision making and evacuation in response to Hurricane
Lili. Natural Hazards Rev 6(4):171179
Lindell MK, Prater CS, Peacock WG (2007) Organizational communication and decision making for
hurricane emergencies. Natural Hazards Rev 8(3):5060
S. V. Ukkusuri et al.
Lindell MK, Kang JE, Prater CS (2011) The logistics of household hurricane evacuation. Nat Hazards 58(3):
10931109
Macal CM, North MJ (2005) Tutorial on agent-based modeling and simulation. In Proceedings of the 2005
Winter Simulation Conference, eds. Kuhl ME, Steiger NM, Armstrong FB, Joines JA, 215. Piscataway,
New Jersey: Institute of Electrical and Electronics Engineers, Inc
Mahmassani HS (2001) Dynamic network traffic assignment and simulation methodology for advanced
systems management applications. Networks Spatial Econ 1:267292
Mesa-Arango R, Hasan S, Ukkusuri S, Murray-Tuite P (2013) Household-level model for hurricane evacu-
ation destination type choice using hurricane Ivan data. Natural Hazards Rev 14(1):1120
Mileti DS, Sorensen JH, OBrien PW (1992) Toward an explanation of mass care shelter use in evacua tions.
Int J Mass Emergencies Disasters 10(1):2542
Mitchell SW, Radwan E (2006) Heuristic prioritization of emergency evacuation staging to reduce clearance
time. Proceedings of the 85th Annual Meeting Transportation Research Board, Washington
Murray-Tuite P (2007) Perspectives for network management in response to unplanned disruptions. J Urban
Plan Dev 133(1):917
Murray-Tuite P, Wolshon B (2013) Evacuation transportation modeling: an overview of research, develop-
ment, and practice. Trans Res Part C:Emerg Technol 27:2545
Naghawi H, Wolshon B (2012) Performance of traffic networks during multimodal evacuations: simulation-
based assessment. Natural Hazards Rev 13(3):196204
Noh H, Chiu YC, Zheng H, Hickman M, Mirchandani P (2009) An approach to modeling demand and supply
for a short-notice evacuation. Trans Res Record, J Trans Res Board 2091:9199
North MJ, Howe TR, Collier NT, Vos JR (2005) The Repast Simphony runtime system. In Proceedings of the
Agent 2005 Conference on Generative Social Processes, Models, and Mechanisms
Pidd M, de Silva FN, Eglese R (1993) CEMPS: a configurable evacuation management and planning
systema progress report. In: Proceedings of the 1993 Winter Simulation Conference
Rathi AK, Solanki RS (1993) Simulation of traffic flow during emergency evacuations: a microcomputer
based modeling system. In: Proceedings of the 1993 Winter Simulation Conference
Sadri AM, Ukkusuri SV, Murray-Tuite P, Gladwin H (2014a) Analysis of hurricane evacuee mode choice
behavior. Trans Res part C: Emerging Technol 48:3746
Sadri AM, Ukkusuri SV, Murray-Tuite P, Gladwin H (2014b) How to evacuate: model for understanding the
routing strategies during hurricane evacuation. J Transp Eng 140(1):6169
Sheffi Y, Mahmassani HS, Powell W (1980) NETVAC: a transportation network evacuation model. Center for
Transportation Studies, Boston
Sherali HD, Carter TB, Hobeika AG (1991) A location-allocation model and algorithm for evacuation
planning under hurricane/flood conditions. Transp Res B 25(6):439452
Su X, Cai H, Luong B, Ukkusuri S (2015) From a link-node-based network representation model to a lane-
based network representation model: two-dimensional arrangements approach. J Comput Civil Eng 29(3)
Subramanian H (1996) Estimation of a car-following model for freeway simulation. Masters thesis,
Massachusetts Institute of Technology, Cambridge, MA
Tarhini H, Bish DR (2015) Routing strategies under demand uncertainty. Networks Spatial Econ, 121
Wang H, Andrews S, Daiheng N, Collura J (2010) Scenario-based analysis of transportation impacts in case of
dam failure flood evacuation in Franklin county, Massachusetts. Proceedings of the 89th Annual Meeting
of the Transportation Research Board, Washington
Wicks DA (1977) INTRAS - a microscopic freeway corridor simulation model. Overview Simulation in
Highway Transportation 1:95107
Williams B, Tagliaferri A, Meinhold S, Hummer J, Rouphail N (2007) Simulation and analysis of freeway
lane reversal for coastal hurricane evacuation. J Urban Plan Dev 133(1):6172
Wu HC, Lindell MK, Prater CS (2012) Logistics of hurricaneevacuation in Hurricanes Katrina and Rita. Trans
Res Part F: Traffic Psychology Behaviour 15(4):445461
Yang Q (1997) A simulation laboratory for evaluation of dynamic traffic management systems, PhD. Thesis,
Massachus etts Institute of Technology
Yao T, Mandala SR, Do Chung B (2009) Evacuation transportation planning under uncertainty: a robust
optimization approach. Networks Spatial Econ 9(2):171189
Yin W, Murray-Tuite P, Ukkusuri SV, Gladwin H (2014) An agent-based modeling system for travel demand
simulatio n for hurricane evacuation. Transp Res C 42:4459
Yu-Ting H, Peeta S (2015) Clearance time estimation for incorporating evacuation risk in routing strategies for
evacuation operations. Networks Spatial Econ 15(3):122
Zhang Z, Parr SA, Jiang H, Wolshon B (2015) Optimization model for regional evacuation transportation
system using macroscopic productivity function. Transp Res B Methodol 81:616630
A-RESCUE: An Agent based Regional Evacuation Simulator
... Agent-Based Modeling (ABM) has been widely used in modeling evacuation for different types of disasters, such as flooding (Dawson, Peppe, and Wang, 2011;Liu and Lim, 2018;Uno and Kashiyama, 2008), tsunamis (Liu, and Lim, 2016;Mas, Koshimura, Imamura, Suppasri, Muhari, and Adriano, 2015;Wang et al., 2016;Mostafizi, Wang, Cox, and Dong, 2019), volcano eruptions (Carver and Quincey, 2016), hurricanes (Ukkusuri, Hasan, Luong, Doan, Zhan, Murray-Tuite, and Yin, 2017), fires (Adam and Dugdale, 2018), and earthquakes (Cimellaro, Ozzello, Vallero, Mahin, and Shao, 2017). ABM simulates the actions and interactions of the individuals, and then the effect of their actions and interactions on the whole system can be addressed. ...
Thesis
Full-text available
The effectiveness of an emergency evacuation plan highly depends on human behaviors and transportation systems. It is essential to study human behavior, predict their decisions, and estimate the capacities of the transportation system to plan for a safe evacuation. In this study, an agent-based model (ABM) is developed to simulate the behavior of people and estimate the endangered population and clearance time during a short-notice tsunami evacuation. The model is implemented for Waikiki, Hawaii, as the study area. This study aims to estimate the vulnerable population and evaluate different travel modes and horizontal and vertical evacuation effectiveness to consider them in emergency evacuation planning. Three types of agents are defined based on their evacuation modes: pedestrians, bicyclists, and car drivers. The response of agents to the evacuation alarm is defined based on survey findings. After a milling time, agents decide to evacuate or not. If they choose to evacuate, two types of shelters can be selected: vertical evacuation shelter or horizontal evacuation to safe zones. Four scenarios are implemented to estimate the fatality rate and clearance time in case of using different strategies: 1) based on the real situation, 2) constructing a new exit path, 3) encouraging people to evacuate on foot, 4) and encouraging people to evacuate vertically. Based on the results of the scenarios, suggestions can be made to improve the current emergency evacuation plan.
... (Ghurye, Krings, & Frias-Martinez, 2016) The study compares the observed human behavior during a disaster with the behavior expected under normal circumstances to understand the causal effects of the disaster event. Yin et al. combined mobile phone location data with agent based simulations (which are widely used in evacuation analysis; e.g., Ukkusuri et al., 2017) to improve the estimation accuracy of evacuation movement, proposing a hybrid approach. (Yin et al., 2020) More computational approaches using data assimilation techniques have been explored for near real-time predictions of post-disaster mobility patterns. ...
Article
Full-text available
Rapid urbanization and climate change trends, intertwined with complex interactions of various social, economic, and political factors, have resulted in an increase in the frequency and intensity of disaster events. While regions around the world face urgent demands to prepare for, respond to, and to recover from such disasters, large-scale location data collected from mobile phone devices have opened up novel approaches to tackle these challenges. Mobile phone location data have enabled us to observe, estimate, and model human mobility dynamics at an unprecedented spatio-temporal granularity and scale. The COVID-19 pandemic, in particular, has spurred the use of mobile phone location data for pandemic and disaster management. However, there is a lack of a comprehensive review that synthesizes the last decade of work and case studies leveraging mobile phone location data for response to and recovery from natural hazards and epidemics. We address this gap by summarizing the existing work, and point to promising areas and future challenges for using mobile phone location data to support disaster response and recovery.
... The use of ABMs for routes has been studied in diverse fields, such as logistics and transport (e.g., Badland et al., 2013;Batet et al., 2012;Dia, 2002;Elbert et al., 2020;Nasir et al., 2014;Nguyen-Trong et al., 2017;Ukkusuri et al., 2017). In relation to ABMs in tourism (or movement of individuals within a fixed area), Table 2 compares 10 aspects of different ABM proposals. ...
Article
Full-text available
Spain ranks second in the world for the number of international tourists. These tourists have different preferences, which influence their choice of tourist routes depending on the activities offered by provinces. There are currently no routes customised according to the preferences of a travel party, which makes the supply of tourist packages complex. We propose an agent-based model, named ABM RoutePlanner, which simulates the behaviour of travel parties through provinces of Spain. The model is developed as an application of the ODD protocol. This paper makes two contributions. First, we describe a model applicable to the identification of appropriate routes for different combinations of tourists’ preferences in some (hypothetical) environments. Second, we present the actual routes that will allow tour operators to define strategic plans that motivate tourists to visit the provinces included in the routes. The simulation model was calibrated with data extracted from TripAdvisor and Spanish tourist surveys.
... ABMs were chosen to represent the human system (Fig. 1b) and the built environment (Fig. 1c) as the models capture individual's decisionmaking processes and interactions between agents, making them excellent tools for investigating complex system dynamics [7,9,31,32]. Another reason is that ABMs have proven capable of simulating evacuation decisions in hurricanes [6,13,18,[24][25][26], and hurricane traffic dynamics [33][34][35][36]. One drawback of ABMs is their high computational expense, which makes them less suitable for operational use (e.g., pros/cons in Ref. [37]). ...
Article
Hurricane evacuations involve many interacting physical-social factors and uncertainties that evolve with time as the storm approaches and arrives. Because of these complex and uncertain dynamics, improving the hurricane-forecast-evacuation system remains a formidable challenge for researchers and practitioners alike. This article introduces a modeling framework built to holistically investigate the complex dynamics of the hurricane-forecast-evacuation system i.e., to determine which factors are most important and how they interact across a range of real or synthetic scenarios. The modeling framework, called FLEE, includes models of the natural hazard (hurricane), the human system (information flow, evacuation decisions), the built environment (road infrastructure), and connections between systems (forecasts and warning information, traffic). In this paper, we describe FLEE's conceptualization and implementation and present proof-of-concept experiments illustrating its behaviors when key parameters are modified. In doing so, we show how FLEE is capable of examining the dynamics of the hurricane-forecast-evacuation system from a new perspective that is informed-by and builds-upon empirical work. This information can support researchers and practitioners in hazard risk management, meteorology, and related disciplines, thereby offering the promise of direct applications to mitigate hurricane losses.
... We had originally developed an agent-based evacuation simulator A-RESCUE to overcome the limitations of standard agent-based traffic simulators [19]. We improved its performance and scalability for this study, the details of which are provided in its second version [20]. ...
... The study compares the observed human behavior during a disaster with the behavior expected under normal circumstances to understand the causal effects of the disaster event. Yin et al. combined mobile phone location data with agent based simulations (which are widely used in evacuation analysis; e.g., 45 ) to improve the estimation accuracy of evacuation movement, proposing a hybrid approach 46 . ...
Preprint
Full-text available
Rapid urbanization and climate change trends are intertwined with complex interactions of various social, economic, and political factors. The increased trends of disaster risks have recently caused numerous events, ranging from unprecedented category 5 hurricanes in the Atlantic Ocean to the COVID-19 pandemic. While regions around the world face urgent demands to prepare for, respond to, and to recover from such disasters, large-scale location data collected from mobile phone devices have opened up novel approaches to tackle these challenges. Mobile phone location data have enabled us to observe, estimate, and model human mobility dynamics at an unprecedented spatio-temporal granularity and scale. The COVID-19 pandemic has spurred the use of mobile phone location data for pandemic and disaster response. However, there is a lack of a comprehensive review that synthesizes the last decade of work leveraging mobile phone location data and case studies of natural hazards and epidemics. We address this gap by summarizing the existing work, and pointing promising areas and future challenges for using data to support disaster response and recovery.
Article
Evacuation planning is an essential part of transportation resilience. It is necessary to understand transportation systems and human behavior to plan for safe evacuation. This study estimates populations exposed to harm, clearance times and fatalities for short-notice evacuation for a catastrophic tsunami event. An Agent-Based Model (ABM) is developed and tested for Waikiki, Hawaii, with three travel modes: pedestrian, bicycle, and motor vehicles. The modeled environment includes the road network, tsunami inundation zones, and building footprints with vertical and horizontal evacuation destinations. Significant loss of life (38,760 persons) is estimated to occur with a catastrophic short-notice tsunami event. A new bridge over the Ala Wai Canal will create an additional exit route that could save 13,860 lives. Shortening tsunami detection and alert times by half would save 4,510 lives over the baseline, do-nothing scenario. Vertical evacuation is most effective and is estimated to save 55,770 lives. In addition to the importance to evacuation planning, the research using ABM supports longer term investment in transportation imnprovements, mitigation and adaptation.
Article
This paper investigates how perceived certainty factors influenced households’ selection of destination and accommodation type during evacuation. Using survey responses from Jacksonville, FL, multinomial logit models were developed for both choices. For the first, greater understanding of hurricane-related graphics decreased households' probability of staying within their community. Households with a member who has special medical needs and those evacuating with a greater number of vehicles were more likely to stay in the eastern portion of their county. Greater perceived certainty about the hurricane impact location decreased households’ probability of evacuating to the south. For the accommodation model, married evacuees and those who received official evacuation notices had increased likelihood of staying in hotels/motels, while those who evacuated a day before landfall were less likely to do so. Greater perceived certainty about hurricane impact time and frequency of communication with social network members increased the probability of staying in a peer’s home.
Article
Taiwan, located in the Pacific Ring of Fire, is highly vulnerable to the destructive impact of powerful earthquakes. When a massive earthquake occurs, failure to take proper measures can lead to massive casualties, collapse of public facilities, and disruption to traffic networks. Among the many post-disaster management strategies, efficient evacuation is absolutely critical as it can potentially reduce the number of casualties substantially. At present, most post-disaster evacuation decisions are made based on past experience, but the rarity of high-magnitude earthquakes impedes the rate at which the quality of the decision-making process can improve. As such, this study primarily aims to develop an efficient and credible evacuation simulation modeling framework which can be used as a powerful tool to bolster evacuation strategies. The simulation evacuation framework we propose is known as the Stochastic Pedestrian Cell Transmission Model (SPCTM). The model is built based on real data of the mapping of roads and emergency shelters in an urban environment, as well as predictive models of pedestrian flow, the number of people needed to be evacuated after an earthquake, and the impact of earthquakes on urban infrastructure. Much of the data is provided by the Taiwan Earthquake Impact Research and Information Application platform (TERIA) developed by the National Science and Technology Center for Disaster Reduction (NCDR). Applying this data, the model divides the traffic network into a cellular structure, applies the concepts of inflow and outflow to represent population movement and uses random variables to simulate the selection behavior of pedestrian agents. In addition to SPCTM, we propose an alternative model which makes slight modifications to SPCTM to offer users a complementary framework in which evacuees are armed with less ideal information about the evacuation situation. An empirical study is run in a district of Taipei, Taiwan based on a simulated earthquake of magnitude 6.5 in which the overall and moment-by-moment simulation evacuation process is presented. The use of SPCTM as a simulation evacuation model is compared to the current government evacuation policy in Taiwan. Further potential benefits of using the proposed simulation evacuation tool are numerous. For example, the frameworks proposed in this study can offer valuable insights to policy-makers on infrastructure design decisions and evacuation strategies before an earthquake occurs, including the best placement of relief centers, to potentially significantly reduce the number of casualties. Also, simulation results can be utilized after a disaster to assist in crucial decisions related to road restoration priorities, distribution of supplies to shelters, etc.
Article
Evacuation modelling has developed over time from simple engineering equations that do not consider behavioral tendencies to more sophisticated models with the potential to represent evacuation behaviors and decisions. This paper aims to lay the foundations for a more realistic representation of human factors in evacuation models, which is needed to ensure the adequacy of the infrastructure, decision processes and safety of evacuation. To provide a clearer picture of the empirical knowledge and modelling for evacuation studies, a generalized timeline is introduced. Recent behavioral evidence from empirical studies in the fields of both pedestrian evacuation and vehicular evacuations are reviewed to investigate the impact of various factors on the evacuee behavior over different phases. The consensus perspective on key behaviors that emerges is then used to review and consolidate the recent advances in evacuation modelling, in particular with respect to the formulations and techniques for representing these behaviors. Within each of these discussions, we pointed to current limitations and make corresponding suggestions on future research directions.
Article
This article reports on the calibration and analysis of a fully disaggregate (agent-based) transport simulation for the metropolitan area of Zurich. The agent-based simulation goes beyond traditional transport models in that it equilibrates not only route choice but all-day travel behavior, including departure time choice and mode choice. Previous work has shown that the application of a novel calibration technique that adjusts all choice dimensions at once from traffic counts yields cross-validation results that are competitive with any state-of-the-art four-step model. While the previous study aims at a methodological illustration of the calibration method, this work focuses on the real-world scenario, and it elaborates on the usefulness of the obtained results for further demand analysis purposes.
Article
The model has two levels of analysis, a macroscopic level and a microscopic level. The macroscopic level simulates the evacuation process on the highway network by looking at the major road arteries as a complete and integrated system. This macroscopic level provides an estimation of the maximum network evacuation time under different disaster intensity levels including severe traffic condition combinations. Furthermore, traffic bottlenecks are spotted and solutions can be introduced into the computer simulation model to study their effectiveness on clearance times. The microscopic level simulates a small highway network in detail that shows candidacy for traffic congestion such as highway network intersection controls, lane blockage due to accidents and others. This simulation level is used to test different traffic control and operational management strategies to improve evacuation process.
Article
Hazardous events, both natural and human-made, present tremendous risks to communities throughout the world. These events typically necessitate the evacuation of local or regional populations to safe destinations or shelters and have warning times ranging from minutes to hours or even days. The size and scope of these events present a challenge to the emergency management or agency personnel who must see to the health and safety of those living or working in their jurisdiction. This study evaluated various heuristic strategies to improve evacuation of an at-risk region by using a representative traffic roadway network. Finding evacuation strategies that reduce clearance time would lead to saving lives, time, and money. For the given test network, population density, or total number of trips, has an effect on overall clearance times; as densities (trips) increase, a greater potential for improved clearance time is indicated. Six different shift strategies were evaluated, each strategy based on origin-to-destination distances. For departure volumes greater than five vehicles per acre (approximately 12 vehicles per hectare), clearance times showed statistically significant improvements when departure times were shifted for groups within the network. In addition, the amount of the departure shift has an effect on clearance time.
Article
This thesis proposes a hybrid route guidance system in which predictive guidance is generated in a centralized layer and revised in a reactive, decentralized layer that resides on-board the vehicle. This hybrid approach is intended to improve guidance quality by balancing the ability of the centralized layer to generate consistent guidance with the ability of the decentralized layer to respond rapidly to incidents. Centralized guidance is computed using a rolling-horizon Dynamic Traffic Assignment routine based on the Method of Successive Averages. This guidance is disseminated to equipped vehicles in the form of prescribed paths, which may be subsequently revised by an on-board decentralized layer. This decentralized layer revises only the local portion of the vehicle's path in order to limit the potential negative impact of its myopic reactive algorithm. The layer uses a simple splitting algorithm in order to heuristically balance demand on alternate paths. Both layers utilize data collected by guided vehicles. The centralized layer uses position data from guided vehicles. The decentralized layer uses local arc travel time data which is shared among guided vehicles.
Article
A simulation-based framework for the modeling of transportation network performance under emergency conditions is presented. The system extends the well-established dynamic traffic assignment (DTA) framework and provides the necessary support for the meaningful study of a wide array of evacuation measures, the development of strategies under different prevailing conditions, and the generation of comprehensive emergency response plans. The system can be used to develop libraries to deal with emergencies and unplanned events, train response personnel and traffic management center operators, provide decision support and assistance for the evacuation of residents from affected areas, and ensure unhindered access to first responders. A variety of practical issues relevant to evacuation modeling are discussed, and the modeling framework is demonstrated by using the Boston, Massachusetts, network as an example. DynaMIT, a state-of-the-art DTA model, is used in the case study to illustrate how the benefits of network management strategies might be ascertained. The paper concludes with future directions, including the integration of simulation modeling as a real-time tool for the management of evolving evacuations.