optimal design of rapid evacuation ... - optimization onlineconstrained urban networks. we propose...
TRANSCRIPT
TRANSPORTMETRICA A
Optimal Design of Rapid Evacuation Strategies in
Constrained Urban Transport Networks
Corresponding Author:
Jose Javier Escribano-Macias
PhD Student in Sustainable Civil Engineering
Department of Civil and Environmental Engineering
Imperial College London
SW7 2BU, UK
Panagiotis Angeloudis
Senior Lecturer in Transport Systems and Logistics
Centre for Transport Studies
Department of Civil and Environmental Engineering
Imperial College London
SW7 2BU, UK
Ke Han
Lecturer in Transport Operations and Logistics
Centre for Transport Studies
Department of Civil and Environmental Engineering
Imperial College London
SW7 2BU, UK
Word Count: 11,541
Optimal Design of Rapid Evacuation Strategies in Constrained Urban Transportation Networks
Abstract
Large-scale evacuations constitute common life-saving exercises that are activated in many
disaster response campaigns. Their effectiveness is often inhibited by traffic congestion,
disrupted and imperfect coordination mechanisms, and the poor state of the underlying
transportation networks. To address this problem, this paper presents a hybrid simulation-
optimisation methodology to optimise evacuation response strategies through demand staging
and signal phasing. We introduce a pre-planning model that evaluates evacuation policies,
using a low-level dynamic traffic assignment model that captures the effects of congestion,
queuing and vehicle spillback. Optimal strategies are determined using derivative-free
optimisation algorithms, applied to an evacuation problem based on a benchmark dataset. The
effects of varying the number of activated paths and the frequency of departure under different
network conditions are observed. Our analysis indicates that combined departure time
scheduling and signal phasing is a promising method to improve evacuation efficiency when
compared to a worst-case benchmark scenario.
Keywords: simulation-optimisation, disaster evacuation, genetic algorithms, dynamic traffic
assignment
1. Introduction
Evacuations are disaster response mechanisms aiming to temporarily relocate populations
before, during and after large-scale disasters while seeking to minimise potential loss of life.
Given the inherent managerial, societal, financial, behavioural, and operational aspects of
evacuations, adequate planning is required to ensure their effectiveness and safety. Despite the
challenge of predicting the number of lives saved through a specific evacuation strategy given
the interrelated factors at play, evidence indicates that countries with extensive disaster
preparedness mechanisms have been able to reduce casualties substantially from natural
disasters (Hallegatte, 2012).
Key among the factors impeding evacuation processes is the unique traffic patterns that
arise due to the sudden surge in travel demand, which contravene nominal traveller behaviours.
Even when extensive warning is given, the mass mobilisation imposes excessive strains on the
transport network due to the small time window, the lack of resources and the number of
evacuees (Thomas and Kopczak, 2007). As a result, network congestion is among the most
significant challenges encountered during emergency evacuations and is aggravated by
spillbacks, shockwaves and gridlocks.
The effects of congestion were illustrated during the Tohoku earthquake and subsequent
tsunami in 2011, where 51% of casualties were evacuating, of which 30% had problems
reaching shelters due to network congestion (Yun and Hamada, 2015). An overview of recent
studies on high profile natural disasters, their impacts and problems encountered during
evacuations is provided in Table 1.
As is the case with ordinary traffic, many evacuation strategies rely on capacity
expansion measures to preclude the occurrence of congestion. Among them, contraflow and
crossing elimination have been found to result in capacity increases of up to 70% and 40%,
respectively (Wolshon, 2001; Cova and Johnson, 2003). However, such techniques rely on
prior familiarity with evacuation plans and are likely to increase safety risks in scenarios where
evacuees are unaware or non-compliant (Xie, Lin and Travis Waller, 2010).
Table 1 List of recent natural disasters and evacuation problems.
Disaster Area Date Deaths Evacuation Problems Reference
T PH 2013 6,300 800,000 Damage to evacuation sites. Mahar et al. (2015)
F IN 2013 4,000 24,000 Evacuation order not issued. Rautela (2013)
Rana et al. (2013)
EQ ID 2012 10 3 million Traffic congestion. Munawir (2012)
News24 (2012)
EQ-TS JP 2011 16,000 400,000 Damage to transport
infrastructure. Hasegawa (2013)
EQ HT 2010 230,000 482,000 Lack of warning, damage to
transport infrastructure.
Margesson & Taft-
Morales (2010)
H US 2005 1,800 400,000 Over 100,000 to 300,000 did
not or could not evacuate. Wolshon (2006)
EQ-TS SA 2004 220,000 N/A Lack of warning, damage to
transport infrastructure. Fehr et al. (2005)
Table Legend
Disaster Type: T โ Typhoon, F โ Floods, EQ โ Earthquake, TS โ Tsunami, H โ Hurricane.
Area Affected: PH โ Philippines, IN โ India, ID โ Indonesia, JP โ Japan, HT โ Haiti,
US โ United States of America, SA โ Southern Asia.
Demand management has been explored as an alternative method to limit congestion effects
by restructuring the distribution of evacuees across the network (Abdelgawad and Abdulhai,
2009). This is achieved by providing alternate safe sites to evacuees or by staging departures
at their source (Chen and Zhan, 2008).
The estimation of expected network congestion levels during an evacuation is not a trivial
task. The occurrence and scale of gridlocks, spillbacks and shockwaves are dependent upon
time and user behaviour, and as such cannot be integrated directly into linear optimisation
algorithms without affecting model accuracy or computational performance (Lu, George and
Shekhar, 2005). As a result, many previous studies on congestion prediction have adopted bi-
level optimisation models that involve Stackelberg games.
In such a model, the upper-layer usually focuses on decisions taken by a โsuper-userโ,
who assumes the role of the decision-maker and possesses perfect knowledge of the evacuation
process. In practice, this role is assumed by emergency management and law enforcement
organisations, with support from transport agencies (Urbina and Wolshon, 2003). Conversely,
the lower-level focuses on modelling the effects of these decisions (Urbina and Wolshon, 2003;
Abdelgawad and Abdulhai, 2009). Typical implementations use traffic assignment models,
which may vary in scale and detail (Table 2).
Table 2 Traffic assignment model types
Model Type Description Examples
Microscopic
Considers individual behaviours and
congestion effects. Computationally
expensive.
OREMS (Li, Yang and Heng, 2006) and
MatSim (Nagel and Flรถtterรถd, 2012)
Mesoscopic Models groups of agents. Compromise
between modelling detail and speed.
DYNASMART-P (Sbayti and
Mahmassani, 2006), DTALite (Zhou and
Taylor, 2014), and models by Szeto &
Lo (2004) and Ziliaskopoulos (2000)
Macroscopic
Vehicle movement modelled as fluid
flows. Less emphasis on individual
interactions.
DYNEV (Goldblatt, 2004) and models
by Friesz et al. (1993) and Han et al.
(2013)
Hobeika & Kim (1998) were among the first to use demand staging for evacuation
modelling of UE and SP assignments in nuclear power plant evacuations. However, their study
did not consider SO assignment, limiting the effectiveness of any evacuation plan.
Sbayti & Mahmassani (2006) developed a bi-level optimisation algorithm where the
upper-level generates all evacuee origin-destination choices and initial time-dependent historic
assignment paths to the lower level. The latter uses a mesoscopic cell transmission model
(CTM) featuring dynamic SO traffic assignment.
Another SO CTM-based solution is presented by Chiu et al. (2007) for no-notice
evacuations. Their formulation integrated departure schedule and traffic assignment decisions
into the single-destination bi-level evacuation optimisation model.
An optimal-control formulation is proposed by Liu et al. (2006), discretising the
evacuation area into staged departure zones, instead of the node-based departure staging
strategies described above. A similar problem is presented by Bish et al. (2014), who formulate
a CTM-based mixed integer linear problem for the reduction network of clearance times with
multiple evacuee types.
Mitchell & Radwan (2006) use a microscopic traffic assignment method with pre-trip
and en-route choice models to optimise network clearance time, highlighting the scalability
limitations derivative of microscopic assignment methods. Although considering single-route
evacuation, Chien & Korikanthimath (2007) presented a model that consider the total duration
and any congestion delays of multistage evacuations.
Chen & Zhan (2008) used agent-based simulation to evaluate the merits of evacuation
staging strategies under different network conditions, finding best results for geometric
transportation networks in high-density areas. Abdelgawad & Abdulhai (2010) proposed a
spatiotemporal multi-modal evacuation framework that minimises travel and wait times for
evacuees.
More recently, Bretschneider & Kimms (2012) presented a pattern-based dynamic flow
model solved using a bi-level optimisation heuristic to minimise evacuation clearance time. Li
et al. (2012) formulated a shortest-path evacuation staging algorithm that minimises evacuation
congestion. Extensions developed by Xie et al. (2014) and Li et al. (2018) considered multiple
destinations and incorporated a time-expanded network flow model respectively.
Kimms & Maiwald (2017) developed a CTM-based evacuation planning model
considering contraflow and staged evacuation. A further study by Kimms & Maiwald (2018)
investigated route resilience by minimising potential capacity using a CTM evacuation model.
An alternative formulation developed by Even et al. (2015) used a tree-design framework
to minimise converging and diverging delays and improve evacuation performance. A similar
framework is proposed by Pillac et al. (2016), where congestion and conflicts are minimised
through demand staging over pre-generated routes.
Another evacuation path generation algorithm is developed by Seekircher (2017), who
considers the selfish behaviour of the evacuees and manipulate evacuee behaviour through
strategic link blockages. The solution procedure consists of a bi-level approach where the
upper-level optimises the route choice and the lower-level manipulates the network structure.
While the optimisation of isolated components (demand staging, signalling or user
behaviour) during evacuation have been explored independently, no study in the reviewed
literature considers simultaneous mechanisms and their interactions, except Kimms & Maiwald
(2017) who explore contraflow assignment and demand staging. Additionally, the choice of
evacuation routes and origin-destination pairs have been commonly addressed as part of the
optimisation model, instead of being considered within the context of road user choice (Liu,
Lai and Chang, 2006; Sbayti and Mahmassani, 2006; Jabari, He and Liu, 2012).
This study focuses on the development of an evacuation planning algorithm in
constrained urban networks. We propose an assignment-optimisation framework composed of
two elements that contribute to the development of an evacuation strategy, namely an
evacuation decision model and a dynamic traffic model. The latter uses an efficient numerical
traffic model based on continuous time-dependent kinematic wave simulation, with
consideration of congestion effects and the impact of user information on road traffic. The
algorithm is structured as a Stackelberg game acting as a simulation-optimisation framework,
pursuing optimal evacuation strategies in the absence of derivative information (Nguyen,
Reiter and Rigo, 2014; Amaran et al., 2016).
Our main contribution to the state of the art comprises the development of an evacuation
decision model that simultaneously controls the evacuation and signal phasing strategy over a
specified time horizon. In addition, contrasting the literature, the model acknowledges the
evacueesโ ability to make independent route choices, only constrained by a pre-defined set of
available paths. In doing so, the study seeks to highlight the benefits of evacuation staging in
constrained road networks under varying conditions, including multiple sink nodes and
extensive link damage.
The mathematical formulation used to determine optimal evacuation strategies is
provided in Section 2, followed by the traffic and evacuation routing models presented in
Section 3. The implementation of the overall algorithm is described in Section 4, alongside any
modifications made to the underlying models to facilitate closer integration. A numerical
example is presented, accompanied by a discussion on performance and scalability.
Conclusions and recommendations for future work are provided in Section 5.
2. Evacuation Strategy Model
The upper-level model in our framework is used to capture the decisions involved in the design
of the evacuation strategy, with the objective to maximise evacuee numbers within a predefined
time span. In this section, we assume that the evacuation strategy adopts a departure staging
mechanism, with users prescribed a set of available routes and instructed when to vacate their
premises. Once users are allowed to depart, they are free to travel towards any of the evacuation
shelters (implemented as sink nodes) using any of the routes that have been made available. A
limitation on the total number of distinct routes selected at every time step is considered. The
following assumptions are adopted:
Assumption 1: Evacuees use their road vehicle as a mode of transportation.
We adopt this assumption to transform the model into a single-commodity assignment problem
where evacuees share the same attributes regarding driving behaviour (Kimms and Maiwald,
2018). We acknowledge that public transport can aid evacuation performance, especially
among specific groups such as limited mobility evacuees (Urbina and Wolshon, 2003) and that
limiting car usage may be beneficial to the evacuation performance (Shimamoto et al., 2018).
Assumption 2: Background traffic and shadow evacuation effects are ignored.
This assumption follows from literature (Tuydes-Yaman and Ziliaskopoulos, 2014; Xie et al.,
2014) and seeks to ensure deterministic results given its intended use as a strategic design tool.
However, we recognise that background traffic is likely to affect the performance of the
evacuation strategy alongside the availability of information systems and the nature of hazard
type (Zheng et al., 2010).
In practice, we expect that evacuation planners would apply the framework in a range
of problem instances, each pertaining to a different network state. Concerning shadow
evacuation or self-evacuation, it can account for 20% of evacuees for uncommon hazards, such
as chemical or nuclear threats (Zeigler et al., 1981; Mitchell, Cutter and Edmonds, 2007). We
assume that the citizens in our model are familiar with existing evacuation procedures.
Assumption 3: Intermediate trips are not considered.
Intermediate trips made to an alternate destination before evacuating increase network
clearance time, and are more common under short warning times (Liu and Murray-Tuite, 2011;
Murray-Tuite and Wolshon, 2013). An illustrative example is provided in the 2012 Indonesia
Earthquake, where the evacuation was undermined by families collecting their children from
the education centres (Munawir, 2012; News24, 2012). We assume that sufficient warning time
is given, minimising the impact of this phenomenon.
Assumption 4: The super-user has perfect information of the traffic network.
The assumption is based on the existence of Intelligent Transport Systems (ITS) that can update
information on the transport network during the evacuation. Therefore, authorities can collect
data on the network state in real-time and provide evacuees with instructions that reflect that
current state of the network. Assuming that sufficient warning is given, the ITS systems are
considered to be in operation during the evacuation. Additionally, no uncertainty related to
evacuee demand is considered, and thus a deterministic model is proposed (Sabouhi et al.,
2018).
Assumption 5: Users cannot change their evacuation route after departure.
During the evacuation, road user compliance with controllers is difficult to estimate and model,
as these depend on social and economic factors, among others (Sorensen, 2000). We partially
address the effects of this assumption by allowing users to select through a range of possible
routes during the traffic assignment stage of our algorithm.
The following notation is used in the remainder of this section:
Indices: Sets:
๐, ๐ links ๐๐ฃ incoming links to node ๐ฃ
๐ paths ๐ช๐ฃ outwards links from node ๐ฃ
๐ก evacuation time step ๐ซ paths
๐ destination node ๐ time horizon
๐ source node ๐ destination nodes
๐ฃ network node ๐ฎ source nodes
๐ links in the network
๐ฑ nodes in the network
Parameters:
๐พ total number of paths [โ] ๐ upper departure rate limit [๐ฃ๐โ/๐ ]
๐พ fitness factor [โ] ๐ธ lower departure rate limit [๐ฃ๐โ/๐ ]
๐๐,๐ path-source node relation [โ] ๐๐,๐ path-destination node relation [โ]
๐๐ destination node capacity [๐ฃ๐โ] ๐ต๐ source node total demand [๐ฃ๐โ/๐ ]
ฮ๐ก๐๐ก estimated travel time of path ๐ at
time ๐ก [๐ ]
Decision variables:
๐ ๐๐ก departure rate [๐ฃ๐โ/๐ ] ๐๐
๐ก activated paths [โ]
๐๐ข๐,๐๐ก cumulative vehicles entering link ๐
[๐ฃ๐โ]
๐๐๐,๐๐ก vehicle flow entering link ๐ [๐ฃ๐โ/๐ ]
๐๐๐๐ค๐,๐๐ก cumulative vehicles exiting link ๐
[๐ฃ๐โ]
๐๐๐ข๐ก,๐๐ก vehicle flow exiting link ๐ [๐ฃ๐โ/๐ ]
2.1.Integer Formulation
The overall objective of this model is to ensure that the maximum number of evacuees from a
given area complete their journeys within the course of a finite time horizon. The level of risk
which the evacuees are subjected to is assumed to increase with time. Therefore, we seek to
ensure that the journeys are completed as early as possible, using elapsed time as a surrogate
for risk level. We adopt a composite function (1) consisting of two factors that reflect these
objectives (1.1-1.2):
๐ = ๐ถ โ ๐พ๐ (1)
๐ถ = โ (1
๐กโ ๐๐ข๐,๐
๐ก
๐โ๐๐
)
๐กโ๐,๐โ๐
(1.1)
๐ =โ( โ ๐๐๐๐ค๐,๐๐ก
๐ โ๐ฎ,๐โ๐ช๐
โ โ ๐๐ข๐,๐๐ก
๐โ๐,๐โ๐๐
)
๐กโ๐
(1.2)
Equation (1.1) penalises departures given the time step at which they are due to
complete their journeys. The ๐๐ข๐,๐๐ก variable is used to capture the cumulative number of vehicle
departures at every time step ๐ก. The second component of the objective (1.2) seeks to minimise
the total wait time ๐, which captures the total duration of aggregate vehicle presence in the
network (which includes delays). A weight factor ๐พ is used to control the priority allocated to
each of the two objectives. A maximum time horizon ๐ is defined along with discretised time
step ฮ๐ก. The model formulation is as follows:
Objective:
๐๐๐ฅ๐๐๐๐ ๐ ๐ (2)
Subject to:
๐๐ข๐,๐๐ก = ฮ๐กโ๐๐๐,๐
๐
๐ก
๐=1
๐ โ ๐,
๐ก โ ๐ (2.1)
๐๐๐๐ค๐,๐๐ก = ฮ๐กโ๐๐๐ข๐ก,๐
๐
๐ก
๐=1
๐ โ ๐,
๐ก โ ๐ (2.2)
โ ๐๐,๐๐๐ข๐,๐๐ก
๐โ๐ซ.๐โ๐,๐โ๐๐
= โ ๐๐,๐ ๐๐๐ข๐ก,๐๐ ฮ๐ก
๐โ๐ซ,๐ โ๐ฎ,๐โ๐ช๐ ,๐โ๐:๐+ฮ๐ก๐๐ โค๐ก
๐ก โ ๐ (2.3)
๐๐๐ข๐ก,๐๐ก = โ๐๐
๐ก ๐ ๐๐ก๐๐,๐
๐โ๐ซ
๐ โ ๐ช๐ ,
๐ก โ ๐ (2.4)
๐ธ โค ๐ ๐๐ก โค ๐
๐ โ ๐ซ,
๐ก โ ๐ (2.5)
โ ๐๐,๐ ๐ ๐๐ก๐๐
๐ก
๐กโ๐,๐โ๐ซ
= ๐ต๐ ๐ โ ๐ฎ (2.6)
โ ๐๐๐๐ค๐,๐๐ก
๐โ๐๐
โค ๐๐ ๐ก โ ๐
๐ โ ๐ (2.7)
๐๐๐ก = {0,1}
๐ โ ๐ซ,
๐ก โ ๐ (2.8)
โ๐๐๐ก
๐โ๐ซ
= ๐พ ๐ก โ ๐ (2.9)
Constraints (2.1) and (2.2) calculate the number of vehicles in every link ๐ among the
arcs ๐ of the network. Constraint (2.3) ensures that vehicles entering path ๐ will arrive at the
sink node within travel time ฮ๐ก๐๐ก . Constraint (2.4) calculates the vehicular flow entering the
network at time step ๐ก given a vehicle inflow rate ๐ ๐๐ก at the activated paths ๐๐
๐ก .
Equation (2.5) ensures that the path rate is bounded by a lower limit ๐ and an upper limit
๐ at every time step ๐ก. Constraint (2.6) equates the total departures throughout the time horizon
๐ to the original demand. Equation (2.7) ensure capacity limitations at the destination nodes
are satisfied. Finally, (2.8) and (2.9) state that ๐๐๐ก is a Boolean variable and that the paths
activated must be limited to ๐พ for each origin, respectively.
The main limitations of the above formulation are its inability to consider signal control
and to evaluate ฮ๐ก๐๐ก , which is varies with vehicle flows due to congestion. Therefore, the value
of ฮ๐ก๐๐ก is treated as a dynamic variable with a non-linear relationship with the decision variables
๐ ๐๐ก and ๐๐
๐ก . Section 3 introduces a dynamic traffic assignment model that evaluates ฮ๐ก๐๐ก for each
time step and models nodes as signalled junctions.
2.2. Linear Model Formulation
A typical linearisation approach for the traffic assignment component of the model consists on
duplicating the network for every discrete time step of a time horizon ๐, increasing the problem
size to impractical levels for large transport networks (Lu, George and Shekhar, 2005; Shen,
Nie and Zhang, 2007; Li, Zhou and Ju, 2010). Mathematical simplifications including no-
overtaking (First-In-First-Out) and traffic holding require the inclusion of additional variables
and non-convex constraints that preclude the adoption of computationally efficient solution
algorithms (Doan and Ukkusuri, 2012).
Given these limitations, we develop a linear mathematical formulation that is solved in
tandem with the dynamic traffic simulator. The evacuation strategy developed using CPLEX
serves as an input for the traffic simulator, which updates path travel times ฮ๐ก๐๐ก considering the
effects of congestion. The linear problem is then solved using the updated network properties,
creating a deviation between the solutions before and after ฮ๐ก๐๐ก is updated. This iterative
process runs until the variation between solutions is negligible, thus estimating the effects of
congestion within the constraints of the linear formulation.
The main limitation of the proposed approach is the use of a static ฮ๐ก๐๐ก throughout the
time window for each iteration. We present the linear formulation, introducing the following
parameters in addition to the notation description in Section 2:
Decision Variables:
๐๐๐ก slack variable to linearise relationship between ๐๐,๐ก and ๐๐ท๐,๐ก [๐ฃ๐โ/๐ ]
๐๐ท๐ก๐ vehicle departures from path ๐ at time-step ๐ก [โ]
๐๐ด๐ก๐ vehicle arrival from path ๐ at time-step ๐ก [โ]
Objective:
๐๐๐ฅ๐๐๐๐ ๐ ๐ = ๐ถ โ ๐พ๐ (3)
๐ถ =โ(1
๐กโ ๐๐ด
๐ก๐
๐ โ ๐
)
๐กโ๐
(3.1)
๐ =โ(โ๐๐ท๐ก๐
๐โ๐
โโ๐๐ด๐ก๐
๐โ๐
)
๐กโ๐
(3.2)
Subject to:
๐๐ด๐ก๐= โ ๐๐ท๐
๐
๐โ๐:๐+ฮ๐ก๐๐ =๐ก
๐ โ ๐ซ,
๐ก โ ๐ (4.1)
๐๐ด๐ก๐=โ๐๐ด๐
๐
๐ก
๐=1
๐ โ ๐ซ,
๐ก โ ๐ (4.2)
๐๐ท๐ก๐=โ๐๐ท๐
๐
๐ก
๐=1
๐ โ ๐ซ,
๐ก โ ๐ (4.3)
๐ ๐๐กฮ๐ก๐ = ๐๐ท
๐ก๐
๐ โ ๐ซ,
๐ก โ ๐ (4.4)
๐๐๐ก๐ธ โค ๐ ๐
๐ก โค ๐๐๐ก ๐
๐ โ ๐ซ,
๐ก โ ๐ (4.5)
๐๐๐ก โ ๐ธ(1 โ ๐๐
๐ก ) โค ๐ ๐๐ก โค ๐๐
๐ก โ ๐(1 โ ๐๐๐ก )
๐ โ ๐ซ,
๐ก โ ๐ (4.6)
๐ต๐ = โ ๐๐,๐ ๐ ๐๐ก
๐กโ๐,๐โ๐
๐ โ ๐ฎ (4.7)
๐๐ โฅ โ๐๐ท๐ก๐ ๐๐,๐
๐โ๐
๐ก โ ๐
๐ โ ๐ (4.8)
โ๐๐๐ก
๐โ๐
๐๐,๐ = ๐พ ๐ โ ๐ฎ,
๐ก โ ๐ (4.9)
๐๐๐ก = {0,1}
๐ โ ๐ซ,
๐ก โ ๐ (4.10)
๐ธ โค ๐๐๐ก โค ๐
๐ โ ๐ซ,
๐ก โ ๐ (4.11)
The objective function (3) is similar to equation (1), as it calculates the cumulative
weighted evacuations and the total wait time. Equation (4.1) ensures that vehicle departures
and arrivals are equal within a path ๐ given an expected travel time ฮ๐ก๐๐ก . Constraints (4.2-4.3)
calculates the cumulative vehicle departures ๐๐ท๐ก๐ and arrivals ๐๐ด
๐ก๐. Equation (4.4) considers
the departing flows ๐ ๐๐ก and the total vehicles entering a path ๐๐ท๐
๐ก . Constraints (4.5) and (4.6)
bounds the departing vehicles ๐ ๐๐ก to ๐ and ๐ respectively while ๐ ๐
๐ก = 0 for non-selected paths.
Equation (4.7) equates the total demand in each source node ๐ to the total departures of
each path, where ๐๐,๐ equates to 1 when source ๐ belongs to path ๐ and 0 otherwise. Equation
(4.8) considers sink node capacity similarly to (2.7). Constraint (4.9) ensures that the total
number of paths activated ๐๐๐ก for each source node equals ๐พ. Finally, equations (4.10) and
(4.11) bound the decision variables ๐๐๐ก and ๐๐
๐ก to a Boolean and to ๐ and ๐ respectively.
3. Dynamic Network Loading Model
This section presents the model that addresses the inability of the above formulation to capture
the effects of congestion on vehicular flow by using a continuous time link-based kinematic
wave model that was initially developed by Han et al. (2016). The kinematic wave model
considers the effects of flow propagation and congestion, including physical queuing and
vehicle spillback, based on a reformulation of the scalar conservation law model (Lighthill and
Whitham, 1955; Richards, 1956), the Hamilton-Jacobi equation and variational theory. Such a
formulation has been independently derived by Jin (2015) and Han et al. (2016) and, when
discretised, leads to the well-known link transmission model (Yperman, Logghe and Immers,
2005). The model is path-based, with path selection and network properties provided
previously by the user.
The traffic flow model consists of three components: a junction model, a source model
and a link model. The junction model can be used to implement traffic signalling in the
network. The source model provides departure rates from specific nodes over the time-horizon
of the evacuation. Finally, the link model is based on the fundamental triangular diagram and
the concepts of demand and supply (Lebacque and Khoshyaran, 1999), allowing the simulation
to calculate in-flows and out-flows of links that are adjacent to the same junction and thus
simulate shockwave propagation and vehicle spillback.
The traffic simulator is described by a system of time-continuous differential algebraic
equations (DAE). It is also referred to as the dynamic network loading procedure in the
dynamic traffic assignment literature. The underlying mathematical model is presented below
following the notations introduced in Section 2 - the reader is referred to Han et al. (2016) for
its derivation and computational examples.
Parameters
๐ถ๐ flow capacity of link ๐ [๐ฃ๐โ] ๐๐๐๐๐
jam density of link ๐ [๐ฃ๐โ/๐]
๐ฟ๐ length of link ๐ [๐] ๐ฃ๐ forward wave speed of link ๐ [๐/๐ ]
๐ค๐ backward wave speed of link ๐ [๐/๐ ]
Variables
๐ท๐๐ก demand of link ๐ [๐ฃ๐โ] ๐๐
๐ก supply of link ๐ [๐ฃ๐โ]
๐๐๐,๐๐ก inflow of link ๐ [๐ฃ๐โ/๐ ] ๐๐๐ข๐ก,๐
๐ก outflow of link ๐ [๐ฃ๐โ/๐ ]
๐๐ข๐,๐๐ก cumulative link entry volume [๐ฃ๐โ] ๐๐๐๐ค๐,๐
๐ก cumulative link exit volume [๐ฃ๐โ]
๐๐๐,๐ก(๐ฅ) path disaggregation variable [โ] ๐๐
๐ก link entry time function [๐ ]
๐ผ๐๐๐ฝ,๐ก
flow split at junction ๐ฝ [โ] ๐๐๐ก effective green time for link ๐ [โ]
๐ด๐ฝ flow distribution at ๐ฝ [โ]
The definitions of demand and supply that are used in this model conform with those
proposed by Lebacque & Khoshyaran (1999), and correspond to the inflow and outflow
capacity of a link. ๐๐ข๐,๐๐ก and ๐๐๐๐ค๐,๐
๐ก denote the cumulative number of vehicles that have entered
and exited the link ๐ by time ๐ก, respectively. In order to explicitly incorporate traveller route
choices, the path disaggregate variable ๐๐๐,๐ก(๐ฅ) represents the proportion of traffic at point ๐ฅ
and time step ๐ก on link ๐ that is associated with path ๐, where ๐ โ ๐. Based on ๐๐ข๐,๐๐ก and ๐๐๐๐ค๐,๐
๐ก ,
the link entry time function can be calculated based on the following flow propagation identity:
๐๐ข๐,๐๐๐๐ก
= ๐๐๐๐ค๐,๐๐ก ,
where ๐ก denotes the link exit time. Finally, the flow split variable ๐ผ๐๐๐ฝ,๐ก
indicates the probability
of traffic exiting upstream link ๐ that are heading towards downstream link ๐, where both ๐ and
๐ are adjacent to junction ๐ฝ. Given the above, and the formulation provided in the previous
section, the DAE system for the dynamic network loading model is as follows:
๐ท๐๐ก =
{
๐๐๐,๐
๐กโ๐ฟ๐๐ฃ๐ if ๐
๐ข๐,๐
๐กโ๐ฟ๐๐ฃ๐ = ๐๐๐๐ค๐,๐
๐ก
๐ถ๐ if ๐๐ข๐,๐๐กโ๐ฟ๐๐ฃ๐ > ๐๐๐๐ค๐,๐
๐ก
๐ โ ๐,
๐ก โ ๐ (5.1)
๐๐๐ก =
{
๐๐๐ข๐ก,๐
๐กโ๐ฟ๐๐ค๐ if ๐๐ข๐,๐
๐ก = ๐๐๐๐ค๐,๐
๐กโ๐ฟ๐๐ค๐ + ๐๐
๐๐๐๐ฟ๐
๐ถ๐ if ๐๐ข๐,๐๐ก < ๐
๐๐๐ค๐,๐
๐กโ๐ฟ๐๐ค๐ + ๐๐
๐๐๐๐ฟ๐
๐ โ ๐,
๐ก โ ๐ (5.2)
๐๐๐๐ค๐,๐๐ก = ๐๐ข๐,๐
๐๐๐ก
๐ โ ๐,
๐ก โ ๐ (5.3)
๐๐๐,๐ก(0) =
๐๐๐ข๐ก,๐๐ก ๐
๐
๐,๐๐๐ก
(0)
๐๐๐,๐๐ก โ๐ such that {๐, ๐} โ ๐
๐ก โ ๐ (5.4)
๐ด๐ฝ(๐ก) = {๐ผ๐๐๐ฝ,๐ก}, ๐ผ๐๐
๐ฝ,๐ก = โ ๐๐๐,๐๐
๐ก
(0)
๐,๐โ๐
๐ โ ๐๐ฝ, ๐ โ ๐ช๐ฝ,
๐ก โ ๐ (5.5)
๐๐๐,๐๐ก = ๐ ๐๐
๐ด๐ฝ [(๐ท๐๐ก)๐โ๐J , (๐๐
๐ก)๐โ๐ชJ
, (๐๐๐ก)๐โ๐J]
๐ โ ๐ช๐ฝ, ๐ก โ ๐ (5.6)
๐๐๐ข๐ก,๐๐ก = ๐ ๐๐
๐ด๐ฝ [(๐ท๐๐ก)๐โ๐J , (๐๐
๐ก)๐โ๐ชJ
, (๐๐๐ก)๐โ๐J]
๐ โ ๐๐ฝ, ๐ก โ ๐ (5.7)
๐๐ข๐,๐๐ก โ ๐๐ข๐.๐
๐กโ1 =ฮ๐ก ๐๐๐,๐๐ก , ๐๐๐๐ค๐,๐
๐ก โ ๐๐๐๐ค๐,๐๐กโ1 = ฮ๐ก ๐๐๐ข๐ก,๐
๐ก ๐ โ ๐, ๐ก โ ๐ (5.8)
Equations (5.1)-(5.2) define the link demand and supply, given the cumulative entry and exit
volumes at earlier times. The relationship between entry and exit times at every link, based on
flow conservation and the First-In-First-Out principle is defined by equation (5.3). Equation
(5.4) propagates the path disaggregation variable between consecutive links, while equation
(5.5) states that the flow distribution at a junction with multiple downstream links can be
uniquely determined by the relevant path disaggregation variables.
Equations (5.6)-(5.7) determine the inflow and outflow of links that are incident to a
junction ๐ฝ based on the Riemann Solver proposed by (Han et al., 2014; Han and Gayah, 2015),
which approximates the effect of on-off signal controls using a continuum approach based on
the kinematic wave theory, and depends on the junction topology and specific traffic
management scenarios such as signal control and priorities.
Finally, (5.8) links the two types of fundamental flow variables: flow and cumulative
volume. This DAE system can be solved in discrete time in a forward fashion; see Garavello
et al. (2016) for computational examples of this system in a number of traffic networks.
The traffic model uses a set of paths as primary inputs. While safety concerns in real-live
evacuation exercises may condition routing, we assume for this study that a reduction of vehicle
exposure in the network will increase user safety. This simplification allows us to use a k-
shortest path algorithm based on Yenโs algorithm to generate a set of multiple route-choices to
the exit node (Yen, 1971). Our implementation of the evacuation routing algorithm is detailed
in the Pseudocode 1 below and incorporates a Dijkstraโs-based solver to determine individual
paths (Dijkstra, 1959).
1. Algorithm Evacuation Routing is
2. Input: graph ๐บ, origins ๐๐ฃ, destinations ๐ท๐ฃ, number of paths ๐พ 3. Output: shortest paths ๐ฟ๐ 4.
5. spurPath ๐๐ โ CreateEmptyList() 6. ๐โ โ Dijkstra(๐บ, ๐๐ฃ , ๐ท๐ฃ) 7. for each path ๐ in ๐ฟ๐:
8. for each vertex ๐ฃ in ๐
9. spur vertex ๐ฃ๐ โ ๐ฃ
10. root path ๐๐ โ ๐[0, ๐ฃ] 11. for each subPath ๐๐ in ๐ฟ๐
12. if ๐๐ == ๐๐ [0, ๐ฃ]
13. ๐บ. RemoveLink(๐๐ [๐ฃ, ๐ฃ + 1] 14. for each pathNode ๐ฃ๐ in ๐๐ except ๐ฃ๐
15. ๐บ. RemoveVertex(๐๐[๐ฃ๐])
16. ๐๐ โ Dijkstra(๐บ, ๐๐ฃ , ๐ท๐ฃ) 17. newPath ๐๐ = ๐๐ + ๐๐ 18. spur path list ๐ฟ๐๐ โ ๐๐
19. ๐บ. Restore() 20. if ๐ฟ๐๐ is empty 21. break
22. ๐ฟ๐๐ . SortByCost()
23. ๐ฟ๐ โ ๐ฟ๐๐ . First()
24. ๐ฟ๐๐ . RemoveFirst()
25. return ๐ฟ๐ Pseudocode 1 Shortest path algorithm pseudocode.
We integrate both models using a feedback loop as presented in Pseudocode 2. The traffic
model requests the initial paths and network conditions, a simulation time-step ฮ๐ก๐ and decision
time-step ฮ๐ก๐ as inputs. Between each ฮ๐ก๐ the model updates the network properties based on
current vehicular flow, maintaining network properties constant during ฮ๐ก๐ .
The traffic model outputs the current network conditions and calls the evacuation routing
model after each ฮ๐ก๐, returning the new shortest paths accounting for congestion. A list of
paths is updated in case the returned shortest paths have not been previously explored, allowing
the algorithm to respond to major changes in the state of network while minimising reductions
to the performance of the algorithm. The cost of all paths in the list is updated as the algorithm
runs for traceability purposes.
The proposed traffic assignment model has been applied to a range of network sizes and
configurations to determine its scalability. Networks A, B, C and D are defined in the
Appendix, while the Sioux-Falls and Enriched Sioux Falls networks are presented in further
detail in Section 4.2. Different junction configurations are tested to determine the effect in
computational time. As shown in Table 3, the traffic assignment model can operate in complex
networks under reasonable runtime. The computation time increases with network size and
network complexity.
1. Algorithm Traffic Model is
2. Input: network properties ๐, simulation time step ฮ๐ก๐ , decision time
step ฮ๐ก๐ initial flows ๐, initial paths ๐พ, time horizon ๐, graph ๐บ,
origins ๐, destinations ๐ท, activated paths ๐ 3.
4. Output: network flows ๐โ 5.
6. ๐ก = 0 7. while ๐ก < ๐ 8. ๐ = EvaluateFlows(๐, ๐, ๐พ) 9. if IsSimulationTimeStep(๐ก, ฮ๐ก๐ ) 10. ๐ = UpdateNetworkProperties(๐, ๐) 11. if IsDecisionTimeStep(๐ก, ฮ๐ก๐) 12. ๐พ = YensAlgorithm(๐บ, ๐, ๐ท, ๐) 13. if AllVehiclesLeftNetwork(๐)
14. return ๐ as ๐โ 15. ๐ก = ๐ก + 1 16. return ๐ as ๐โ
Pseudocode 2 Integrated traffic model.
Table 3 Traffic model scalability.
Network Nodes Links Origins Destinations Junctions Time
(seconds)
A 6 10 1 2 D 0.91
B 6 10 2 1 C 0.81
C 8 16 1 1 D-C 0.81
D 12 30 3 1 M 2.51
SF 24 76 4 1 M 3.88
ESF 282 334 21 2 M 33.5
Table Legend
Network Type: A, B,C,D โ Provided in Appendix, SF โ Sioux Falls, ESF โ Enriched Sioux Falls.
Junction Type: D โ Diverging, C โ Converging, M โ Multiple junction structures.
4. Formulation of Optimal Evacuation Strategy Model
In this section, we introduce and evaluate the evacuation optimisation model. We present the
final mathematical formulation to evaluate the effectiveness of the evacuation scheme as a
combination of equation sets (2) and (5):
Objective:
๐๐๐ฅ๐๐๐๐ ๐ ๐ = ๐ถ โ ๐พ๐ (6)
๐ถ = โ (1/๐ก โ ๐๐ข๐,๐๐ก
๐โ๐๐
)
๐กโ๐,๐โ๐
(6.1)
๐ =โ( โ ๐๐๐๐ค๐,๐๐ก
๐ โ๐ฎ,๐โ๐ช๐
โ โ ๐๐ข๐,๐๐ก
๐โ๐,๐โ๐๐
)
๐กโ๐
(6.2)
Subject to:
๐๐ข๐,๐๐ก = ฮ๐ก๐ โ๐๐๐,๐
๐
๐ก
๐=1
๐ โ ๐,
๐ก โ ๐ (7.1)
๐๐๐๐ค๐,๐๐ก = ฮ๐ก๐ โ๐๐๐ข๐ก,๐
๐
๐ก
๐=1
๐ โ ๐,
๐ก โ ๐ (7.2)
๐ท๐๐ก =
{
๐๐๐,๐
๐กโ๐ฟ๐๐ฃ๐ if ๐
๐ข๐,๐
๐กโ๐ฟ๐๐ฃ๐ = ๐๐๐๐ค๐,๐
๐ก
๐ถ๐ if ๐๐ข๐,๐๐กโ๐ฟ๐๐ฃ๐ > ๐๐๐๐ค๐,๐
๐ก
๐ โ ๐,
๐ก โ ๐ (7.3)
๐๐๐ก =
{
๐๐๐ข๐ก,๐
๐กโ๐ฟ๐๐ค๐ if ๐๐ข๐,๐
๐ก = ๐๐๐๐ค๐,๐
๐กโ๐ฟ๐๐ค๐ + ๐๐
๐๐๐๐ฟ๐
๐ถ๐ if ๐๐ข๐,๐๐ก < ๐
๐๐๐ค๐,๐
๐กโ๐ฟ๐๐ค๐ + ๐๐
๐๐๐๐ฟ๐
๐ โ ๐,
๐ก โ ๐ (7.4)
๐๐,๐๐ก,0 =
๐๐๐ข๐ก,๐๐ก ๐
๐,๐
๐๐๐ก,0
๐๐๐,๐๐ก โ๐ such that {๐, ๐} โ ๐
๐ก โ ๐ (7.5)
๐ด๐ฝ๐ก = {๐ผ๐ฝ,๐๐
๐ก }, ๐ผ๐ฝ,๐๐๐ก = โ ๐๐,๐
๐๐๐ก,0
๐โ๐,๐
๐ โ ๐๐ฝ,
๐ โ ๐ช๐ฝ,
๐ก โ ๐
(7.6)
๐๐๐,๐๐ก = ๐ ๐๐
๐ด๐ฝ [(๐ท๐๐ก)๐โ๐J , (๐๐
๐ก)๐โ๐ชJ
, (๐๐๐ก)๐โ๐J]
๐ โ ๐ช๐ฝ,
๐ก โ ๐ (7.7)
๐๐๐ข๐ก,๐๐ก = ๐ ๐๐
๐ด๐ฝ [(๐ท๐๐ก)๐โ๐J , (๐๐
๐ก)๐โ๐ชJ,
, (๐๐๐ก)๐โ๐J]
๐ โ ๐๐ฝ,
๐ก โ ๐ (7.8)
โ๐๐๐ก
๐โ๐J
= 1 ๐ฝ โ ๐ฑ (7.9)
๏ฟฝฬ ๏ฟฝ๐๐๐ โค ๐๐๐ก โค ๏ฟฝฬ ๏ฟฝ๐๐๐ฅ
๐ โ ๐๐ฝ,
๐ก โ ๐
๐ฝ โ ๐ฑ
(7.10)
๐๐๐ข๐ก,๐๐ก =โ๐๐
๐ก ๐ ๐๐ก
๐โ๐
๐๐,๐ ๐ โ ๐ช๐ ,
๐ก โ ๐ (7.11)
๐ธ โค ๐ ๐๐ก โค ๐
๐ โ ๐ซ,
๐ก โ ๐ (7.12)
โ ๐๐,๐ก ๐ ๐๐ก
๐กโ๐,๐โ๐
= ๐ต๐ ๐ โ ๐ฎ (7.13)
โ ๐๐๐๐ค๐,๐๐ก
๐โ๐๐
โค ๐๐ ๐ก โ ๐
๐ โ ๐ (7.14)
๐๐๐ก = {0,1}
๐ โ ๐ซ,
๐ก โ ๐ (7.15)
โ ๐๐๐ก
๐ โ๐ซ
= ๐พ ๐ก โ ๐ (7.16)
The objective function presented in Equation (6) remains analogous to Equation (2). It
aims to maximise the number of evacuees, promoting early evacuations, while reducing the
time vehicles remain in the network. Constraints (7.1-7.2) and (7.11-7.16) are carried over from
Section 2. Equations (7.1) and (7.2) describe the calculation of total vehicles in every link ๐ผ๐
among the arcs of the network ๐. Equation (7.11) determines the vehicle flow entering the
network from each origin at a defined rate ๐ ๐๐ก . Constraints (7.12-7.13) propose lower and upper
bounds on the departure rates based on available demand. Sink node capacity is captured by
constraint (7.14). Finally, the number of activated paths is constrained by Equations (7.15-
7.16).
The evacuation process is controlled by decision variables ๐๐๐ก, ๐ ๐
๐ก and ๐๐๐ก , where ๐๐
๐ก
determines the effective green time at critical junctions for all incoming links and is bounded
by (7.9-7.10). The evacuation staging strategy is defined by the number of users allowed to
evacuate at specified time-steps, ๐ ๐๐ก , and paths allowed to be accessed by the evacuees, ๐๐
๐ก .
Constraints (7.3-7.8) are carried over from the traffic model presented in Section 3.
Equations (7.3-7.4) define supply and demand for each link similar to (5.1-5.2), and (7.5-7.8)
reflect (5.4-5.7), determining flow distribution through junctions using the Riemann Solver.
The overall approach is structured as a Stackelberg game acting as a simulation-
optimisation framework (refer to Pseudocode 3), and proceeds as follows:
1. In the upper level, a metaheuristic algorithm generates a) an initial set of departure
vehicle rates ๐ ๐๐ก , b) signal phasing strategies ๐๐
๐ก based on (7.9-7.10) and (7.12) and c) a
set of K-shortest paths given network properties.
2. Following this, we execute the lower-level traffic model, evaluating a) the sets of
network flows ๐ as per (7.1-7.8, 7.11), b) activated paths ๐๐ described by (7.13-7.16) ,
and c) an overall fitness value ๐ as defined by equations (6-6.2).
3. Following every lower-level execution, we invoke the upper-level algorithm to obtain a
new set of departure rates and signal phases.
4. Steps 2-3 are repeated until the stopping criteria are met (either convergence or
maximum number of iterations).
1. Algorithm EvacuationOptimisationAlgorithm is
2. Input: network properties ๐, simulation time step ฮ๐ก๐ , decision
time step ฮ๐ก๐ initial flows ๐, time horizon ๐, iterations ฯ, graph ๐บ,
origins ๐, destinations ๐ท, number of paths ๐ 3. Output: optimal evacuation strategy ๐ โ 4.
5. initial evacuation strategy ๐ โ Generation()
6. initial network flows ๐ โ DetermineFlows(๐ ๐)
7. initial paths ๐พ โ YensAlgorithm(๐บ, ๐, ๐ท, ๐) 8. best evacuation strategy ๐ ๐
โ = โโ
9. while current iteration ๐ < ๐
10. vehicle flows ๐ โ TrafficModel(๐, ๐ฅ๐ก๐ , ๐ฅ๐ก๐, ๐, ๐พ, ๐, ๐บ, ๐, ๐ท, ๐) 11. solution fitness ๐ ๐ โ EvaluateSolution(๐)
12. if ๐ ๐โ โค ๐
13. ๐ โ โ ๐ 14. ๐ โ GenerateNewSolution(๐ ) 15. ๐ = ๐ + 1 16. return ๐ as ๐ โ
Pseudocode 3 Evacuation optimisation algorithm pseudocode.
4.1. Performance of Linear Formulation
The linearised problem presented in equations (3)-(4.9) is solved using the set of simple
networks provided in the Appendix. The key performance parameters explored include the
fitness value as described in equation (6); the weighted cumulative evacuation, defined as
โ (๐๐ข๐,๐๐ก /๐ก) ๐กโ๐ in the same equation; the total number of departures โ ๐๐๐๐ค๐,๐
๐ก๐กโ๐ , which
denotes the number of vehicles that are allowed to leave their residence; and the total number
of finished evacuations โ ๐๐ข๐,๐๐ก
๐กโ๐ . Finally, we compute the mean travel time for each
evacuation strategy.
Figure 1 Percentage improvement of bi-level model approach in comparison with linear model. Networks A-D are presented
in the Appendix, while SF refers to the Sioux Falls network presented in Section 4.2.
We present the results in Figure 1 comparing the performance of the linear problem and
the bi-level solution algorithm presented in Section 4. In most networks, the linear formulation
provides lower fitness score than the bi-level model. Similar performance levels are only
obtained in the smallest network A โ likely due to its simple layout, which only contains a
diverging junction. However, as the complexity of the network increases the deviation between
both solution increases. As an example, in network B which contains a single merging junction,
there is a 30% difference in fitness values between both methods. We attribute this behaviour
to the inability of the linear problem to consider dynamic travel time as a result of traffic
congestion.
Despite the model being solved with the traffic simulator iteratively until convergence,
optimisation techniques that can consider congestion effects during its evaluation provide a
better optimum. Therefore, the linear formulation proposes an insufficient framework to
evaluate and optimise evacuation strategies.
4.2. Test Network
The methodology is applied to the Sioux Falls Network (SFN), consisting of 24 nodes and 76
arcs (Bar-Gera, 2016). In this problem, a danger zone covers nodes 10, 11, 14 and 15, and a
single evacuation point is established in node 1 as shown in Figure 2a. Further iterations of the
model in Section 4.4 will contain additional sink nodes, damaged nodes and key junctions
presented in Figure 2a.
The minimum and maximum effective green time is set to ๏ฟฝฬ ๏ฟฝ๐๐๐ = 0.1 and ๏ฟฝฬ ๏ฟฝ๐๐๐ฅ = 0.6
for this problem instance. These values are informed based on the network topology (see Figure
2a), in which all nodes contain at least 4 connecting links, with most having at least 3 incoming
links. The above constraints ensure that effective green time is given to all links.
1
8
4 5 63
2
15 19
17
18
7
12 11 10 16
9
20
23 22
14
13 24 21
3
1
2
6
8
9
11
5
15
122313
21
16 19
17
2018 54
55
50
48
29
51 49 52
58
24
27
32
33
36
7 35
4034
41
44
57
45
72
70
46 67
69 65
25
28 43
53
59 61
56 60
66 62
68
637673
30
7142
647539
74
37 38
26
4 14
22 47
10 31
a) b)
Figure 2 Experiment test networks. The evacuation area is enclosed in a rectangle, with the evacuation points highlighted
with a triangle, and additional sink nodes with a dashed triangle. The signal phasing of key junctions highlighted with a
dashed square are controlled by the algorithm. a) Sioux-Falls network (SFN) with closed link presented in a circle. b)
Enriched Sioux-Falls network (ESFN)
The model aims to evacuate users generated in the danger zone and take them to the safe
node. All nodes within the evacuation area are assumed to have an equal pre-defined demand
of 3,200 vehicles, flooding the network in case of simultaneous departure.
A large-scale version of the SFN is used to determine its scalability and applicability
(Figure 2b), consisting of 282 nodes and 334 arcs (Chakirov and Fourie, 2014). A small
alteration to the evacuation procedure is introduced to compensate for the larger network size
โ we introduce a secondary sink node in node 2.
It is important to note that, due to the new network structure, the source nodes increase
from 4 to 20 nodes to create the same danger zone, each with a demand of 640 vehicles. For
the remainder of this paper, we will refer to this network as Enriched Sioux Falls Network
(ESFN).
We propose a selection of Derivative-Free Algorithms (DFO) to evaluate in Table 4.
Similar to simulation-optimisation strategies algorithms, DFOs are categorised into local
greedy and global elitist algorithms. Within local search, direct methods continuously evaluate
trial solutions, and model-based search creates a surrogate model to perform a guided search.
Global deterministic and model-based search are analogous to the local direct and model-based
search counterparts but partition the search space. Finally, stochastic methods allow
intermediate lower quality solutions to avoid local optima.
We consider both local and global methods in our test, including deterministic and
stochastic methods. A scenario with 2 activated paths ๐ and a decision time step size ฮ๐ก๐ of 10
seconds was selected to test all algorithms, using the SFN.
As shown in Figure 3, local methods provide far from optimal solutions due to their
inability to escape local optima. On the performance of the global algorithms, the Simulated
Annealing and the Genetic Algorithm outperform other solvers. The latter provides the best
solution, with the highest number of departures, evacuation and shortest mean travel time.
Given the results of this analysis, we chose to adopt the Genetic Algorithm as the
optimisation solver used by the bi-level approach. This algorithm consists of three main
components: selection, crossover and mutation.
We use a binary tournament to pair solutions and select the best as indicated by their
fitness value. Although this process discards well-performing solutions as does not compare
candidates to the population, it can be parallelised to improve computational performance
(Noraini and Geraghty, 2011).
The crossover operator combines two parent solutions to create two new unique
solutions. We use a BLX-alpha crossover method as it allows exploration of solution further
than the parents (Eshelman and Schaffer, 1993). The size of the search space explored is
defined by parameter ๐ผ, where a higher value increases the search space.
Finally, mutation alters the parent solutions to create unique solutions outside the range
of the parent solutions. We use the non-uniform mutation (Michalewicz, 1992), which reduces
the mutation range after every generation, encouraging convergence as the algorithm runs. The
decrease in range rate is controlled by the ๐ฝ parameter, where a greater value of ๐ฝ increases
the rate of decrease.
Figure 3 Key performance indicators for different Derivative Free Algorithms.
Table 4 Derivative-Free Optimisation algorithms used.
DFO Algorithm Type Solver Author
Nelder-Mead Simplex (NM) Local Direct fminsearch Mathworks
Implicit Filtering (IF) Local Model-Based imfil Kelley (2011)
Branch and Fit (BF) Global
Deterministic snobfit
Huyer & Neumaier
(2008)
Particle Swarm (PS) Global Stochastic Pswarm Vaz & Vicente (2007)
Simulated Annealing (SA) Global Stochastic Original -
Genetic Algorithm (GA) Global Stochastic Original -
4.3. Parameter Tuning
We proceed with manual tuning to determine suitable control parameters for the genetic
algorithm. The control parameters considered include ๐ผ and ๐ฝ, which control crossover and
mutation search spaces, and ๐๐ and ๐๐, which represent probability parameters for crossover
and mutation respectively. A total of 9 scenarios are considered, with their parameters
summarised in Table 5. The evacuation parameters decision time step ฮ๐ก๐, and activated paths
๐พ remain fixed in all cases. The scenario naming convention describes the evacuation
parameters used such that K1 denotes ๐พ = 1 and D10 indicates that ฮ๐ก๐ = 10.
As observed in Figure 4, higher values of ๐ผ and ๐ฝ provide higher fitness scores than the
original K2D10. Both parameters provide initially a large search space that gradually decreases
in later iterations.
It can be observed that increasing crossover rate let to improvements in performance, as
well as reducing the mutation rate. The first provides more solution combinations, while the
second reduces the stochasticity of solutions. Despite mutation providing a means to deviate
from local optima, high mutation probabilities may result in a high-degree stochastic search
algorithm rather than elitist search mechanism. Conversely, providing a very low value does
not allow the algorithm to explore sufficient are of the solution surface to find a close to the
optimal solution, which is the case in m2. Table 5 Sioux Falls evacuation calibration scenarios.
Scenario ๐พ ฮ๐ก๐ ๐ผ ๐ฝ ๐๐ ๐๐
K2D10-n 2 10 0.5 0.5 0.5 0.1
K2D10-a1 2 10 0.6 0.5 0.5 0.1
K2D10-a2 2 10 0.8 0.5 0.5 0.1
K2D10-b1 2 10 0.5 0.6 0.5 0.1
K2D10-b2 2 10 0.5 0.8 0.5 0.1
K2D10-c1 2 10 0.5 0.5 0.6 0.1
K2D10-c2 2 10 0.5 0.5 0.8 0.1
K2D10-m1 2 10 0.5 0.5 0.5 0.05
K2D10-m2 2 10 0.5 0.5 0.5 0.01
Figure 4 Key performance indicators for parameter calibration of the Genetic Algorithm. Refer to Table 5 for scenario
names. Scenario โnโ refers to the standard K2D10-n.
4.4. Results and discussion
The algorithm is run under the scenarios presented in Table 9 under the network configurations
presented in Section 4.2. An initial run is executed without considering signal phasing to
evaluate the effectiveness of evacuation staging. A benchmark problem is used to simulate a
worst-case solution in which all demand is released simultaneously at the beginning of the
evacuation.
The model was run on a modest workstation with an Intel Core i7-4790 CPU and 8GB
RAM. The runtime increases given a higher value of ๐, but the variation is not substantial, as
all the scenarios required approximately 5 and 16 hours to complete in the SFN and the ESFN
respectively. This level of performance was deemed to be acceptable for pre-planning purposes
yet unsuitable for real-time implementation. Further efforts to accelerate the algorithm are to
increase parallelisation of the genetic algorithm or the use of high-performance cloud
computing facilities.
Table 6 Sioux Falls evacuation calibration scenarios.
Scenario ๐ ฮ๐ก๐ ๐ผ ๐ฝ ๐๐ ๐๐
Benchmark 5 10 - - - -
K1D10 1 10 0.6 0.8 0.6 0.05
K1D20 1 20 0.6 0.8 0.6 0.05
K2D10 2 10 0.6 0.8 0.6 0.05
K2D20 2 20 0.6 0.8 0.6 0.05
K3D10 3 10 0.6 0.8 0.6 0.05
K3D20 3 20 0.6 0.8 0.6 0.05
The results are summarised in Table 7. The benchmark scenario involves an
uncoordinated evacuation which saturates the network immediately after the evacuation
warning is emitted. As expected, this is the worst performing problem instance - the evacuation
process is not able to recover from the immediate surge in demand, with the rate of vehicle
arrivals at the evacuation point decreasing over time (Figure 5a). This scenario also yields the
lowest total departures and evacuations among all scenarios, meaning that congestion occurs
close to the source nodes, preventing vehicles from accessing the network.
The remaining instances performed better due to a combination of reduced decision time-
steps and increased activated route paths. The provision of additional paths increases the
network capacity, thus allowing a higher rate of evacuations to take place. However, the
evacuation improvement is reduced for ๐ > 2 due to greater convergence between the paths.
An alternate run is carried out with signal phasing control at node 3, which provided
improvements in fitness value, departure and arrival of evacuees in all the scenarios, at the cost
of a marginal increase in mean travel time.
The same experiment is carried out under an alternate Sioux Falls network containing
two more sink nodes (Figures 2a), as well as one where links 29 and 48 are unavailable. In
addition, every network is tested with signal phasing control at nodes 3 and 8 to evaluate its
benefit. The main benefit of adding more sink nodes is the increased proportion of evacuees
arriving to their destination, as well as a reduction in wait time. As a result, despite reduction
in the total departures and arrivals, an increase in fitness value is observed, particularly for
scenarios ๐ โฅ 2 activated paths. This is further reflected in a reduced mean evacuee travel time.
a) b)
c) d)
Figure 5 Performance parameters for the SFN case. a) Cumulative weighted sum of evacuees over time. b) Proportion of
evacuated users over time in case K2D10. c) Travel time and distance ratio. d) Travel distance deviation between users.
The introduction of signal phasing results in a general increase in departures and arrival
rates in all scenarios, yet mean travel time and distance deviation remain consistent (Figure 5c
and 5d). With the addition of a destroyed link at 29 and 48, longer evacuation times are seen
throughout the network. Figure 5c indicates that the increased travel time is directly related to
the increased path distance as a result of the link destruction, given that travel time and distance
ratio varies similarly in all network configurations.
Concerning the arrival and departure rates, a significant increment is observed for ๐ = 1
scenarios. However, the opposite behaviour is observed for larger values of ๐, resulting in a
reduction in completed evacuations.
Finally, as shown in Figure 5b, the damage network case results in reduced proportion of
completed evacuations throughout the evacuation process. Note the proportion of evacuations is
conceived using the number of users that have entered the network. Hence, a reduction in the
proportion of completed evacuations is possible if the numbers of entries to the network exceed
the number of exits.
With respect to the sink arrival rates, in all cases sink node 1 receives the greatest
proportion of users (Figure 7). As expected, the disparity increases when we introduce the
destroyed link, particularly in the absence of signal control optimisation. This is a result of
userโs route choice given perceived travel time throughout the network.
Another observation relates to the constant evacuation rate when ๐ก > 100, as seen in
Figure 5a. Note that a similar behaviour is observed in all network configurations. Therefore,
the solutions not only significantly improve upon a do-nothing approach, but also create a
stable evacuation process that allows users to exit the network at a constant rate. Interestingly,
the do-nothing solution performs similarly to the rest at the beginning of the evacuation, under
free-flow network conditions. Evacuation rate promptly decreases as congestion starts
affecting the evacuation process.
The algorithm provides more stable results in the ESFN case study, where we observe an
improvement in the fitness value when increasing the number of activated paths and when
reducing the decision time step (Table 8). A more significant increase is observed when
introducing signal phasing, becoming more significant in the case where ๐ = 3.
The inclusion of additional sink nodes and signal phasing result in an increase in fitness,
total departures and evacuations in expense of travel time. Both also result in longer trips, and
greater deviation in the distance travelled between the evacuees (Figure 6c and 6d). As
observed in the SFN, signal phasing provides further improvements in evacuation completion
and fitness values.
a) b)
c) d) Figure 6 Performance parameters for the ESFN case. a) Cumulative weighted sum of evacuees over time. b) Proportion of
evacuated users over time in case K2D10. c) Travel time and distance ratio. d) Travel distance deviation between users.
The results suggest that demand management is an effective method to optimise
evacuations as all solutions improve evacuation upon the benchmark scenario and create stable
evacuation processes. The behaviour is represented by the constant gradient in Figure 5a and
Figure 6a.
It is interesting to note that the benchmark solution significantly reduces in effectiveness
as the evacuation is carried out, which may be particularly misleading for coordinating agencies
on the effectiveness of the evacuation. Of particular interest is the behaviour Figure 6a, where
the benchmark evacuation outperforms K1 strategies at the beginning of the evacuation.
Figure 7 Percentage evacuees accessing each node. K2D10 scenario.
a) b)
Figure 8 GA performance โ fitness value variation over runtime under no signal phasing. a) SFN. b) ESFN.
On the evacuation performance under each network, it is evident that more vehicles
successfully evacuated the SFN, and the proportion of completed evacuation among departing
users is significantly smaller in the ESFN. This is a result of the smaller network density ESFN,
which contains several junctions (nodes) within what would be a single link in the SFN.
The provision of additional paths leads to reduced performance gains. The demand
monitoring scheme provides a more significant evacuation improvement than the subsequent
path provision. In all network configurations (except under the damaged network) greater
increase in fitness values is seen for ๐ โค 2. The variability suggests that higher path
convergence exists in the ESFN, and the performance increase is a result of the additional
capacity. Further performance gains are achieved using signal phasing despite the generalised
increase in travelling distance deviation, being more significant in the ESFN.
Figures 8a and 8b indicate that in both networks the algorithm converges into a stable
solution. As expected, problem instances with a more substantial number of activated paths
require a longer time to converge due to a higher number of decision variables. Although not
presented, a similar behaviour is observed when considering signal phasing and alternative
network conditions. Potential solutions include the use of a modified version of the k-shortest
path algorithm, or the consideration of alternative evacuation strategies that utilise crossing
elimination, which would reduce network congestion.
While the presented algorithm provides a platform to model urban evacuation strategies
using staging and signal phasing for pre-planning purposes, some improvements would provide
more robust results. Despite the high car ownership rates in Sioux Falls (United States Census
Bureau, 2017), the assumption that only a single mode of transport is used is unrealistic in
urban networks (Abdelgawad and Abdulhai, 2010). In addition, we consider a single type of
evacuee and do not identify priority groups.
While previous work limit destination choice by evacuee type (Bish et al. 2014),
distinctions between users regarding route choice and driving behaviours would provide more
realistic results. Furthermore, despite the increase in evacuation performance due to demand
staging, the study does not consider potential harm, physical or psychological, to evacuees are
ordered to wait at departure nodes (Hu et al., 2017). Finally, although some choice is given to
the users, the algorithm assumes that vehicles will not deviate from their initial route choice
and will comply with indications from evacuation controllers. Modelling compliance requires
research on educational and social factors, as well as providing greater uncertainty.
Table 7 Sioux Falls network results.
Network
Fitness Value
(vehicles/
time step)
Weighted
evacuations
(vehicles/
time step)
Total
departures
(vehicles)
Total
evacuations
(vehicles)
Mean travel
time
(time steps)
Value % Value % Value % Value % Value %
SF
N
BM 5,341 0 5,458 0 5,991 0 4,934 0 204.6 0
K1D10 8,664 62 8,831 62 11,302 89 9,771 98 192.1 6
K1D20 8,554 60 8,723 60 11,595 94 9,878 100 182.2 11
K2D10 9,091 70 9,259 70 11,393 90 10,254 108 187.6 8
K2D20 8,994 68 9,157 68 12,298 105 10,156 106 184.3 10
K3D10 9,199 72 9,347 71 11,679 95 10,150 106 180.6 12
K3D20 9,172 72 9,341 71 11,570 93 10,308 109 181.3 11
SF
N w
ith
sig
nal
ph
asin
g
BM 5,341 0 5,458 0 5,991 0 4,934 0 204.6 0
K1D10 8,604 61 8,775 60 11,075 85 9,460 92 187.2 9
K1D20 8,682 63 8,854 62 11,629 94 9,755 98 185.9 9
K2D10 9,164 72 9,318 71 11,890 98 10,319 109 188.3 8
K2D20 9,091 70 9,147 68 11,845 98 10,223 107 190.7 7
K3D10 9,218 73 9,401 72 11,501 92 10,223 107 186.5 9
Table 8 Enriched Sioux Falls network results.
Network
Fitness Value
(vehicles/
time step)
Weighted
evacuations
(vehicles/
time step)
Total
departures
(vehicles)
Total
evacuations
(vehicles)
Mean travel
time
(time steps)
Value % Value % Value % Value % Value %
ES
FN
BM 1,233 0 1,517 0 3,920 0 1,491 0 401.8 0
K1D10 1,949 58 2,321 53 8,025 105 3,584 140 460.7 -15
K1D20 1,933 57 2,273 50 7,495 91 3,557 139 470.5 -17
K2D10 2,185 77 2,534 67 6,856 75 3,705 148 447.5 -11
K2D20 2,187 77 2,521 66 6,819 74 3,647 145 449.2 -12
K3D10 2,269 84 2,572 70 6,185 58 3,562 139 429.1 -7
K3D20 2,234 81 2,546 68 6,346 62 3,487 134 437.7 -9
ES
FN
wit
h
sig
nal
lin
g
BM 1,233 0 1,517 0 3,920 0 1,491 0 401.8 0
K1D10 2,257 84 2,431 60 5,106 30 3,593 141 291.7 27
K1D20 2,224 80 2,388 57 5,313 36 3,586 141 318.9 21
K2D10 2,270 84 2,541 68 6,419 64 3,697 148 417.6 -4
K2D20 2,318 88 2,523 66 6,036 54 3,713 149 327.5 18
K3D10 3,004 144 3,266 115 6,824 74 4,457 199 386.1 4
K3D20 2,975 141 3,253 114 7,327 87 4,535 204 356.2 11
K3D20 9,268 74 9,463 73 11,963 100 10,328 109 190.2 7 BM 6,687 0 6,783 0 6,702 0 5,923 0 177.3 0
SF
N-m
K1D10 9,294 39 9,477 40 10,580 58 9,036 63 212.4 6
K1D20 9,483 42 9,634 42 10,071 50 9,625 65 183.7 17
K2D10 9,344 40 9,565 41 10,429 56 9,747 71 197.7 -12
K2D20 9,288 39 9,746 44 10,798 61 10,149 53 191.3 -8
K3D10 10,613 59 10,089 49 12,367 85 10,850 63 166.6 -20
K3D20 10,605 59 12,209 80 12,800 91 12,176 65 147.7 -4
SF
N-m
wit
h s
ign
al
ph
asin
g
BM 6,687 0 6,783 0 6,702 0 5,923 0 177.3 0
K1D10 10,108 51 10,277 52 10,344 54 9,291 57 202.1 -14
K1D20 10,143 52 10,265 51 10,918 63 9,364 58 144.4 19
K2D10 10,057 50 10,245 51 11,476 71 9,489 60 193.5 -9
K2D20 10,346 55 10,435 54 11,455 71 9,270 57 194.5 -10
K3D10 10,484 57 10,653 57 11,466 71 10,213 72 196.8 -11
K3D20 10,579 58 11,556 70 12,460 86 11,391 92 211.0 -19
SF
N-m
d
BM 6,286 0 6,376 0 6,771 0 5,872 0 181.0 0
K1D10 9,645 53 9,770 53 11,732 73 10,167 73 163.4 10
K1D20 10,310 64 10,435 64 10,569 56 9,578 63 171.5 5
K2D10 9,750 55 9,948 56 11,911 76 10,330 76 194.4 -7
K2D20 9,144 45 9,338 46 11,593 71 10,334 76 189.4 -5
K3D10 11,412 82 11,628 82 12,800 89 11,316 93 195.2 -8
K3D20 11,926 90 12,101 90 12,800 89 12,023 105 168.6 7
SF
N-m
d w
ith
sig
nal
ph
asin
g
BM 6,286 0 6,376 0 6,771 0 5,872 0 181.0 0
K1D10 10,281 64 10,442 64 12,428 84 11,190 91 184.9 -2
K1D20 10,729 71 10,886 71 12,800 89 11,050 88 167.1 8
K2D10 10,012 59 10,070 58 12,609 86 11,166 90 197.9 -9
K2D20 10,047 60 10,432 64 12,430 84 10,752 83 188.1 -4
K3D10 12,091 92 12,291 93 12,800 89 12,363 111 183.6 -1
K3D20 12,266 95 12,460 95 12,800 89 12,454 112 184.6 -2
ES
FN
-m
BM 1,751 0 2,015 0 4,039 0 1,346 0 272.4 0
K1D10 3,778 116 3,984 98 6,488 61 4,790 256 306.4 12
K1D20 3,701 111 4,004 99 7,993 98 4,937 267 365.1 34
K2D10 4,011 129 4,341 115 8,515 111 5,009 272 353.5 30
K2D20 3,923 124 4,219 109 8,118 101 4,950 268 346.9 27
ES
FN
-m w
ith
sig
nal
lin
g
BM 1,751 0 2,015 0 4,039 0 1,346 0 272.4 0
K1D10 4,257 143 4,481 122 7,060 75 5,475 307 310.1 14
K1D20 4,160 138 4,425 120 7,549 87 5,374 299 332.2 22
K2D10 4,573 161 4,886 142 9,233 129 5,835 333 333.9 23
K2D20 4,395 151 4,680 132 8,449 109 5,773 329 320.0 17
5. Conclusion
In this paper, we develop a bi-level modelling framework that is used to optimise evacuation
strategies using a network demand management approach. A driving principle that was adopted
in its design is that centralised control is difficult to apply in emergency evacuations, therefore
giving users an increased level of control over their choice of paths, when compared to previous
studies. An efficient traffic flow model is adopted, implemented as a black-box model that
captures traffic behaviour patterns during the evacuation.
The methodology was tested in two different networks under a variety of scenarios. The
analysis of the results indicates that demand management is an effective method to improve
the efficiency of evacuations โ with improvements observed across all key performance
indicators when compared to the benchmark case. Further improvements in evacuation
parameters were observed when controlling signal phasing simultaneously. A range of
Derivative-Free Optimisation algorithms was evaluated, with Genetic Algorithms shown to
outperform other options and adopted as the solver that underpins the overall solution search
process.
Even though a degree of personal choice (for routing decisions) is considered by this
framework, further work is essential to establish the impact of varying rates of departure time
compliance and information quality. Our future research in this area will explore the
applicability of probabilistic decision-making techniques, including the use of myopic routing
models. Monte Carlo simulation can be used to model a range of user behaviours, which is
expected to come at the expense of longer runtimes. Furthermore, the use of a variety of
transport modes should be considered to aid population groups with accessibility problems or
lack of access to private vehicles.
Acknowledgements
The research was supported by the UK Engineering and Physical Sciences Research Council
(EPSRC) as part of the Sustainable Civil Engineering Centre for Doctoral Training (Grant
number EP/L016826/1).
References
Abdelgawad, H. and Abdulhai, B. (2009) โEmergency evacuation planning as a network
design problem: a critical reviewโ, Transportation Letters: International Journal of
Transportation Research, 1(December), pp. 41โ58. doi: 10.3328/TL.2009.01.01.41-58.
Abdelgawad, H. and Abdulhai, B. (2010) โManaging Large-Scale Multimodal Emergency
Evacuationsโ, National Evacuation Conference, 2(2), pp. 122โ151. doi:
10.1080/19439962.2010.487636.
Amaran, S. et al. (2016) โSimulation optimization: a review of algorithms and applicationsโ,
Annals of Operations Research, 240(1), pp. 351โ380. doi: 10.1007/s10479-015-2019-x.
Bar-Gera, H. (2016) Transportation Network Test Problems.
Bish, D. R., Sherali, H. D. and Hobeika, a G. (2014) โOptimal evacuation planning using
staging and routingโ, Journal of the Operational Research Society, 65(1), pp. 124โ140. doi:
10.1057/jors.2013.3.
Bretschneider, S. and Kimms, A. (2012) โPattern-based evacuation planning for urban areasโ,
European Journal of Operational Research, 216(1), pp. 57โ69. doi:
10.1016/j.ejor.2011.07.015.
Chakirov, A. and Fourie, P. J. (2014) Enriched Sioux Falls Scenario with Dynamic And
Disaggregate Demand. Singapore.
Chen, X. and Zhan, F. B. (2008) โAgent-based modelling and simulation of urban
evacuationโฏ: relative effectiveness of simultaneous and staged evacuation strategies.โ, Journal
of the Operational Research Society, 59(1), pp. 25โ33. doi: 10.1057/palgrave.jors.2602321.
Chien, S. I. and Korikanthimath, V. V. (2007) โAnalysis and Modeling of Simultaneous and
Staged Emergency Evacuationsโ, Journal of Transportation Engineering, 133(3), pp. 190โ
197. doi: 10.1061/(ASCE)0733-947X(2007)133:3(190).
Chiu, Y. C. et al. (2007) โModeling no-notice mass evacuation using a dynamic traffic flow
optimization modelโ, IIE Transactions (Institute of Industrial Engineers), 39(1), pp. 83โ94.
doi: 10.1080/07408170600946473.
Cova, T. J. and Johnson, J. P. (2003) โA network flow model for lane-based evacuation
routingโ, Transportation Research Part A: Policy and Practice, 37(7), pp. 579โ604. doi:
10.1016/S0965-8564(03)00007-7.
Dijkstra, E. W. (1959) โA note on two problems in connexion with graphsโ, Numerische
Mathematik, 1(1), pp. 269โ271. doi: 10.1007/BF01386390.
Doan, K. and Ukkusuri, S. V. (2012) โOn the holding-back problem in the cell transmission
based dynamic traffic assignment modelsโ, Transportation Research Part B: Methodological.
Elsevier Ltd, 46(9), pp. 1218โ1238. doi: 10.1016/j.trb.2012.05.001.
Eshelman, L. J. and Schaffer, J. D. (1993) โReal-Coded Genetic Algorithms and Interval-
Schemataโ, (July), pp. 187โ202. doi: 10.1016/B978-0-08-094832-4.50018-0.
Even, C., Pillac, V. and Hentenryck, P. Van (2015) โConvergent Plans for Large-Scale
Evacuationsโ, Proceedings of the 29th AAAI Conference on Artificial Intelligence (AAAI-15),
pp. 1121โ1127.
Fehr, I. et al. (2005) โManaging Tsunami Risk in the Aftermath of the 2004 Indian Ocean
Earthquake & Tsunamiโ, Risk Management Solutions, pp. 1โ24.
Friesz, T. L. et al. (1993) โA variational inequality formulation of the dynamic network user
equilibrium problemโ, Operations Research, 41(1), pp. 80โ91.
Garavello, M., Han, K. and Piccoli, B. (2016) Models for Vehicular Traffic on Networks.
Springfield, MO: Americal Institute of Mathematical Sciences.
Goldblatt, R. B. (2004) โEvacuation Planning, Human Factors and Traffic Engineering
Perspectivesโ, in Proceedings of the 2004 European Transport Conference, pp. 1โ21.
Hallegatte, S. (2012) โA Cost Effective Solution to Reduce Disaster Losses in Developing
Countries. Hydro-Meteorological Services, Early Warning, and Evacuationโ, World Bank
policy research working paper, (May), p. 22. doi: 10.1596/1813-9450-6058.
Han, K. et al. (2014) โOn the continuum approximation of the on-and-off signal control on
dynamic traffic networksโ, Transportation Research Part B: Methodological. Elsevier Ltd,
61, pp. 73โ97. doi: 10.1016/j.trb.2014.01.001.
Han, K., Friesz, T. L. and Yao, T. (2013) โExistence of simultaneous route and departure
choice dynamic user equilibriumโ, Transportation Research Part B: Methodological, 53, pp.
17โ30. doi: 10.1016/j.trb.2013.01.009.
Han, K. and Gayah, V. V. (2015) โContinuum signalized junction model for dynamic traffic
networks: Offset, spillback, and multiple signal phasesโ, Transportation Research Part B:
Methodological. Elsevier Ltd, 77, pp. 213โ239. doi: 10.1016/j.trb.2015.03.005.
Han, K., Piccoli, B. and Szeto, W. Y. (2016) โContinuous-time link-based kinematic wave
model: formulation, solution existence, and well-posednessโ, Transportmetrica B, 4(3), pp.
187โ222. doi: 10.1080/21680566.2015.1064793.
Hasegawa, R. (2013) โDisaster Evacuation from Japanโs 2011 Tsunami Disaster and the
Fukushima Nuclear Accidentโ, Iddri, (May), p. 54.
Hobeika, A. G. and Kim, C. (1998) โComparison of traffic assignments in evacuation
modelingโ, IEEE Transactions on Engineering Management, 45(2), pp. 192โ198. doi:
10.1109/17.669768.
Hu, Z. H. et al. (2017) โPost-disaster relief operations considering psychological costs of
waiting for evacuation and relief resourcesโ, Transportmetrica A: Transport Science, 13(2),
pp. 108โ138. doi: 10.1080/23249935.2016.1222559.
Huyer, W. and Neumaier, A. (2008) โSNOBFIT - Stable Noisy Optimization by Branch and
Fitโ, ACM Transactions on Mathematical Software, 35(2), pp. 1โ25. doi:
10.1145/1377612.1377613.
Jabari, S. E., He, X. and Liu, H. X. (2012) โHeuristic Solution Techniques for No-Notice
Emergency Evacuation Traffic Managementโ, in Network Reliability in Practice, pp. 241โ
259. doi: 10.1007/978-1-4614-0947-2.
Jin, W. L. (2015) โContinuous formulations and analytical properties of the link transmission
modelโ, Transportation Research Part B: Methodological, 74, pp. 88โ103. doi:
10.1016/j.trb.2014.12.006.
Kelley, C. (2011) Usersโ Guide for imfil Version 1.0, North Carolina State University,.
Kimms, A. and Maiwald, M. (2017) โAn exact network flow formulation for cell-based
evacuation in urban areasโ, Naval Research Logistics, 64(7), pp. 547โ555. doi:
10.1002/nav.21772.
Kimms, A. and Maiwald, M. (2018) โBi-objective safe and resilient urban evacuation
planningโ, European Journal of Operational Research. Elsevier B.V., 269(3), pp. 1122โ1136.
doi: 10.1016/j.ejor.2018.02.050.
Lebacque, J. P. and Khoshyaran, M. M. (1999) โModelling vehicular traffic flow on networks
using macroscopic modelsโ, in Vilsmeier, R., Benkhaldoun, F., and Hanel, D. (eds) Finite
volumes for complex applications II. Duisburg, pp. 551โ558.
Li, Q., Yang, X. and Heng, W. (2006) โIntegrating Traffic Simulation Models with
Evacuation Planning System in a GIS Environmentโ, in 2006 IEEE Intelligent Transportation
Systems Conference. Toronto, Canada, pp. 590โ595. doi: 10.1109/ITSC.2006.1706805.
Li, S., Zhou, Q. H. and Ju, Y. (2010) โA mixed integer linear program for the single
destination system optimum dynamic traffic assignment problem with physical queueโ,
Promet - Traffic & Transportation, 22(4), pp. 245โ249.
Li, X. et al. (2012) โA novel method for planning a staged evacuationโ, Journal of Systems
Science and Complexity, 25(6), pp. 1093โ1107. doi: 10.1007/s11424-012-0257-4.
Li, X., Li, Q. and Claramunt, C. (2018) โA time-extended network model for staged
evacuation planningโ, Safety Science. Elsevier, 108(May 2016), pp. 225โ236. doi:
10.1016/j.ssci.2017.08.004.
Lighthill, M. J. and Whitham, G. B. (1955) โOn Kinematic Waves. II. A Theory of Traffic
Flow on Long Crowded Roadsโ, Proceedings of the Royal Society A: Mathematical, Physical
and Engineering Sciences, 229(1178), pp. 317โ345. doi: 10.1098/rspa.1955.0089.
Liu, S. and Murray-Tuite, P. (2011) Analysis and Evaluation of Household Pick-up and
Gathering Behavior in No-Notice Evacuations. Virginia Polytechnic Institute and State
University.
Liu, Y., Lai, X. and Chang, G. (2006) โCell-Based Network Optimization Model for Staged
Evacuation Planningโ, Transportation Research Record: Journal of the Transportation
Research Board, 1964, pp. 127โ135.
Lu, Q., George, B. and Shekhar, S. (2005) โCapacity Constrained Routing Algorithms for
Evacuation Planning: A Summary of Resultsโ, Advances in spatial and temporal databases,
(81655), pp. 291โ307. doi: 10.1007/11535331_17.
Mahar, A. et al. (2015) โInternational Journal of Disaster Risk Reduction Devastating storm
surges of Typhoon Haiyanโ, International Journal of Disaster Risk Reduction, 11, pp. 1โ12.
doi: 10.1016/j.ijdrr.2014.10.006.
Margesson, R. and Taft-Morales, M. (2010) Haiti Earthquakeโฏ: Crisis and Response,
Congressional Research Service. Available at: https://www.fas.org/sgp/crs/row/R41023.pdf.
Michalewicz, Z. (1992) Genetic Algorithms + Data Structures = Evolution Programs. 3rd
edn. New York, NY: Springer.
Mitchell, J. T., Cutter, S. L. and Edmonds, A. S. (2007) โImproving Shadow Evacuation
Management: Case Study of the Grianteville, South Carolina, chlorine spillโ, Journal of
Emergency Management, 5(1), pp. 28โ34.
Mitchell, S. W. and Radwan, E. (2006) โHeuristic Prioritization of Emergency Evacuations
Staging to Reduce Clearance Timeโ, in Proceedings of the 85th Annual Meeting
Transportation Research Board. Washington, DC, USA.
Munawir, R. (2012) Historic quake off Indonesia causes panic, but no huge tsunami,
National Post. Available at: http://nationalpost.com/news/historic-quakes-off-indonesia-
cause-panic-but-no-huge-tsunami.
Murray-Tuite, P. and Wolshon, B. (2013) โEvacuation transportation modeling: An overview
of research, development, and practiceโ, Transportation Research Part C: Emerging
Technologies, 27, pp. 25โ45. doi: 10.1016/j.trc.2012.11.005.
Nagel, K. and Flรถtterรถd, G. (2012) โAgent-Based Traffic Assignment: Going From Trips to
Behavioral Travelersโ, in Bhat and Pendyala (eds) Travel Behaviour Research in an Evolving
World, pp. 261โ293.
News24 (2012) Indonesia revises quake death toll, News24. Available at:
https://www.news24.com/World/News/Indonesia-revises-quake-death-toll-20120413.
Nguyen, A.-T., Reiter, S. and Rigo, P. (2014) โA review on simulation-based optimization
methods applied to building performance analysisโ, Applied Energy, 113, pp. 1043โ1058. doi:
10.1016/j.apenergy.2013.08.061.
Noraini, M. and Geraghty, J. (2011) โGenetic algorithm performance with different selection
strategies in solving TSPโ, in World Congress on Engineering. London, UK, pp. 4โ9.
Pillac, V., Van Hentenryck, P. and Even, C. (2016) โA conflict-based path-generation
heuristic for evacuation planningโ, Transportation Research Part B: Methodological, 83, pp.
136โ150. doi: 10.1016/j.trb.2015.09.008.
Rana, N. et al. (2013) โRecent and past floods in the alaknanda valley: Causes and
consequencesโ, Current Science, 105(9), pp. 1209โ1212.
Rautela, P. (2013) โLessons learnt from the Deluge of Kedarnath, Uttarakhand, Indiaโ, Asian
Journal of Environment and Disaster Management, 5(2), p. 167. doi:
10.3850/S1793924013002824.
Richards, P. (1956) โShock Waves on the Highwayโ, Operations Research, 4(1), pp. 42โ51.
Sabouhi, F. et al. (2018) โA robust possibilistic programming multi-objective model for
locating transfer points and shelters in disaster reliefโ, Transportmetrica A: Transport
Science. Taylor & Francis, 0(0), pp. 1โ28. doi: 10.1080/23249935.2018.1477852.
Sbayti, H. and Mahmassani, H. (2006) โOptimal Scheduling of Evacuation Operationsโ,
Transportation Research Record, 1964(1964), pp. 238โ246. doi: 10.3141/1964-26.
Seekircher, K. (2017) Evacuation Planning under Selfish Evacuation Routing. Duisburg-
Essen University.
Shen, W., Nie, Y. and Zhang, H. M. (2007) โDynamic network simplex method for designing
emergency evacuation plansโ, Transportation Research Record, pp. 83โ93. doi:
10.3141/2022-10.
Shimamoto, H. et al. (2018) โEvaluation of tsunami evacuation planning considering vehicle
usage and start timing of evacuationโ, Transportmetrica A: Transport Science, 14(1โ2), pp.
50โ65. doi: 10.1080/23249935.2017.1290708.
Sorensen, J. (2000) โHazard warning systems: review of 20 years of progressโ, Natural
Hazards Review, 1(2), pp. 119โ125. doi: 10.1061/(ASCE)1527-6988(2000)1:2(119).
Szeto, W. Y. and Lo, H. K. (2004) โA cell-based simultaneous route and departure time
choice model with elastic demandโ, Transportation Research Part B: Methodological, 38(7),
pp. 593โ612. doi: 10.1016/j.trb.2003.05.001.
Thomas, A. and Kopczak, L. (2007) โLife-Saving Supply Chainsโ, in Building Supply Chain
Excellence in Emerging Economies, pp. 93โ111. doi: 10.1080/15470620802059315.
Tuydes-Yaman, H. and Ziliaskopoulos, A. (2014) โModeling demand management strategies
for evacuationsโ, Annals of Operations Research, 217(1), pp. 491โ512. doi: 10.1007/s10479-
014-1533-6.
United States Census Bureau (2017) โAmerican Community Survey 1-year Estimateโ.
Urbina, E. and Wolshon, B. (2003) โNational review of hurricane evacuation plans and
policies: A comparison and contrast of state practicesโ, Transportation Research Part A:
Policy and Practice, 37(3), pp. 257โ275. doi: 10.1016/S0965-8564(02)00015-0.
Vaz, A. I. F. and Vicente, L. N. (2007) โA particle swarm pattern search method for bound
constrained global optimizationโ, Journal of Global Optimization, 39(2), pp. 197โ219. doi:
10.1007/s10898-007-9133-5.
Wolshon, B. (2001) โโโOne-Way-Outโโ: Contraflow Freeway Operation for Hurricane
Evacuationโ, Natural Hazards Review, 2(3), pp. 105โ112.
Wolshon, B. (2006) โEvacuation planning and engineering for Hurricane Katrinaโ, The
Bridge, 36(1), pp. 27โ34.
Xie, C., Lin, D. Y. and Travis Waller, S. (2010) โA dynamic evacuation network optimization
problem with lane reversal and crossing elimination strategiesโ, Transportation Research
Part E: Logistics and Transportation Review. Elsevier Ltd, 46(3), pp. 295โ316. doi:
10.1016/j.tre.2009.11.004.
Xie, J. et al. (2014) โNear optimal allocation strategy for making a staged evacuation plan
with multiple exitsโ, Annals of GIS. Taylor & Francis, 20(3), pp. 159โ168. doi:
10.1080/19475683.2014.942363.
Yen, J. Y. (1971) โFinding the K Shortest Loopless Paths in a Networkโ, Management
Science, 17(11), pp. 712โ716. doi: 10.1287/mnsc.17.11.712.
Yperman, I., Logghe, S. and Immers, B. (2005) โThe Link Transmission Modelโฏ: an Efficient
Implementation of the Kinematic Wave Theory in Traffic Networksโ, in Proceedings of the
10th EWGT Meeting and 16th Mini-Euro Conference, pp. 122โ127.
Yun, N. Y. and Hamada, M. (2015) โEvacuation behavior and fatality rate during the 2011
Tohoku-oki earthquake and tsunamiโ, Earthquake Spectra, 31(3), pp. 1237โ1265. doi:
10.1193/082013EQS234M.
Zeigler, D. J. et al. (1981) โEvacuation from a Nuclear Technological Disasterโ, American
Geographical Society, 71(1), pp. 1โ16.
Zheng, H. et al. (2010) โModeling of Evacuation and Background Traffic for Optimal Zone-
Based Vehicle Evacuation Strategyโ, Transportation Research Record: Journal of the
Transportation Research Board, 2196(1), pp. 65โ74. doi: 10.3141/2196-07.
Zhou, X. and Taylor, J. (2014) โDTAlite: A queue-based mesoscopic traffic simulator for fast
model evaluation and calibrationโ, Cogent Engineering. Cogent, 1(1), pp. 1โ19. doi:
10.1080/23311916.2014.961345.
Ziliaskopoulos, A. K. (2000) โA Linear Programming Model for the Single Destination
System Optimum Dynamic Traffic Assignment Problemโ, Transportation Science, (October
2016), pp. 36โ49.
Appendix
We provide below the network configurations that were introduced in Section 4.1 for the
evaluation of the formulation in larger problem instances.
Networks
a) Network A.
b) Network B.
c) Network C.
d) Network D.
Figure A Sample networks for linear and bi-level algorithm testing. Source and sink nodes are marked with a square and
dashed circle, respectively.