Heuristics for Scheduling Evacuation Operations in Case of Natural
Disaster
Kaouthar Deghdak, Vincent T’kindt and Jean-Louis Bouquard
Universit
´
e Franc¸ois–Rabelais Tours, Laboratoire d’Informatique (EA 6300),
Equipe Ordonnancement et Conduite (ERL CNRS 6305), 64 Avenue Jean Portalis, 37200 Tours, France
Keywords:
Evacuation, Greedy Heuristic, Matheuristic.
Abstract:
In this paper, we consider a large-scale evacuation problem after an important disaster. The evacuation is
assumed to be done by means of a fleet of buses, thus leading to schedule the evacuation operations by buses
(Bus Evacuation Problem, BEP). We propose time indexed formulations, as well as heuristic algorithms like
greedy algorithms and a matheuristic. This matheuristic uses the former formulation to improve the best solu-
tion obtained by the greedy heuristics. In computational experiments, we analyse and evaluate the efficiency
of the proposed solution algorihms.
1 INTRODUCTION
After an important disaster, the evacuation of peo-
ple from the damaged area to a safety area be-
comes necessary. From the literature, it turns out
that two main application scenarios have been con-
sidered: the scenario of building evacuation and the
scenario of region evacuation. Evacuation of build-
ings is extensively discussed in (Chalmet et al., 1982)
and (Hamacher and Tjandra, 2001). Several reviews
tackle the evacuation problem for large regions, e.g,
(Sattayhatewa and Ran, 2000), (Church and Sexton,
2002), (Sbayti and Mahmassani, 2006) and (Bish
et al., 2013).
The evacuation models proposed in the lit-
erature can be classified into two categories:
simulation-based models and optimization-based
models. Simulation-based models are developed to
define, analyse and evaluate evacuation plans, e.g,
(Sheffi et al., 1982) and (Kwon and Pitt., 2005).
These models are generaly based on the Dynamic
Traffic Assignment (DTA) methodology, e.g, (Zil-
iaskopoulos., 2000), (Mahmassani, 2001) and (Sbayti
and Mahmassani, 2006). Optimization-based mod-
els are mainly based on dynamic network flow mod-
els, like universal maximum flow, minimum dynamic
cost flow, maximum dynamic flow, quickest flow and
quickest path, e.g, (Yamada, 1996) and (Hamacher
and Tjandra, 2001). The goal of the evacuation prob-
lems is to find the optimal evacuation plan for an
emergency situation. A proper evacuation plan is im-
perative to reduce fatalities. To solve these problems,
numerous optimization models have been proposed,
in which we find three objective functions:
1. Minimizing the Total Evacuation Time (TET):
TET is the period during which the evacuees will
be exposed to risk until reaching a safe destina-
tion.
2. Minimizing the Clearance Time (CT): CT is the
time until the last vehicle or evacuee leaves the
damaged area.
3. Maximizing the number of people exiting the
damaged area in a given time period.
The first objective function is equivalent to the
second. They are both mentionned because they
are presented in differenent ways in the literature.
In some cases, the third objective function becomes
more important than the first two. For instance, when
the total evacuation time or the clearance time are
fixed in advance by the authorities.
(Bish, 2011) introduced and studied a new model
for bus-based evacuation planning. The choice of
buses as a transportation mean has been motivated by
the fact that bus-based evacuation is logistically com-
plex, expensive, produced unacceptable levels of con-
gestion and more dangerous than the bus-based evac-
uation should be.
The work dealt with in this paper follows the same
line. We consider the evacuation of people due to
a naturel disaster like earthquakes, where evacuees
have to change their centre of lives from several days
294
Deghdak K., T’kindt V. and Bouquard J..
Heuristics for Scheduling Evacuation Operations in Case of Natural Disaster.
DOI: 10.5220/0004922102940300
In Proceedings of the 3rd International Conference on Operations Research and Enterprise Systems (ICORES-2014), pages 294-300
ISBN: 978-989-758-017-8
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
to several months with the eventual goal of returning
back to their respective home. In particular, we as-
sume that the locations of gathering points (i.e where
people are evacuated, outside the damaged area), the
locations of collection points (i.e where people are
gathered waiting to be evacuated) and the capacities
of the transportation network are known. The goal
is to define a macroscopic plan of evacuation which
means that people are considered homogeneously, i.e.
the evacuees are assumed have the same behaviour
and have to be transported from the collection points
to the gathering points in a minimal amount of time.
The evacuation is done by means of a set of homoge-
neous buses.
The remainder is organized as follows. In Section
2, we describe the Bus Evacuation Problem in details
and provide a mixed-integer programming formula-
tion. In Section 3, we present greedy heuristics for
the problem, and an iterative local search heuristic, re-
ferred to as matheuristic, in Section 4. Computational
results are presented in Section 5. Finally, Section 6
concludes the paper.
2 PROBLEM DEFINITION AND
MATHEMATICAL
FORMULATION
In this section, we first describe the Bus Evacuation
Problem (BEP) and next provide a time-indexed for-
mulation. This formulation is used in the following to
evalute the quality of the proposed heuristics and to
develop a matheuristic algorithm.
2.1 Problem Definition
Consider a network (N , A), where N and A respec-
tively denote the set of nodes and edges. N is com-
posed of two subsets of nodes: CP and GP. CP is
a set of collection points where evacuees are initially
located and GP is a set of gathering points. An edge
(i, j) A exists iff evacuees can be transported from
collection point i to gathering point j. Each collection
point i has a demand e
i
i CP, which is the number of
buses required to evacuate all people at point i. Each
gathering point j has a capacity cap
j
, j GP, wich is
the number of evacuees that can be brought to j: it is
expressed as a number of buses. Each edge (i, j) has a
non-negative length p
i, j,t
, which is the travel time be-
tween each couple (CP
i
,GP
j
) and it is time-dependent.
This means that it can be increased or decreased, over
time, depending on the state of the network. This is
a consequence of the evolution of the transportation
network through time due to events like earthquake
replicates, road repairs, roads’ congestion.... We as-
sume that events that modify the transportation net-
work, may occur: we consider that such k events hap-
pen at a known time d
l
and can change the value of
travel times between a collection point i towards gath-
ering point j:
p
i, j,t
=
a
0
i, j
if t
i j
]0,d
1
]
a
1
i, j
if t
i j
]d
1
,d
2
]
.
.
.
a
k1
i, j
if t
i j
]d
k1
,d
k
]
where a
x
i, j
is the travel time if the evacuees are
transported from a collection point i to gathering point
j in the x
th
time interval, i.e. it’s starting time t
i j
]d
x1
,d
x
]. The number of finite intervals [d
l1
,d
l
] is
determined by a preliminary forecasting of the evolu-
tion of the transportation network.
A fleet of K identical buses is given and used to
evacuate people. The problem we consider is to find
a schedule such that all evacuees are transported from
the collection points CP to the gathering points GP,
minimizing the total evacuation time denoted by C
max
.
The Figure 1 depicts an example of a bus evacu-
ation network at time t, with three collection points,
and two gathering points. The number of buses re-
quired for evacuees in the collection points are 2, 4
and 5, while the gathering points’ capacities are 6 and
9 buses of evacuees.
GP1
CP1
CP2
CP3
GP2
6
9
2
4
5
7
5
9
10
6
2
Figure 1: Exemple of BEP network at a given time t.
2.2 A Time-indexed Formulation
The Bus Evacuation Problem can be modeled as a cu-
mulative scheduling problem with additional resource
HeuristicsforSchedulingEvacuationOperationsinCaseofNaturalDisaster
295
constraints: the gathering points are cumulative par-
allel machines and the jobs to schedule are people
to evacuate. The additional ressources are the buses
which are disjunctive resources. As mentioned earlier
in Section 2.1, an important feature of our evacua-
tion scheduling problem is that the processing times
(i.e.travelling times) of the jobs are dependent on the
starting times of the job in the schedule.
Consider that we have M parallel and cumulative
machines, each machine j corresponding to a gath-
ering location GP
j
, with a capacity cap
j
expressed
as a number of buses that can bring evacuees. N =
iCP
e
i
evacuation operations have to be scheduled
on the machines, and each operation i is associated to
a collection point CP
`
. Similarly to the calculation of
the machine’s capacities, it is assumed that one evacu-
ation operation coressponds to the evacuation of peo-
ple by a full bus. We suppose that the capacity of the
reconfigured network is sufficient after the disaster to
enable a fluent evacuation. This hypothesis enables
to approximate the bus routting by considering only
travel times.
To model the Bus Evacuation Problem we have
proposed a time-indexed mathematical formulation.
Usually, time-indexed formulations on scheduling
problems yield simple and efficient models despite
the presence of a pseudo-polynomial number of vari-
ables (see (Berghman et al., 2011) among others).
Throughout the paper, we make use of the notation
[S] := {1,...,S}. Let us turn to the model for our
evacuation scheduling problem in which T is the time
horizon. The decision variables are:
i [N], j [M],t [T ];
x
i, j,t
=
1 if a bus starts the evacuation operation
i towards j at [t,t + 1[.
0 otherwise.
and C
max
, the duration of the schedule.
The proposed (IP) formulation is as follows:
minC
max
. (1)
Subject to:
C
max
(t + p
i, j,t
)x
i, j,t
i [N], j [M],t [T ] (2)
t[T ]
i[N]
x
i, j,t
cap
j
j [M] (3)
i[N]
j[M]
t
0
[0,t]
p
i, j,t
0
+t
0
>t
(4)
x
i, j,t
0
K t [T ] (5)
t[T ]
j[M]
x
i, j,t
= 1 i [N] (6)
x
i, j,t
{0, 1} (7)
C
max
N (8)
Constraints (2) define the value of criterion C
max
,
which is then minimized by the objective function (1).
Constraints (3) are the gathering point capacity con-
straints: we cannot exceed the capacities of gather-
ing points. Constraints (5) are the bus capacity con-
straints : we cannot exceed the number of buses we
have. Constraints (6) ensure that each operation is
processed once and only once. Constraints (7) and
(8) are the logical binary and non-negative integer re-
strictions on the x
i, j,t
and C
max
variables, respectively.
Notice that when the processing times are not
time-dependant and the capacities of the gathering
points are sufficiently large enough, then the BEP
reduces to the identical parallel machine scheduling
problem with makespan minimization. The latter
being strongly N P -hard (see (Garey and Johnson,
1979)), we can deduce that BEP is so.
3 GREEDY HEURISTICS
In this section, we describe several greedy heuris-
tic algorithms to solve the Bus Evacuation Problem.
These heuristics aim to compute a solution in a very
fast way. We outline the need for such heuristic proce-
dures to solve real-size instances which are very large.
Figure 2 presents the general algorithm of the pro-
posed greedy heuristic, which takes upon entry a se-
quencing rule R . Let oprlist be the set of the oper-
ations that will be assigned on machines, sortlist(i,j)
be the set of sorted operations according to the rule
R at a given time t, and C
m
max
be the makespan of a
machine m. The greedy heuristic’s solution is stored
in schedulelist. Each operation added in schedulelist
will be deleted from oprlist and sortlist(i,j).
The function event, presented in the pseudo code
of the greedy algorithm (Figure 2), checks if there
is an event requiring to recalculate the priority list
sortlist. Each rule R has a specific event function.
In the following, we have tested eight greedy
heuristic versions.
1. Version 1: this version uses the rule R
1
. To each
operation opr
i
, let tno
i
be the total number of op-
erations of the same job j than opr
i
. Then, the
operations are sorted by non-decreasing order of
their value tno
i
. The function event always re-
turns false because the tno
i
of each operation opr
i
does not change throughout the algorithm.
ICORES2014-InternationalConferenceonOperationsResearchandEnterpriseSystems
296
Generic Greedy Algorithm (R )
Input: /* A sorting rule R */
schedulelist
/
0, sortlist
/
0,
oprlist {opr
1
,opr
2
,...., opr
N
};
t 0, k |K|.
Arrange operations in the sortlist(i,j) using the rule R .
Repeat
Add the k first operations in schedulelist (i,j,t);
Delete k operations added in schedulelist from
oprlist and sortlist ;
if t=0 then k 0;
Let m [M] be a machine with the minimum C
max
;
t C
m
max
;
k k + 1;
if event() then
sortlist
/
0 ;
Arrange operations in the sortlist(i,j) using
the rule R ;
end if
Until |oprlist| = 0 and t > T
return Feasible solution schedulelist .
Figure 2: Greedy-Heuristic algorithm.
2. Version 2: this version uses the rule R
2
. In this
rule, operations opr
i
are sorted by decreasing or-
der of their value tno
i
. The function event always
returns false because the tno
i
of each operation
opr
i
does not change throughout the algorithm.
3. Version 3: this version uses the rule R
3
. The oper-
ations are sorted according to the shortest process-
ing time first rule (SPT) known in scheduling the-
ory (Smith, 1956). This rule has been adapted to
take the time-dependency of the processing time
p
i, j,t
into acount. The function event returns true
if there are modifications in the values of opera-
tions’ processing time.
4. Version 4: this version uses the rule R
4
. The oper-
ations are sorted in non-decreasing order of their
ratio
nos
j
tno
j
, with nos
j
the number of operations of a
job j already scheduled and tno
j
the total number
of operations of a job j. If several operations have
the same priority, we break ties by means of the
SPT rule. The function event returns true if there
is, at least, an operation opr
k
that has a smaller
ratio than the first operation in sortlist.
5. Version 5: this version uses the rule R
5
. In this
rule, the operations are sorted in non-decreasing
order of their ratio
tno
j
nto
nos
j
tno
j
, with nto is the total
number of operations. The function event returns
true if there is, at least, an operation opr
k
that has
a smaller ratio than the first operation in sortlist.
6. Version 6: this version uses the rule R
6
. To each
operation opr
i
, let nuo
i
be the number of currently
unscheduled operations of the same job j than
opr
i
. Then, the operations are sorted by decreas-
ing order of their value nuo
i
. The function event
returns true if there is an operation opr
k
that has
a larger number of currently unscheduled opera-
tions than the first operation in sortlist.
7. Version 7: this version uses the rule R
7
. To each
operation opr
i
, let nuo
i
be the number of cur-
rently unscheduled operations of the same job j
than opr
i
. Then, the operations are sorted in non-
decreasing order of their value nuo
i
. The func-
tion event returns true if there is an operation opr
k
that has a smaller number of currently unsched-
uled operations than the first operation in sortlist.
8. Version 8: this version uses the rule R
8
. In this
rule, we randomly assign a priority for each job.
The function event returns true if there are modi-
fications in the values of jobs’ priority.
4 A MATHEURISTIC
In this section, we propose a local search method
called matheuristic. The general idea of matheuristics
is to exploit the strength of both metaheuristic algo-
rithms and exact methods as well, leading to a hybrid
approach (Della Croce et al., 2011). It is a heuristic
based on mathematical programming.
We use the time-indexed formulation introduced
in Section 2.2 to construct a matheuristic for
BEP. Let be an initial heuristic solution given as
the best solution obtained by the greedy heuris-
tics. The matheuristic tries to improve that solu-
tion by exploring its neighborhood as follows. Let
be a feasible schedule (heuristic solution) ¯x =<
¯x
i, j,t
,i [N], j [M],t [T ] >, where ¯x
i, j,t
= 1, if op-
eration i is processed on machine j at time t. We de-
fine a neighberhood N ( ¯x,r,h) by choosing a date r
in the schedule and a size parameter h. Let
˜
S(r,h) be
the index set of the operations scheduled in the time
interval [r, r + h[. We call such a subset of operations
an ”operation-window”.
The best solution in the neighberhood N ( ¯x, r,h) is
computed by minimizing the makespan C
w
max
, subject
to (3)-(8) and by adding the following constraint:
x
i, j,t
= ¯x
i, j,t
i /
˜
S(r,h), j [M],t [0,r[ (9)
We call this reduced minimization problem the
window reoptimization problem, and it is solved by
a mathematical solver like CPLEX. The additional
constraints (9) forces the changes to occur within the
HeuristicsforSchedulingEvacuationOperationsinCaseofNaturalDisaster
297
operation-window. If we have an improvement in
C
w
max
, then, in the new solution ˜x, all the operations
which started after the time r + h in the initial solu-
tion ¯x will be left time shifted keeping their previous
positions and respecting the model constraints.
If no improved solution is found a new operation-
window (i.e. new value of r) is selected to be opti-
mized until all possible windows have been selected.
The search is stopped if no window reoptimization
problem has an optimal solution which improves the
current solution or if a predefined time limit is ex-
ceeded.
The algorithm of the matheuristic is given in Fig-
ure 3:
The Matheuristic Algorithm
Input:
¯x =<heuristic solution calculated by greedy algoritthm>.
vect =<vector of starting times of all operations in ¯x >.
Repeat
Set improved f alse;
Set i 0;
Repeat
r vect[i] ;
˜x Neighborhood( ¯x,r, h)
if (C
max
( ¯x) > C
max
( ˜x)) then
¯x ˜x
improved true
For r r + h to C
max
r + h do
¯y Neighborhood( ¯x,r,h)
if (C
max
( ¯x) > C
max
( ¯y)) then
¯x ¯y
end if
end For
recompute vect from ¯x
end if
Until i |vect| or improved or time limit expired
Until not improved or time limit expired
Return Feasible solution ¯x.
The Neighborhood Procedure
Calculate
˜
S(r,h);
Reschedule operations in tha time window [r,r + h[ by
minimizing C
w
max
( ¯x) subject to (3)-(8), and (9);
Let ˜y be the solution obtained;
Return solution ˜y.
Figure 3: Matheuristic algorithm.
5 COMPUTATIONAL
EXPERIMENTS
In this section, we focus on the experimental evalua-
tions of the (IP) solution, and the heuristic algorithms.
We first describe how the experimentation have been
setup.
Environment. All experiments were run on a com-
puter with a 8-core Intel processor, running at 2.60
GHz with 20MB cache, 8 GB RAM and Windows 7.
We wrote our code in C++ , and used the commercial
IP solver CPLEX v.12.2. CPLEX was pinned to one
core for the solution of the time-indexed formulation.
Dataset. This work is part research project, called
DSS-Evac-Logistic, partially granted by the french
research agency ANR. In this project, we consider the
real-world instance of Nice (France), as a case study.
So, the data sets are randomly generated in such a way
that we always have a feasible and realistic instances
for Nice city. The number of machines takes values
M ∈{2, 4, 6, 8, 10} and the capacity cap
j
of each ma-
chine draws at random from {20,21,22,...,40}. The
number of jobs takes values L ∈{10, 20, 30, 40} and
the number e
l
of operations of job l are generated
randomly from
h
1
4
0.9
j
cap
j
L
;
7
4
0.9
j
cap
j
L
i
. This gener-
ation ensures that the total operations’ number is less
than the total machines’ capacity. We assume that the
time period of evacuation is [D, D+1[, when D is the
day of disaster. In the following, we take the time dis-
cretisation as a quarter of an hour, which implies that
T is equal to 192 quarters.
We assume that we have 6 events happening
and changing the value of the operations’ process-
ing time. To do so, we generated two degradation
dates: T
D
1
,T
D
2
[0,192[, common to all operations,
and three improvement dates: T
A
1
U[0,T
D
1
],T
A
2
U]T
D
1
,T
D
2
], T
A
3
U]T
D
2
,192[, specific to each oper-
ation. Operations’ processing time are drawn ran-
domly from {2, 3, 4, ..., 8}. Finally, the number of
buses K =
e
l
¯
P
180
, where ¯p is the average processing
time, which ensures that we have enough buses to do
the evacuation in 180 quarters of an hour.
For each instance, we run the following algo-
rithms:
Exact Solution Algorithm: we have tested the so-
lution of the (IP) model by Cplex solver. In the
remainder, this solution algorithm will be denoted
by IP. Notice that in this model we use the best
C
max
value found by the heuristics, to set the value
of the time horizon T. Besides, a time limit of
1800 seconds and a memory limit of 1 GB have
been given to CPLEX solver.
Heuristic Algorithms:
We have tested the eight versions of the greedy
heuristic, and for each data set we keep the best
solution found.
We have tested the matheuristic with a time
limit of 600 secondes. Previous preliminary ex-
perimentations have shown that the best results
ICORES2014-InternationalConferenceonOperationsResearchandEnterpriseSystems
298
are obtained with the window size h = 25. The
initial solution is the best solution found by the
greedy heuristics
We solve the time-indexed formulation with a
time limit equal to that of the matheuristic. This
heuristic is referred to as IPheuristic.
Table 1 presents the runnig times and the number
of nodes explored by CPLEX when solving the IP.
Column #oprs presents the number of operations, col-
umn #mach presents the number of machines, column
#bus presents the number of buses, column #inst tot
presents the number of instances for each problem
size, column #inst opt presents the number of in-
stances that have been solved to optimality by IP. No-
tice that the instances are randomly generated based
on a number of jobs, and a number of machines.
However, the size of the instances solved by IP de-
pends on the total number of operations, i.e.
l
e
l
.
consequently, after having generated instances we
have gathered them in classes of ”equivalent size in-
stances”, but in term of number of operations, num-
ber of machines and number of buses. Columns IP
(time) report the minimum, the average and the max-
imum running times of CPLEX when solving IP. IP
(nodes) columns report the minimum, the average and
the maximum explored nodes. As the results in Ta-
ble 1 illustrate, for almost all the instances with less
than 60 operations the IP find the optimal solutions.
Unfortunately, when the number of operations varing
between 90 and 110, IP can not solve the instances to
optimality.
Table 1: Running times and number of explored nodes.
#oprs #mach #bus #inst tot #inst opt
IP (time(s)) IP (nodes)
min avg max min avg max
[20,35[ 2 [1,2] 32 32 15 23,72 87 0 217,7 901
[35,40[ 2 2 17 16 14 63,06 200 0 243,5 784
[40,60[ 2 [2,3] 31 30 17 106,8 384 0 367,7 4251
[40,60[ 4 [2,3] 6 5 85 323 859 0 207,8 602
[60,70[ 4 3 13 6 175 921,2 1710 0 359,5 872
[70,80[ 4 [3,4] 17 1 1220 1220 1220 1015 1015 1015
[80,90[ 4 4 26 1 610 610 610 147 147 147
[90,110[ 4 [4,5] 18 0 / / / / / /
In the following, we investigate the quality of the
proposed heuristics.
Table 2 displays the results of IPheuristic com-
pared to the best heuristic solutions delivred by
IPheuristic or matheuristic solutions (the matheuris-
tic is at least as good than the greedy heuristic) and its
running time. Let dev(A) be the deviation associated
to a heuristic A. We have:
dev(A) =
C
max
(A) min(C
max
(A),C
max
(bestheuristic))
min(C
max
(A),C
max
(bestheuristic))
100
IPheuristic performs well when the number of op-
erations is less than 40 and requires less than 600 sec-
ondes to provide the best solutions. For the instances,
for which the number of operations belongs to [40,
70[ we observe a deterioration of the quality of this
heuristic. For instance, in some cases, the IPheuris-
tic solution is far from the best heuristic solution with
a deviation of 95%. Furthermore, we found that for
larger instances, IPheuristic becomes less competi-
tive, and fails to produce good solutions whitin a time-
limit of 600 secondes.
Table 2: Evaluation of the IPheuristic.
#oprs #mach #bus
dev% CPU time (s)
min avg max min avg max
[20,35[ 2 [1,2] 0 0 0 17 49 101
[35,40[ 2 2 0 0 0 98 217,8 600
[40,60[ 2 [2,3] 0 0 0 53 277,8 600
[40,60[ 4 [2,3] 0 1.35 42,06 575 597 600
[60,70[ 4 3 0 1 3,84 600 600 600
[70,80[ 4 [3,4] 38,64 71,24 115,25 600 600 600
[80,90[ 4 4 26,42 64,37 105,1 600 600 600
[90,110[ 4 [4,5] 23,77 58,73 95.35 600 600 600
The results presented in Table 3 show the qual-
ity of the solutions obtained by the greedy heuristics
against the best heuristics solutions. All the proposed
priority rules are tested on the small and large in-
stances, and we have found that the greedy heuristics
which use the rules R
3
and R
4
outperform the other
rules. We note that running these heuristics takes a
negligible amount of time.
Table 3: Evaluation of the greedy heuristics.
#oprs #mach #bus
dev %
min avg max
[20,35[ 2 [1,2] 5,26 12,89 27,42
[35,40[ 2 2 7,5 14,58 25,29
[40,60[ 2 [2,3] 9 15,85 30,84
[40,60[ 4 [2,3] 4,34 8,91 13,04
[60,70[ 4 3 0 2,45 12,35
[70,80[ 4 [3,4] 0 0,72 3,15
[80,90[ 4 4 0 0,92 6,94
[90,110[ 4 [4,5] 0 1,57 9,57
Table 4 provides the deviation of the matheuristic
against the best heuristic solution. For instances with
less than 70 operations in size, the matheuristic suc-
ceeds in providing the best solution on some instances
since the minimal deviation value is 0. For instances
with more than 60 operations in size, the matheuristic
outperforms, on the average, the IPheuristic. For the
instances with a number of operations in [60, 70[, the
matheuristic provides the best solution for 12 out of
13 instances. When the number of operations belongs
to [70, 80[, the best solutions are always delivred by
the matheuristic within a time limit of 600 secondes.
To summarize all the results introduced so far, we
can conclude that the matheuristic is the best heuris-
tic algorithm, even if for large instances the time limit
must be increased to enable improvement over the
greedy heuristics. Noteworthy, these ones seem to be
a good compromise between quality and CPU time,
especially for larger instances.
HeuristicsforSchedulingEvacuationOperationsinCaseofNaturalDisaster
299
Table 4: Evaluation of the matheuristic.
#oprs #mach #bus
dev % CPU time (s)
min avg max min avg max
[20,35[ 2 [1,2] 0 5,2 10,29 32 157,8 370
[35,40[ 2 2 0 8.6 24,14 97 326,2 603
[40,60[ 2 [2,3] 0 7,97 20 166 399,7 600
[40,60[ 4 [2,3] 0 3.57 11,11 600 600 600
[60,70[ 4 3 0 0,38 5,05 600 600 600
[70,80[ 4 [3,4] 0 0 0 600 600 600
[80,90[ 4 4 0 0 0 600 600 600
[90,110[ 4 [4,5] 0 0 0 600 600 600
6 CONCLUSIONS
In this work, we have studied the problem of schedul-
ing evacuation operations that we have called the
bus evacuation problem (BEP). We have provided
a time-indexed formulation for this problem. Next,
we have provided three heuristics, the first one is a
set of greedy heuristics. The second one is a local
search called matheuristic based on the mathemati-
cal formulation provided, which improves the best
geedy heuristic solutions. The third one is provided
by CPLEX within an imposed time limit. During the
conference, we will present all the computationnal re-
sults for small and larger instances, i.e. instances with
6, 8, 10, 20, 30 machines and more than 110 opera-
tions.
ACKNOWLEDGEMENTS
This research has been supported by ANR-11-SECU-
002-01, project DSS EVAC LOGISTIQUE (CSOSG
2011).
REFERENCES
Berghman, L., Leus, R., and Spieksma (2011). Optimal
solutions for a dock assignment problem with trailer
transportation. Annals of Operations Research.
Bish, D. (2011). Planning for a bus-based evacuation. OR
Spectrum, 33(3):629–654.
Bish, D. R., Sherali, H. D., and Hobeika, A. G. (2013). Op-
timal evacuation planning using staging and routing.
J Oper Res Soc, pages –.
Chalmet, L., Francis, R., and Saunders, P. (1982). Net-
work models for building evacuation. Fire Technol-
ogy, 18(1):90–113.
Church, R. and Sexton, R. (2002). Modeling Small Area
Evacuation: Can Existing Transportation Infrastruc-
ture Impede Public Safety? University of California
Santa Barbara, National Center for Geographic Infor-
mation and Analysis, Vehicle Intelligence and Trans-
portation Analysis Laboratory.
Della Croce, F., Grosso, A., and Salassa, F. (2011).
A matheuristic approach for the total completion
time two-machines permutation flow shop problem.
6622:38–47.
Garey, M. R. and Johnson, D. S. (1979). Computers
and Intractability: A Guide to the Theory of NP-
Completeness. W. H. Freeman & Co, New York, NY,
USA.
Hamacher, H. W. and Tjandra, S. A. (2001). Mathematical
modeling of evacuation problems: A state of the art. In
Pedestrian and Evacuation Dynamics (Schreckinberg,
M. and Sharma, S. D. eds), 1964:227–266.
Kwon and Pitt. (2005). Evaluation of emergency evacua-
tion strategies for downtown event traffic using a dy-
namic network model. Journal of the Transportation
Research Board, 1922:149155.
Mahmassani, H. S. (2001). Dynamic network traffic assign-
ment and simulation methodology in advanced system
management applications. Networks and Spatial Eco-
nomics, 1:67.
Sattayhatewa, P. and Ran, B. (2000). Developing A Dy-
namic Traffic Management Model For Nuclear Power
Plant Evacuation. Transportation Research Board,
79th Annual Meeting.
Sbayti, H. and Mahmassani, H. S. (2006). Optimal schedul-
ing of evacuation operations. Journal of the Trans-
portation Reasearch Board, 1964:238–246.
Sheffi, Y., Mahmassani, H., and Powell., W. B. (1982). A
transportation network evacuation model. Transporta-
tion Research Part A, 16(3):209–218.
Smith, W. (1956). Various optimizers for single-stage pro-
duction. 3(1-2):59–66.
Yamada, T. (1996). A network flow approach to a city emer-
gency evacuation planning. International Journal of
Systems Science, 27(10):931–936.
Ziliaskopoulos., A. K. (2000). A linear programming model
for the single destination system optimum dynamic
traffic assignment problem. Transportation Science,
34:37.
ICORES2014-InternationalConferenceonOperationsResearchandEnterpriseSystems
300