Available online at www.sciencedirect.

com

Electronic Notes in Discrete Mathematics 52 (2016) 317–324 www.elsevier.com/locate/endm

Work-troop scheduling for road network

accessibility after a major earthquake

Celso Satoshi Sakuraba

Departamento de Engenharia de Produção, Universidade Federal de Sergipe
Av. Marechal Rondon S/N, Jardim Rosa Elze 49100-000 São Cristóvão, Brazil

Andréa Cynthia Santos 1 Christian Prins

ICD-LOSI, UMR CNRS 6281, Université de Technologie de Troyes
12 rue Marie Curie, CS 42060, 10004 Troyes Cedex, France

Abstract
The road network accessibility is an important issue for earthquake relief operations,
since several roads may be damaged obstructing the access to certain areas. This
work proposes a mathematical model and two heuristics for the road repairing
work-troops scheduling in order to increase accessibility to the population as fast
as possible after a major earthquake. Solutions for randomly generated instances
given by the model are used to evaluate the heuristics’ performance. The heuristics
are tested on a graph with more than ten thousand vertices and edges from Port-
au-Prince 2010 earthquake in Haiti.
Keywords: post-disaster relief, network rehabilitation, multiflow formulation,
heuristics.

1
Email: andrea.duhamel@utt.fr
This work is part of the OLIC (Optimisation de la Logistique d’Intervention pour les
Catastrophes majeures) project, funded by the Conseil Supérieur de la Formation et de
la Recherche Stratégiques (CSFRS), France.

http://dx.doi.org/10.1016/j.endm.2016.03.042
1571-0653/© 2016 Elsevier B.V. All rights reserved.
318 C.S. Sakuraba et al. / Electronic Notes in Discrete Mathematics 52 (2016) 317–324

1 Introduction
Although resources like helicopters are also deployed on post-disaster logistics
after major earthquakes, the majority of the population receives assistance and
supplies from road relief teams. Thus, road network accessibility becomes an
important issue, since a large amount of the city structures might be strongly
affected, with rubble blocking several roads. In such situations, it is necessary
to elaborate a plan for Work-Troops (WT), composed of bulldozers, excava-
tors and dump trucks, to clean and repair the roads, allowing relief teams to
reach the maximum amount of the population in a minimum time. In this
context, the Work-troops Scheduling Problem (WSP) consists of defining a
road repairing plan for WT per time period to increase accessibility as fast as
possible to places where the population gather after an earthquake.
Humanitarian logistics and emergency management have been considered
in several papers as shown in the review presented by [2]. A static version of
the problem, where the order in which edges are repaired is not considered, is
presented in [4]. A max-flow based model to minimize the weighted sum of the
shortest times to reach each destination is proposed, together with a GRASP
combined with VND. A dynamic version of the problem is proposed by [1],
where two mathematical models are introduced to determine a restoration
schedule and the WT allocation among origins. The model requires a pre-
calculation of shortest paths (SP), and is applied to two regions of Turkey.
Another dynamic representation of the problem is introduced by [5,6] using
a time-space network where each node represents an origin, intersection or
damaged area at a certain time. The highway repair presented in [5] considers
the rescheduling of WT due to changes in the demand and offer of repairs. A
hybrid global search based on ant colony metaheuristic is proposed. A model
to minimize the maximum time necessary to both restore all damaged roads
and distribute all the relief is presented in [6].
In this study, the WSP is formalized by means of a multi-flow mathematical
model. We proposed two simple heuristics to solve real size instances with
over ten thousand vertices and edges in a reasonable amount of time. Tests
are performed on simulated instances and on the graph of the Port-au-Prince
earthquake, 2010 (See Figure (1)).

2 Problem definition
The WSP is defined in a graph G = (V, E), where V is the set of vertices
(|V | = n) and E the set of edges (|E| = m). Let O ⊂ V be the subset
 C.S. Sakuraba et al. / Electronic Notes in Discrete Mathematics 52 (2016) 317–324 319

Fig. 1. Port-au-Prince road network after the 2010 earthquake. Circles, triangles
and dark edges represent origins, destinations and blocked roads, respectively.

of origins, where WT are available initially, and D ⊂ V be the subset of

destinations corresponding to population gathering points (O ∩ D = ). Each
origin has a number qi ∈ N∗ of available WT and each destination has a
population pi ∈ N∗ .
Let us consider B ⊂ E the subset of blocked edges (roads to be repaired).
Each edge [i, j], with i < j, is associated with a set of three parameters
described as follows. The distance dij ∈ R∗+ is the length of the edge. The
repair rij ∈ N represents the amount of working periods necessary for WT to
repair that edge. For example, if rij = 4 the edge is repaired after one WT
spends four time periods working on it, or two WT spend two time periods
each. The maximum number of WT that can work simultaneously at each
extremity of edge [i, j] due to physical limitations imposed by the road width
is given by an upper bound uij ∈ N∗ . We define T as the number of time
periods considered by the model, after which all edges are repaired.
The mathematical formulation (1) to (19) considers the dynamic aspect of
the problem, i.e., the network evolution along time. Then, SP to destinations
320 C.S. Sakuraba et al. / Electronic Notes in Discrete Mathematics 52 (2016) 317–324

and blocked edges accessibility to WT are computed at each time period. The
model has two main components per time period: a flow from an origin to each
destination through its SP and a WT flow from the origins to an extremity of
a blocked edge. Both flows are defined in a directed network obtained from G
by replacing each edge [i, j] ∈ E by two arcs (i, j) and (j, i). We denote the set
of arcs as A and define B  ⊂ A as the subset of all arcs replacing blocked edges
[i, j] ∈ B by blocked arcs (i, j) and (j, i). We consider that a WT flow cannot
pass through an arc (i, j) ∈ B  , while for the SP flow, dij is set to a large
integer number M . In addition, the SP flows are defined using an artificial
root vertex r, connected to all origins by arcs (r, i), with dri = rri = 0, ∀i ∈ O.
The model uses the following variables: at the end of each time period
t, the SP flow through arc (i, j) is represented by variable fijt , and variable
xtij = 1 if arc (i, j) is used on some SP, 0 otherwise. Besides, variable zijt = 1 if
edge [i, j] is available and 0 otherwise, and the variable sti represent the length
of the shortest path from the root r to a vertex i. During each time period
t, the WT flow on arc (i, j) is represented by variable yijt , while variable wij t

represent the number of WT entering arc (i, j) to repair it. Note that we can
have WT repairing edges from both extremities, so for each edge [i, j] we have
t t
wij and wji .

T

min Z = pi sti s.t. (1)
i∈D t=1
⎧
⎪
⎪ |D|, if i = r
⎪
⎨
 
fijt − t
fji = 0, ∀j ∈ V \D , t = 1 . . . , T (2)
⎪
⎪
i:(i,j)∈A i:(j,i)∈A ⎪
⎩ 1, ∀j ∈ D
fijt
xtij ≥ ∀(i, j) ∈ A, t = 1, . . . , T (3)
|D|
xtij ≤ fijt ∀(i, j) ∈ A, t = 1, . . . , T (4)
 
t t
yij + wik ≤ qi ∀i ∈ O, t = 1, . . . , T (5)
j:(i,j)∈A k:(i,k)∈B 
  
t t t
yij − yji = wjk ∀j ∈ V \O, t = 1, . . . , T (6)
i:(i,j)∈A i:(j,i)∈A k:(j,k)∈B 

0
zij =0 (7)
[i,j]∈B

T
zij = |B| (8)
[i,j]∈B
 C.S. Sakuraba et al. / Electronic Notes in Discrete Mathematics 52 (2016) 317–324 321

t

t t  t 
rij zij ≤ (wij + wji ) ∀[i, j] ∈ E, t = 1, . . . , T (9)
t =1

t t t−1
yij + yji ≤ qi zij ∀[i, j] ∈ E, t = 1, . . . , T (10)
i∈O
t
wij ≤ uij ∀(i, j) ∈ A, t = 1, . . . , T (11)
sti − stj + (M + dij )xtij ≤M ∀i, j ∈ V |(i, j) ∈ A, t = 1, . . . , T (12)
stj ≥ M (xtij − zij
t
) ∀(i, j) ∈ A, t = 1, . . . , T (13)
t
fij ≥ 0 ∀(i, j) ∈ A, t = 1, . . . , T (14)
sti ≥ 0 ∀i ∈ V, t = 1, . . . , T (15)
t
wij ≥0 ∀(i, j) ∈ A, t = 1, . . . , T (16)
t
xij ∈ {0, 1} ∀(i, j) ∈ A, t = 1, . . . , T (17)
t
yij ≥0 ∀(i, j) ∈ A, t = 1, . . . , T (18)
t
zij ∈ {0, 1} ∀[i, j] ∈ E, t = 0, . . . , T (19)

The objective function (1) is the weighted sum of the SP distances to each des-
tination over all time periods. Equations (2) ensure there is a unitary flow from
the root r to each destination, while Inequalities (3) and (4) link variables fijt to
xtij . Constraints (5) and (6) represent the WT flow, limited to the number of WT
available at each origin qi and allocating them to a blocked edge. Equations (7) and
(8) determine the initial and final availability of blocked edges, while Inequalities
(9) set the edges availability according to the number of working periods spent on
repairing it until the present time period. Inequalities (10) make WT use only avail-
able edges to reach blocked edges. Inequalities (11) limit the number of WT on each
extremity of a blocked edge according to the road physical limitation. Inequalities
(12) compute the distances of the SP to each destination from its nearest origin, by
accumulating the distances of the SP from this origin to each intermediate vertex
in the path. Inequalities (13) make the distance of the SP to a vertex equal to M
if all the paths to that vertex are blocked. Variables are defined from (14) to (19).

3 Heuristics
We proposed two heuristics based on three steps: a preliminary allocation of WT
in the first period, a classification of blocked edges and the allocation of WT on
the remaining periods. The first step is the same for both heuristics: Dijkstra’s
algorithm [3] is applied to compute the SP from the root r to all destinations
considering edges’ repair values instead of their lengths. Then, all blocked paths
among the SP obtained by the algorithm are checked to decide if they can be
322 C.S. Sakuraba et al. / Electronic Notes in Discrete Mathematics 52 (2016) 317–324

rehabilitated in the first period, i.e., if they contain only one blocked edge and if
its upper bound allows enough WT to repair it immediately. If so, such paths are
repaired immediately by allocating a number rij of WT to the reachable extremity
of the blocked edge [i, j] in the path.
The heuristics distinguish from each other by the way blocked edges are classified
to be repaired after the initial step mentioned previously. Both of them consider
edge lengths, with a large number M indicating the length of a blocked edge. The
first heuristic H1 classifies the edges by the gain on the objective function (1) once
they are repaired. To calculate this gain efficiently, Dijkstra’s algorithm is initially
run to calculate the SP between the root and all vertices, and the values of this SP
are kept. Then it is applied again, twice for each blocked edge, using its extremities
as origin points. The SP to all destinations are updated using the distances to the
extremities of each blocked edge and the new SP is calculated from this extremities.
Blocked edges are classified in nonincreasing order of their gain.
The second heuristic H2 ranks the edges according to the number of times they
appear in SP to destinations. In this heuristic, Dijkstra’s algorithm is applied once
for each origin in the original graph, counting the number of times each blocked edge
is in a SP from an origin to a destination. This value becomes the edge ranking, and
edges are classified from its highest to the smallest ranking. Edges not contained in
SP to any destinations are randomly treated afterwards.
For both heuristics, after classifying the blocked edges, WT are allocated ac-
cording to their classification. For each time period, every WT available is allocated
to the extremity of the first edge in the classification it can reach, ensuring it does
not surpass the maximum number of WT allowed (uij ). During the first time pe-
riod, we only allocate WT that had not been already allocated in the initial steps.
Whenever an edge is repaired the values of the SP are updated.

4 Results
Experiments were addressed using a set of 80 simulated instances and a real instance.
The simulated instances were randomly generated over a graph with n = 10, m =
20, |O| = 2, |D| = 3 and T = 12 time periods. Ten different blocked edge configura-
tions were generated for each value of |B| = {5, 10, 15, 20}. For each
 configuration,
two instances were generated with the sum of all repair times R = [i,j]∈B rij = 40
and 45. The real instance consists of the graph of Port-au-Prince after the 2010
earthquake, which has the following parameters: n = 16, 657, m = 19, 558, |O| =
4, |D| = 62, |B| = 536, R = 842 and T = 106. All instances have qi = 2, ∀i ∈ O.
Tests were made on an Intel Core i7-4600M CPU with 2.90 GHz and 16 GB of
RAM. Codes were written in C and CPLEX 12.6 was used to solve the simulated
instances to optimality.
The first columns of Table 1 summarize the results obtained by CPLEX within
 C.S. Sakuraba et al. / Electronic Notes in Discrete Mathematics 52 (2016) 317–324 323

nb time accessibility
R |B| t (s) nb nodes
exceeded CPLEX H1 H2
5 0 12,63 9090,50 1,50 3,40 1,50
10 1 882,92 160058,10 1,40 4,60 1,50
40
15 1 936,65 72434,90 2,10 6,20 3,50
20 0 83,65 14348,00 2,80 10,10 3,60
5 0 23,30 11980,70 1,70 3,40 1,70
10 2 1518,22 223138,50 1,80 5,80 2,30
45
15 3 2277,90 108982,00 2,50 9,70 3,80
20 1 818,18 39447,40 3,20 12,00 3,90
Table 1
Results obtained on simulated instances

a time limit of two hours. The two first columns present the parameters R and
|B| used, and the third column (nb time exceeded) shows the number of instances
that exceeded the time limit. The two following columns show the average time
in seconds (t(s)) and number of nodes of the B&B tree (nb nodes). 90% of the
instances were solved to optimality, and the experiments suggest instances with 10
or 15 blocked edges are somehow difficult to be solved to optimality by CPLEX.
Considering that in real situations the goal is to reach the population as fast as
possible, the time period from which all population becomes accessible was used to
measure the quality of the solutions obtained. This metric is referred henceforth as
“accessibility”, and its average value obtained by CPLEX and by the heuristics for
each instance set is shown in the last three columns of the table. By calculating
the difference between them, we observed that H1 and H2 could make the whole
population accessible in the same time period as CPLEX for 41 and 54 instances,
respectively. For the remaining instances, H1 obtained a difference between 9 and
11 time periods for 37 instances, with an average difference of 4.7 time periods.
H2 obtained an average difference of 0.6 time periods, with 77 instances within a
difference up to three time periods from CPLEX solutions. Both heuristics needed
less than one second to solve any of the instances.
For the Port-au-Prince instance, all destinations are already accessible from the
first time period. However, both heuristics were able to reduce the lengths of SP
to destinations by two thirds at the end of the first period. H1 needed about 43
minutes to obtain the first period solution and 13 hours to obtain the final solution,
while H2 performed the same tasks in about 22 seconds and 20 minutes. In real
324 C.S. Sakuraba et al. / Electronic Notes in Discrete Mathematics 52 (2016) 317–324

situations, the time to calculate the allocation for the first time period plays an
important role, since once WT are sent the algorithm has a whole time period to
determine the next period allocation.

5 Concluding remarks
In this work, we presented a mathematical model and two heuristics for the Work-
troops Scheduling Problem. Optimal solutions were obtained for the majority of
graphs with 10 vertices and 20 edges generated randomly. Among the heuristics, the
ranking heuristic had a better performance achieving similar results than CPLEX
for most of the simulated instances in terms of accessibility. This heuristic was also
able to find a solution for the graph of Port-au-Prince in a reasonable time, being
recommended to be applied to real situations.
For future works, optimizations will be considered in terms of algorithm com-
plexity, together with smart ways to improve solutions for real instances within a
reasonable computational time.

References

[1] Aksu, D.T. and Ozdamar, L., A mathematical model for post-disaster road
restoration: Enabling accessibility and evacuation, Transportation Research
Part E: Logistics and Transportation Review 61 (2014), 56–67.

[2] Caunhye, A.M., Nie, X. and Pokharel, S., Optimization models in emergency
logistics: A literature review, Socio-Economic Planning Sciences 46 (2012), 4–
13.

[3] Dijkstra, E. W., A note on two problems in connexion with graphs, Numerische
Mathematik 1 (1959), 269–271.

[4] Duque, P. M. and Sörensen, K., A GRASP metaheuristic to improve accessibility

after a disaster, OR Spectrum 33 (2011), 525–542.

[5] Yan, S., Chu, J.C. and Shih, Y.L., Optimal Scheduling for Highway Emergency
Repairs Under Large-Scale Supply–Demand Perturbations, IEEE Transactions
on Intelligent Transportation Systems 15 (2014), 2378–2393.

[6] Yan, S. and Shih, Y.L., Optimal scheduling of emergency roadway repair and
subsequent relief distribution, Computers & Operations Research 36 (2009),
2049–2065.