Socio-Economic Planning Sciences xxx (xxxx) xxx

Contents lists available at ScienceDirect

Socio-Economic Planning Sciences

journal homepage: [Link]

Logistics optimization through a social approach for food distribution

Rafael Granillo-Macías
Universidad Autónoma del Estado de Hidalgo, Escuela Superior de Ciudad Sahagún, Área Académica de Ingeniería Industrial, Tepeapulco, Hidalgo, México

A R T I C L E I N F O A B S T R A C T

Keywords: The management of supply chains is becoming more important in economic and social environments. Currently,
Social sustainability the social sustainability is a factor that must be considered to design governmental strategies and policies. The
Cluster objective of this research paper was to show, with a case study an approach to optimize distribution and delivery
Distribution
logistics of food in a social assistance program of school breakfast using mathematical models that include
Logistics
transportation distance, optimal locations, and vehicle routing through different clusters. By using qualitative
variables like poverty levels, food insecurity and social exclusion, different clustering methods are proposed with
the purpose of identifying the common characteristics in the studied population; and at the same time, reducing
the distribution complexity. The results show an efficient approach to design a supply chain that includes eco­
nomic and social factors. The new model developed in this paper can be used to plan social assistance
governmental programs, to identify the specific needs and characteristics of the beneficiaries, minimizing the
total cost of the distribution network logistics when delivering food for school meals.

1. Introduction School Meals, daily serves 3.5 million students of almost 60,000 schools,
including some indigenous communities [33]).
According to Ref. [34]; the management of social problems in supply In Mexico, one of the food assistance programs of wider coverage at a
chains is getting the attention of more professionals and researchers. The national level is the School Feeding Program (PAE) [23]), which aims at
distribution of food to fight poverty and food insecurity, is one of the contributing to the food security of the school population in need of
most relevant problems in supply chains. social assistance, by providing them with breakfast. Particularly, this
Based on data from Food and Agriculture Organization of the United program represents a feeding strategy in Mexico because, according to
Nations [13]) and the World Food Program (WFP), in Latin America, the System for Family Development ( [8], 6 out of 10 Mexican families
extreme poverty and food insecurity leads to 17.4 million children receive some type of food support; and from this universe, 18.2% of the
working to bring food to their families, meaning that they have to drop population benefits from the PAE. Every day, through governmental
out school. bodies coordinated by the PAE, 6 million breakfasts are delivered [9].
One of the strategies to fight this social problem are food programs The target population of PAE are girls, boys and adolescents in risk
based on school meal plans which are applied by different countries and vulnerability conditions, attending public schools of the national
worldwide. education system, located in indigenous, rural or marginalized areas
Food programs generally tackle social and economic objectives to according to the DIF, which coordinates the distribution of the meals
improve enrollment, retention, participation, and the nutritional state of based on different variables from a logistics perspective, like the target
children in schools [21]. Through this strategy, it is expected to cover population being delimited in a specific time and space, in this case the
the food needs in different schools. These food programs serve 368 school itself [9].
million children and represent an investment of up to 75,000 million In general, the approach of these feeding programs based on school
dollars [33]). meals are established under the premise that in order for the costs to
In Latin America, 85 million children benefit from these programs, remain low, the operation must be simple, meaning that the organiza­
making them the most important programs in the region. For example, tion and logistics must be very efficient in order to meet the needs of the
Bogota has the most important logistics in Colombia, regarding school school population. In the supply chain, the final cost of the delivery of
food, with a daily distribution of 730,000 breakfast meals and 163,000 goods and services directly depends on the cost of transportation and the
hot lunches. The government of Peru, through its National Program of logistics routing [15].

E-mail address: rafaelgm@[Link].

[Link]
Received 26 August 2020; Received in revised form 11 November 2020; Accepted 15 November 2020
Available online 26 November 2020
0038-0121/© 2020 Elsevier Ltd. All rights reserved.

Please cite this article as: Rafael Granillo-Macías, Socio-Economic Planning Sciences, [Link]
R. Granillo-Macías Socio-Economic Planning Sciences xxx (xxxx) xxx

Generally, the optimization problem regarding the distribution costs algorithm around medoids (PAM), being this last one the most common
implies the use of tools and models with the objective of improving the of the k-medoids methods [17].
decision-making at a strategic, tactical and operational level, with the The objective of this article was to show through a case study of the
main scope of suggesting the routes that mainly minimize the delivery PAE, an approach to optimize de logistics of food delivery for the school
time and costs. breakfast, through mathematical models, considering aspects like
The strategic decisions like the Facility Location Problems (FLP) as transportation distance, optimal location, vehicle routing, and qualita­
well as the tactical decisions like the Vehicle Routing Problems (VRPs) tive variables to form different clusters.
have had a major attention in the different enhance proposals for dis­
tribution [20]. [31]; suggest that the facility location and the vehicle 2. Materials and méthods
routing are factors that relate and influence on each other, leading to
suggest approaches that include the simultaneous optimization of the The case study was a population that benefits from the distribution of
decisions on facility location and routing. school breakfast in the State of Hidalgo, Mexico, that according to the
The problems that involve routing and programming multiple vehi­ National Population Council [7]), is ranked 4th in marginalization at a
cles, starting from one point to multiple locations under time frame national level, making it one of the main beneficiaries of the PAE.
restrictions, are considered Vehicle Routing Problems with Time Win­ The methodology that was used for this proposal is based on: (1) the
dows (VRPTW) [32]. identification of the target population (locations), (2) the calculation of
An adaptation of the Generalized Vehicle Routing Problem (GVRP) is the transportation distance between locations, (3) the formation of
proposed by Ref. [25]; this author considers a mathematical model of clusters based on qualitative and quantitative variables, (4) the facility
GVRP based on integer programming; formulating groupings reduce the locations, and (5) the routing to determine the best distribution.
complexity in a routing problem. Considering large real problems [1],
propose a distributed vehicle routing problem with time windows, based 2.1. Identification of population
on cluster first, route second methods. The vehicles autonomously
generate a number of clusters equal to the number of available vehicles As a first step for this case study, basic level schools were selected,
[24]. proposed a novel VRP model that permits reductions loss due to located in the Southeastern region of the State of Hidalgo, based on data
the perishable. In this model, a two-phase strategy is proposed, first a obtained from the Ministry of Public Education in Hidalgo (SEP, 2019).
clustering based on a modified p-median model and then a VRP for each In total, 188 schools distributed in 8 regions (Almoloya, Apan, Cuaute­
cluster. pec, Emiliano Zapata, Singuilucan, Tepeapulco, Tlanalapa, and Zem­
[22] propose an algorithm based on the maximum-minimum dis­ poala) were located through Geographic Information Systems (GIS).
tance method using three stages: (1) cluster customer points into the These regions are located in indigenous, rural and marginalized zones.
lowest number of groups, (2) according to the maximum vehicle ca­ The school enrollment in these schools is within a range of 839 and
pacity, the load demand in each group is adjusted to create suitable 6,793 students.
customer points in each clustering group by adopting ‘push-out’ and The student population that is located in the selected regions for this
‘pull-in’ operations and (3) a tabu search is used to minimize the total study is also characterized for presenting a level of marginalization,
travel distance. going from a low to a high marginalization. This level of marginalization
[27] present a mathematical model of mixed integer linear pro­ is a first element to identify the target population that consequently can
gramming, for the optimization of delivery routes to different clients, benefit from the PAE. To know the level of marginalization, CONAPO
based on the VRP version, called Selective Vehicle Routing Problem indicators were used (2019), which indicate the level of marginalization
(SVRP) in which clients are divided according to characteristics specific in each municipality based on 3 strata: low, medium and high
grouping and rules of relevance. marginalization.
A particular case is the Clustered Vehicle Routing Problem (CluVRP) Another qualitative variable that was evaluated in this locations was
which is a variant of VRP, in which the clients are divided into clusters the indicator of food security obtained from the National Health and
[26]. Nutrition Survey [11]), based on the Mexican Food Security Scale
The problem of CluVRP was suggested by Sevaux and Sorensen [29]; (EMSA) which aims at identifying the perception of food security and
developed as a model of package delivery, which has generated a starvation feel around the homes. The EMSA identifies four levels of food
growing interest due to its theoretical challenges and wide areas of insecurity: severe food insecurity, moderate food insecurity, low food
application [3,26]. The essence of CluVRP formulated by Sevaux and insecurity, and food security.
Sorensen [29]; is to divide the clients into zones (clusters) and then do
the routing individually for each cluster. 2.2. Calculation of distance
[12]; also describe a method to divide the graph nodes into clusters,
for the case of CluVRP. The combination of CluVRP with FLP, generates Once the nodes were geographically located, which formed the dis­
the problem called Location Routing Problem (LRP), which is formu­ tribution network, the distance was calculated between these nodes with
lated in two stages integrated by the set of clients and the distribution the purpose of forming a matrix of distances. The traditional way to
centers (DCs). determine distance between 2 nodes is the calculation through
Since the LRP is classified within the NP Hard problems, an approach Euclidean distances; however, one technological alternative with a wide
for their solution are the models based on clusters [2]. In the models functionality to determine in a precise way the distance between an
based on CluVRP, the cluster analysis approach they suggest, relies on origin and a destination are the geospatial applications and web services
the process that consists on grouping elements into clusters with the [30].
objective of getting elements with a maximum similarity in the same The process used to obtain the distances was performed as follows:
groups and minimal similarity between elements of different groups. (1) through GIS were obtained the geographical latitude and longitude
The cluster analysis is mainly used to explore in data mining, and in for each school, (2) it was generated a data base with the information of
other fields like autonomous learning, pattern recognition, image anal­ the locations, (3) by using the obtained data base, the request was
ysis, recovery of information, bioinformatics, data compression, and encoded with the server API, Matrix Distance Google through the Hy­
computer graphics [5,19,31]. pertext Transfer Protocol (HTTP), (4) the request is executed in the
In general, the algorithms used to group and analyze clusters can be server API which performs the calculation to prove that the origin and
classified into: hierarchical and partitioning. Particularly, among the destination are possible, this means, that there is a terrestrial connection
most used methods are: k-means clustering and the partitioning between them, which implies to verify the available infrastructure for

2
R. Granillo-Macías Socio-Economic Planning Sciences xxx (xxxx) xxx

the transportation, (5) the results obtained are delivered in an standard k ∈ K set of vehicles
open format, in this case in JavaScript Object Notation (JSON), and (6) ok ∈ VD set of distribution centers of origin for one vehicle
the collection of information was carried out in a worksheet to form the i, j ∈ A set of origins and destinations
matrix of distances.
The obtained information about the transportation distances be­ Variables:
tween locations will allow to understand the available routes for logis­
tics in a more effective way to later be used in the planning of the xkij vehicle k that goes across i, j⋅(1 ⋅if ⋅xkij > 0, 0 in other case)
distribution network by the PAE.
yki vehicle k visits and serves the vertex i ⋅(1 ⋅if ⋅yki > 0, 0 in other case)
zkij amount of transported breakfasts by vehicle k in the arch i,j
2.3. Cluster formation
tik time it takes for the vehicle k to get to vertex i
vi amount of penalties of time Windows of vertex i
Authors like [32]; pointed out that to design the logistics of a dis­
tribution network multiple factors about the clients must be considered;
Parameters:
among these factors are: the geographical location, accessibility, trans­
portation conditions, among others.
The fact that there is a great number of nodes (schools) complicates ckij travel cost in the arc i,j of vehicle k
things, both for the logistics as well as the distribution network, having f k fixed cost for the use of vehicle k
an impact on the food delivery efficiency within the PAE. An approach to dij distance i,j
reduce the complexity and improve the performance of the solution of q i amount of delivery service in the vertex i
̂
distribution problems at large scale, is a break down into several stages d ij duration of driving in the arch i,j
̂
[21,31]. Specifically, for this case study, a cluster approach was used, si time of service required by the client (school)
which aims at creating k clusters that allow to minimize the complexity Wk time limit of the vehicle’s work k
of the distribution, with the particular characteristic that this clustering
Qk vehicle capacity k
is based on the food insecurity variables, the level of enrollment and the
θ equal 1 if the time window is not met, 0 otherwise
level of marginalization. Including these qualitative variables to form
L penalty cost by time unit
clusters, allows (from the food security perspective) to identify groups of
ϕ equal 1 if the vehicle has to return to a specific distribution,
schools that are highly similar, favoring the level of the service and the
0 otherwise
satisfaction of the users with particular characteristics.
M large constant
To form the k clusters with demand n nodes, k-means methods, hi­
ai , bi time interval for the client (school)
erarchical clustering and PAM algorithm were used, converting cate­
gorical variables into discrete data, using the Ward’s proximity measure.
The objective function (a), for this problem, minimizes the cost of
Each of these methods generated different cluster arrangements, which
vehicle travel, the fixed cost for the use of vehicles, and the penalty that
were evaluated in a later stage to identify the best configuration of k
might be generated for not complying with the delivery of breakfasts on
clusters.
time (time windows).
∑∑ ∑∑ ∑
2.4. Facility locations Minimize ckij xijk + f k xok k ,j + L vi (a)
(i,j)∈A k∈K j∈VC k∈K i∈V

In the logistics design, the multi-echelon distribution systems have The restrictions for this model are.
turned into an alternative deliver in urban zones [14]. In this case, a ∑
two-echelon distribution model is proposed (the first echelon is the yki = 1 ∀i ∈ VM (b)
distribution center and the second echelon are the schools), based on the
k∈K

LRP approach. The network is proposed in 3 levels: in level 0 of distri­ ∑

yki ≤ 1 ∀i ∈ Vc \VM (3)
bution, there is the warehousing from where the sending of the products k∈K
will begin; then, level 1 includes the distribution centers, and finally, in
∑ ∑
level 2 there is the school population. xkij ≤ xjik ∀j ∈ Vc , k ∈ K (4)
The Weber model for facility location suggested for the two first j∈V\{i} j∈V\{i}
levels, is based on the Euclidean equation, defined in Equation (1)
∑ ∑
( ) ( )2 ( )2 zkji − zkij = q̂i yki ∀i ∈ VC , k ∈ K (5)
d22 x, ai = x1 − ai1 + x2 − ai2 (1) j∈V\{i} j∈V\{i}

for all x = (x1 , x2 ) ∈ R2 and all ai : = (ai1 , ai2 ), i = 1, …, m, thus the ∑ ∑

location problem is given by Equation (2) zkok ,j = q i yki
̂ ∀k ∈ K (6)
i∈VC i∈VC
m (
∑
m
( ) ∑ ( )2 ( )2 ) ( ) ( )
d22 x, ai = x1 − ai1 + x2 − ai2 →minx∈R2 (2)
i=1 i=1
tik + ̂dij + si xkij − W k 1 − xijk ≤ tjk ∀(i, j) ∈ A : j ∈ VC , k ∈ K (7)

where x are the location points of schools and a the optimal location. ai ≤ tik ≤ bi − si + vi ∀i ∈ VC , k ∈ K (8)

2.5. Vehicle routing with time windows vi ≤ Mθ ∀i ∈ VC (9)

The VRPTW, used for the optimization of the suggested logistics tokk = rk ∀k ∈ K (10)
network, was formulated based on a mixed integer linear program. In ( )
this model, VD is the set formed by the distribution centers, Vc is the set tik + si + ̂
d ij xki,rk ≤ brk + vrk + M(1 − ϕ) ∀(i, j) ∈ A : i ∈ VC , k ∈ K (11)
formed by the schools, and V = VD ∪ VC . The decision parameters and
variables related with the design model are:
Rates and sets: zkij ≤ Qk ∀(i, j) ∈ A, k ∈ K (12)

3
R. Granillo-Macías Socio-Economic Planning Sciences xxx (xxxx) xxx

∑
dij xkij ≤ Dk ∀(i, j) ∈ A, k ∈ K (13)
(i,j)∈A

∑
̂ ̂ k ∀(i, j) ∈ A, k ∈ K
d ij xkij ≤ D (14)
(i,j)∈A

∑ ∑
si yki + ̂
d ij xkij ≤ W k ∀(i, j) ∈ A, k ∈ K (15)
i∈VC (i,j)∈A

xkij ∈ {0, 1} ∀(i, j) ∈ A, k ∈ K (16)

yki ∈ {0, 1} ∀i ∈ VC , k ∈ K (17)

vi ≥ 0 ∀i ∈ VC (18) Fig. 1. Locations.

zkij ≥ 0 ∀(i, j) ∈ A, k ∈ K (19) coordinates in latitude and longitude obtained in GIS were used with the
Restrictions (b) and (3) guarantee that every school is visited at least application of API, Matrix Distance Google to calculate transportation
once. Restriction (4) allows to keep the flow in the network design, this distances and time between locations. Through positions of m-origins
means that an entrance flow is needed if there is an exit flow. In the two- and n-destinations it was determined a matrix mxn of transportation
echelon design, similarly, restrictions (5) and (6), guarantee the distances and times, in this case, it was used the parameter of distances
connection and the delivery flow toward the schools. Restriction (7), for locations, thus the matrix is formed by 35, 344 distances between
according to the proposal of [10]; allows eliminating subtours. The origins i and destinations j.
limits inferior a and superior b of the time Windows for each school is Subsequently, the clustering was performed considering the vari­
determined by restriction (8), while (9) considers the violation to the ables of: school enrollment (M) in 3 levels low, medium and high; food
time-lapse of delivery established by the time window. Restrictions insecurity (A) used for the EMSA in food insecurity levels low, moderate
related to vehicle k in this model are expressed in (10) and (11), which and severe; and finally, the level of marginalization (G) in levels low,
indicate the time the vehicles start to work and also guarantee that they medium and high. Each of these variables was categorized for its cluster
return to the distribution center in the required time. The capacity of analysis in a range of 1–3, where, for this case, a value of three repre­
vehicles k is imposed in (12). Restrictions (13), (14) and (15) are related sents the highest level.
with the transportation distance, the time of driving and the time limit of Through the clustering approach proposed by Ref. [6] and by using
each vehicle’s work k, respectively. The binary decision variables are multivariate statistical techniques, it was calculated with the NbClust
defined by restrictions (16) and (17). Finally, the non-negative variables package of the R the amount of optimal k clusters to form the logistics
are mentioned in (18) and (19). It is important to mention that the network design. As a result, the optimal number of clusters was four
distances for VRPTW model, are obtained from API Matrix Distance (Fig. 2); however, with the purpose of exploring other possible clusters
Google, to calculate precisely, the distances between an origin and a that improve the network design, clusters with k = 5 and k = 6 were also
destination. considered.
To find the solution to this VRPTW, it was implemented a variant of In the formation of clusters three methods were applied with the
the Adaptive Large Neighborhood Search (LNS) algorithm suggested by purpose of identifying the best arrangements. The k-means, hierarchical
Ref. [10]; through worksheets in Microsoft Excel. The LNS algorithm is clustering and PAM were implemented with k = 4, 5 and 6 clusters. For
based on, according to Ref. [10]; in the following stages: (1) obtain in­ example, for k = 4 using k-means the arrangement of clusters was
formation about the location of the starting points of distribution, the configured as graphically shown in Fig. 3.
location of schools, the transportation distances between the points and The silhouette analysis made in each cluster allowed validating the
the characteristics of vehicles, (2) build an initial solution by adding consistency of data within the arrangement, by using the scheme based
schools to the route, choosing schools that minimize the transportation on the Nbclust module of software R. Fig. 4 shows the result for k = 4 by
cost in an iterative way, (3) improve the suggested solution by using four using the k-means method; in this case, the clusters were formed by 32,
local search operators suggested by Ref. [16]; (4) register the solution to 33, 67, and 56, schools that correspond to clusters 1, 2, 3 and 4,
the former step as the best solution known and (5) repeat the process until respectively. The average value of si was of 0.46 in this clustering
the allowed time restriction of the Central Processing Unit (CPU) is met. configuration.
The LNS algorithm works very well for up to 100 clients (in this case, the To interpret each configuration a means test was performed with the
number of locations for each cluster) and returns acceptable results for
larger instances [10], in which the algorithm performance is better
when a distance constraint.

3. Results

The selected schools were situated as shown in Fig. 1. By using the

GIS tool the latitude and longitude of each school, where the school
breakfasts are offered, were geographically determined.
These schools are located in different geographical regions (munic­
ipalities), where each point shown in Fig. 1, also represents a node of
breakfast demand specifically related with the school enrollment in each
of these locations. For this case study, based on SEP data [28]; newly
admitted population was used, this is, students of first grade in the
southeast of the State of Hidalgo. On average, it is estimated an
enrollment of about 152 students per school. The geographic Fig. 2. Clusters.

4
R. Granillo-Macías Socio-Economic Planning Sciences xxx (xxxx) xxx

mean values of (μ) in each clustering (centroid). The obtained value of μ

allowed identifying the profile of the clusters, meaning, the character­
istics that are predominant for each clustering. For example, in the k = 4
case for cluster one with the k-means algorithm, it is observed a high
predominance toward food insecurity for this group of schools; also for
cluster two formed by 33 schools, presents a higher domain of the var­
iable related to the school enrollment (Table 1). In this same configu­
ration with k = 4 but using the hierarchical clustering method it was
obtained that for cluster one, the level of marginalization is the most
important variable, followed by the aspects related with food insecurity.
These values are similar to those obtained for the same cluster using the
PAM algorithm.
In total, nine possible configurations of clusters were considered to
Fig. 3. Formation of clusters k = 4 with k-means. design the distribution network, the results of these configurations are
shown in Table 1.
Once the schools and its demand points were identified, by using
equation (2), it was calculated the optimal location in level zero, that is,
the warehousing where the products included in the school breakfasts
are found. The resulting optimal location is shown in Fig. 5. This point
geographically presents the location that minimizes the transportation

Fig. 4. Silhouette analysis k = 4 with k-means.

Fig. 5. Optimal location.

Table 1
Results of configurations.
Cluster k-means Hierarchical clustering PAM

N μ s N μ s N μ s

M A G M A G M A G

1 32 1.25 3.00 1.38 0.52 63 1.00 1.63 2.54 0.40 48 1.23 1.96 2.00 0.24
2 33 2.36 1.27 1.21 0.36 42 1.29 2.48 1.00 0.44 48 1.02 2.02 3.00 0.44
3 67 1.01 2.01 2.72 0.42 57 1.67 1.05 1.12 0.44 42 1.29 2.48 1.00 0.43
4 56 1.00 1.29 1.18 0.53 26 1.15 3.00 2.54 0.52 50 1.60 1.00 1.00 0.55
k=4

1 28 1.46 2.82 1.11 0.45 63 1.00 1.63 2.54 0.36 32 1.09 1.69 2.00 0.49
2 25 2.48 1.04 1.16 0.48 42 1.29 2.48 1.00 0.35 48 1.02 2.02 3.00 0.39
3 66 1.06 2.35 2.53 0.34 27 2.41 1.11 1.26 0.37 48 1.17 2.67 1.25 0.32
4 46 1.00 1.35 1.00 0.65 30 1.00 1.00 1.00 1.00 30 2.40 1.20 1.13 0.36
5 23 1.00 1.00 2.57 0.61 26 1.15 3.00 2.54 0.52 30 1.00 1.00 1.00 1.00
k=5

1 60 1.00 2.42 1.47 0.29 40 1.00 2.00 2.53 0.58 44 1.14 2.05 2.00 0.18
2 24 2.46 1.00 1.17 0.45 42 1.29 2.48 1.00 0.35 48 1.02 2.02 3.00 0.42
3 35 1.03 2.40 3.00 0.63 27 2.41 1.11 1.26 0.35 21 1.24 2.00 1.00 0.66
4 16 2.13 2.44 1.38 0.27 23 1.00 1.00 2.57 0.53 25 2.48 1.04 1.16 0.46
5 23 1.00 1.00 2.57 0.56 30 1.00 1.00 1.00 1.00 20 1.25 3.00 1.00 0.66
6 30 1.00 1.00 1.00 1.00 26 1.15 3.00 2.54 0.45 30 1.00 1.00 1.00 1.00
k=6

N = number of schools M = School enrollment A = food insecurity G = level of marginalization.

5
R. Granillo-Macías Socio-Economic Planning Sciences xxx (xxxx) xxx

distance between the school simple for this case study. This point pre­
sents the strategic location where the government office operating the
PAE should initially concentrate the goods for their distribution.
Therefore, this location is used as the starting point of the distribu­
tion toward the suggested clusters, thus from this point the products
must be sent to the different clusters. Due to the characteristics of this
supply chain of school breakfast, another location is necessary where the
products are received, working as a distribution center for each school
clustering.
This way, a location for the distribution center was defined in each
cluster which corresponds to level one of the distribution network. From
this level one location, the distribution logistics start toward the final
client (the schools), therefore equation (2) is applied again to get the
optimal location for the distribution center in each clustering. For Fig. 7. Optimal location for cluster 1 with k-means.
example, for the case k = 4 using the PAM algorithm, it was obtained
that for cluster 1, the optimal location of the distribution center was
defined as shown in Fig. 6, while by using k-means for the same cluster
the obtained configuration is the one shown in Fig. 7.
In this configuration using PAM, the point disposed as distribution
center through equation (2) is close to a location very near to one of the
48 schools that form cluster 1, while in the case of k-means with 32
schools, the location changes significantly.
Using the hierarchical clustering with k = 4 value, it was obtained
the configuration shown in Fig. 8 for cluster 1.
In this case, con hierarchical clustering, the configuration formed by
63 schools, locates the distribution center very near to the schools in
cluster 1, which presents an advantage for distribution compared to the
k means method.
Subsequently, it was performed the routing using VRPTW in each
cluster. The model VRPTW, was solved through the LSN algorithm based
Fig. 8. Optimal location for cluster 1 with hierarchical clustering.
on [10]; under the following operation considerations of PAE and with
data from the National Institute of Ecology and Climate Change [18])
(Table 2): Table 2
Particularly, the time window is an indicator of the level of the Considerations for the VRPTW.
service given to the schools, in this case, if a breakfast is delivered within
Time window 5:00–11:59 h
the time window, the delivery is considered 100% fulfilled; however, if Service time 15 min
the breakfasts are delivered out of the time window, the fulfillment will Vehicle capacity (per transport unit) 1500 breakfasts
be 0% causing a penalty for the delivery. Fixed costs (per transport unit) 175 usd
Cost per unit of distance traveled 0.1 usd/km
Through the VRPTW model, the different configurations suggested
with k = 4, 5 and 6 were solved, considering the k-means, PAM and
hierarchical clustering methods. The results indicated that among all the
configurations, the arrangement with k = 4 through a hierarchical
clustering represents the best scenario for distribution. Fig. 9 shows a
comparative scheme of the costs obtained based on the different con­
figurations, being the clustering with k-means the most expensive ones.
Regarding the level of service in each configuration, the methods
PAM and hierarchical clustering with k = 4 were the best ones with 87%
and 86%, respectively. Fig. 10 shows the behavior of the level of service,
which decreases when passing from cluster four to five. When consid­
ering the relation between the cost of the distribution network for PAE

Fig. 9. Distribution costs for the different clusters.

and the level of service, it is observed that the hierarchical clustering

method with k = 4 turns to be the most convenient.
An additional factor for the distribution network design is the
number of vehicles that are necessary, in this case the results showed
that with k = 4 are needed between 23 and 25 vehicles, being the hi­
erarchical clustering the best configuration. Fig. 11 shows the vehicles
needed according to the different suggested configurations.
In general, the best solution that was found for the distribution
Fig. 6. Optimal location for cluster 1 with PAM. problem of the PAE was through hierarchical clustering with k = 4, thus

6
R. Granillo-Macías Socio-Economic Planning Sciences xxx (xxxx) xxx

Fig. 12 graphically shows the obtained routes for these clusters.

Fig. 12 shows that for cluster 1 there are seven routes; for cluster 2
there are seven routes; for cluster 3 there are six routes; and finally, for
cluster 4 there are three routes.

4. Discussion

The best economic performance must not be the only objective

regarding logistics, but it must also be considered the social impact. The
results of this study show that through the cluster approach for the
VRPTW case suggested by authors like [26,31]; and including social
variables as suggested by Yawar et al. [34] it is possible to optimize a
two-echelon distribution network. The proposal of a two-echelon dis­
Fig. 10. Level of service for the different clusters. tribution system allowed to reduce the complexity of logistics and to
improve the level of service as suggested by Ref. [4]; who showed that,
by using clusters the complexity is reduced in benefit of logistics and
optimization.
In the clusters to identify the target population, it was statistically
proved through this study that, the methods based on the PAM algo­
rithm and hierarchical clustering are good approaches for clusters with
qualitative variables.
Compared to the proposals of [12,26,31]; the contribution of this
research allow identifying the common characteristics (qualitative var­
iables) in the target population, included in the distribution network
design. Identifying the characteristics of the target population in the
supply chain also promote the strategies that must be implemented as
suggested by Jahre et al. [19]; Yawar et al. [20,34]. For example, if in a
cluster it is detected a high level of food insecurity, the strategy should
include mitigating this social problem through different types of
breakfast with nutritional values according to the needs of this
Fig. 11. Number of vehicles for the different clusters. population.
Specifically, in the case of the PAE, the clusters will serve as an in­
dicator to establish if the breakfast distribution should be promoted,
depending on the characteristics and needs of the target population that
form this supply chain.

Fig. 12. Routes for clusters with k = 4 through hierarchical clustering.

7
R. Granillo-Macías Socio-Economic Planning Sciences xxx (xxxx) xxx

The integration of this proposal with a VRPTW model, allows to [7] Conapo, Consejo Nacional de Población. Consejo nacional de Población. México:
Índice de Marginación por entidad federativa 2015; 2019 [On-líne]. Available in:
guarantee that the deliveries of school breakfast are carried out in a
[Link]
timely manner, leading to a better level of service for the students. apitulosPDF/1_4.pdf.
[8] Dif, Sistema Nacional para el Desarrollo Integral de la Familia. Estrategia integral
de asistencia social alimentaria [On-line]. Available in: [Link]
5. Conclusions
dgadc/direccion-de-alimentacion/estrategia-integral-de-asistencia-social-ali
mentaria/; 2018.
The proper planning in the design, analysis and logistics, is a factor [9] Dif, Sistema Nacional para el Desarrollo Integral de la Familia. La participación
that has an impact on the development of different fields like the eco­ social en el programa desayunos escolares [On-line]. Available in: [Link]
[Link]/dgadc/wp-content/uploads/2018/11/Participaci%C3%B3n-Social-PDE-2
nomic, social and territorial field. From a systemic approach, the pro­ [Link]; 2018.
jection and planning of the activities for the logistics is an essential part [10] Erdoğan G. An open source spreadsheet solver for vehicle routing problems.
for a comprehensive understanding of human activity. The issues related Comput Oper Res 2017;84:62–72. [Link]
[11] Encuesta Nacional de Salud y Nutricion de Medio Camino 2016. (ENSANUT 2016).
to social sustainability require an understanding of their complexity, Mexico: Instituto Nacional de Salud Publica; 2016 [On-Line]. Available in, https
including the design of strategies and policies (mainly governmental) for ://[Link]/encuestas/ensanut2016/[Link].
the supply chain, with the purpose of closing the gaps in the manage­ [12] Exposito-Izquierdo C, Rossi A, Sevaux M. A two-level solution approach to solve
the clustered vehicle routing problem. Comput Ind Eng 2016;91:274–89. https://
ment of social problems. [Link]/10.1016/[Link].2015.11.022.
This article suggests a way to design a two-echelon distribution [13] Fao, Food and Agriculture Organization. Panorama de la seguridad alimentaria y
network to deliver school breakfasts, aiming at satisfying the feeding nutricional en América Latina y el Caribe [On-line]. Available in: [Link]
org/americas/publicaciones-audio-video/panorama/2018/es/; 2018.
needs of students located in different schools. The presented approach
[14] Gonzalez-Feliu J. Vehicle routing in multi-echelon distribution systems with cross-
allows the decision makers in governmental programs to design strate­ docking: a systematic lexical-metanarrative analysis. Comput Inf Sci 2013;6(3):
gies to improve breakfast (nutrients) in the clusters, worked according 28–47. [Link]
[15] Granillo-Macías R, González-Hernández I, Martínez-Flores J, Caballero-Morales,
the level of marginalization, food inequality and enrollment.
Olivares-Benítez E. Hybrid model to design a distribution network in contract
This proposal also showed a new model based on technological tools farming. Dyna 2019;86(208):102–9. [Link]
like GIS, which minimizes the total cost of the distribution network lo­ v86n208.72056.
gistics, considering social sustainability factors. [16] Groër C, Golden B, Wasil E. A library of local search heuristics for the vehicle
routing problem. Math. Prog. Comp 2010;79(2):79–101. [Link]
By using qualitative variables such as the levels of poverty, food s12532-010-0013-5.
insecurity and social marginalization, it was possible to form clusters, [17] Harikumar S, Surya PV. K-medoid clustering for heterogeneous DataSets. 4th
with the purpose of identifying the common characteristics of the school international conference on eco-friendly computing and communication systems.
Procedia Comput. Sci. 2015;70:226–37. [Link]
population and also reduce the distribution complexity, leading to an procs.2015.10.077.
optimization of logistics. [18] Inecc, Instituto Nacional de Ecología y Cambio Climático. Portal de indicadores de
It was proven that the classifications based on the PAM algorithm eficiencia energética y emisiones vehiculares [On-line]. Available in: [Link]
[Link]/ecovehiculos/ecovehiculos/[Link]?vehiculo_id=8194;
and hierarchical clustering can be used to design clusters with qualita­ 2019.
tive variables, offering an added value for a decision-making based on [19] Jahre M, Jensen LM. Coordination in humanitarian logistics through clusters. Int J
priorities, demand needs and the inclusion of the target population, for Phys Distrib Logist Manag 2010;40:657–74. [Link]
09600031011079319.
the design of distribution networks.
[20] Ma X, Chen X, Ding C, Wang Y. Sustainable station-level planning: an integrated
As future works, and considering the social approach of this pro­ transport and land use design model for transit-oriented development. J Clean Prod
posal, the analysis of variables could lead to identify the food needs to 2018;170(1):1052–63. [Link]
[21] Mahadevan B, Sivakumar S, Dinesh-Kumar D, Ganeshram K. Redesigning midday
form clusters, where health indicators like the level of malnutrition are
meal logistics for the akshaya patra foundation: OR at work in feeding hungry
included. Regarding the transportation, the variable of type of food that school children. Interfaces 2013;43(6):530–46. [Link]
is being distributed could be included to manage the logistics according inte.2013.0714.
to the cold chain and the expected quality at the consumption points. [22] Min JN, Jin C, Lu LJ. Maximum-minimum distance clustering method for split-
delivery vehicle-routing problem: case studies and performance comparisons. Adv.
This approach could even lead to the humanitarian logistics to identify Prod. Eng. Manag. 2019;14(1):125–35. [Link]
the needs and risks along the operation and distribution, prioritizing apem2019.1.316.
social factors. [23] Morales-Ruán, Ma del C, Shamah-Levy T, Mundo-Rosas V, Cuevas-Nasu L, Romero-
Martínez M, Villalpando S, Rivera-Dommarco JÁ. Programas de ayuda alimentaria
en México, cobertura y focalización. Salud Publica Mex 2013;55(2):S199–205.
Funding [24] Orjuela-Castro JA, Orejuela-Cabrera JP, Adarme-Jaimes W. Last mile logistics in
mega-cities for perishable fruits. J Ind Eng Manag 2019;12(2):318. [Link]
org/10.3926/jiem.2770.
This work was supported by Programa para el Desarrollo Profesional [25] Pop PC, Kara I, Marc AH. New mathematical models of the generalized vehicle
Docente (PRODEP), México, [grant number UAEH-EXB-152]; and Na­ routing problem and extensions. Appl Math Model 2012;36(1):97–107. https://
tional Council of Science and Technology (CONACyT) [grant number [Link]/10.1016/[Link].2011.05.037.
[26] Pop PC, Fuksz L, Marc AH, Sabo C. A novel two-level optimization approach for
CVU-205095], México. clustered vehicle routing problem. Comput Ind Eng 2018;115:304–18. [Link]
org/10.1016/[Link].2017.11.018.
References [27] Sabo C, Pop PC, Horvat-Marc A. On the selective vehicle routing problem.
Mathematics 2020;8(5):771. [Link]
[28] Sep, Secretaria de Educación Pública Hidalgo. Estadísticas básicas [On-line].
[1] Abbatecola L, Fanti MP, Pedroncelli G, Ukovich W. A distributed cluster-based
Available in: [Link]
approach for pick-up services. IEEE Trans Autom Sci Eng 2019;16(2):960–71.
html; 2019.
[Link]
[29] Sevaux M, Sörensen K. Hamiltonian paths in large clustered routing problems.
[2] Barreto S, Ferreira C, Paixao J, Sousa-Santos B. Using clustering analysis in a
Proceedings of EU/MEeting workshop on metaheuristics for logistics and vehicle
capacitated location-routing problem. Eur J Oper Res 2007;179(3):968–77.
routing 2008;8:411–7.
[Link]
[30] Socharoentum M, Karimi AH. A comparative analysis of routes generated by Web
[3] Battarra M, Erdogan G, Vigo D. Exact algorithms for the clustered vehicle routing
Mapping APIs. Cartogr Geogr Inf Sci 2015;42(1):33–43. [Link]
problem. Oper Res 2014;62(1):58–71. [Link]
15230406.2014.976656.
[4] Belgin O, Karaoglan I, Altiparmak F. Two-echelon vehicle routing problem with
[31] Wang Y, Assogba K, Liu Y, Ma X, Xu M, Wang Y. Two-echelon location-routing
simultaneous pickup and delivery: mathematical model and heuristic approach.
optimization with time windows based on customer clustering. Expert Syst Appl
Comput Ind Eng 2018;115:1–16. [Link]
2018;104:244–60. [Link]
[5] Caruso G, Gattone SA, Fortuna F, Di Battista T. Cluster Analysis for mixed data: an
[32] Wang Y, Ma X, Lao Y, Wang Y. A fuzzy-based customer clustering approach with
application to credit risk evaluation. Soc Econ Plann Sci 2020:100850. [Link]
hierarchical structure for logistics network optimizacion. Expert Systems with
org/10.1016/[Link].2020.100850.
[6] Charrad M, Ghazzali N, Boiteau V, Niknafs A. NbClust: an R package for
determining the relevant number of clusters in a data set. J Stat Software 2014;61:
1–36. [Link]

8
R. Granillo-Macías Socio-Economic Planning Sciences xxx (xxxx) xxx

Applications. Expert Syst Appl 2013;104:244–60. [Link] [34] Yawar SA, Seuring S. Management of social issues in supply chains: a literature
eswa.2013.07.078. review exploring social issues, actions and performance outcomes. J Bus Ethics
[33] Wfp, Word Food Programme. The impacto of school feeding programmes [On- 2017;141(3):621–43. [Link]
line]. Available in: [Link]
rogrammes; 2019.