Computers & Industrial Engineering 67 (2014) 223–233

Contents lists available at ScienceDirect

Computers & Industrial Engineering

journal homepage: www.elsevier.com/locate/caie

The impact of food perishability issues in the vehicle routing problem q

P. Amorim ⇑, B. Almada-Lobo
INESC TEC, Faculdade de Engenharia, Universidade do Porto, Rua Dr. Roberto Frias, s/n, 4600-001 Porto, Portugal

a r t i c l e i n f o a b s t r a c t

Article history: Highly perishable food products can lose an important part of their value in the distribution process. We
Received 27 October 2012 propose a novel multi-objective model that decouples the minimization of the distribution costs from the
Received in revised form 30 October 2013 maximization of the freshness state of the delivered products. The main objective of the work is to exam-
Accepted 10 November 2013
ine the relation between distribution scenarios and the cost-freshness trade-off. Small size instances
Available online 23 November 2013
adapted from the vehicle routing problem with time windows are solved with an -constraint method
and for large size instances a multi-objective evolutionary algorithm is implemented. The computational
Keywords:
experiments show the conflicting nature of the two objectives.
Routing
Food perishability
Ó 2013 Elsevier Ltd. All rights reserved.
Freshness
Multi-objective optimization

1. Introduction meals are ordered on the day before and customers are very
demanding in terms of expected delivery time and freshness of
Product perishability may manifest itself in a whole set of dif- the food received. Hence, if a customer orders duck with sauce,
ference forms. Products subject to perishability range from daily he is expecting that when he receives the meal, it will seem as
newspapers that lose their value soon after the day they are report- the sauce was just made and poured above the duck. On the other
ing, to flowers that look wilted sometimes even before reaching hand, if the customer orders an assortment of cheeses, he will be
stores, or blood used for transfusions. This last example sparked less sensible to how much time the product stayed in the truck be-
the study of the perishable inventory (Millard, 1960). Decision fore reaching him.
models valid for this broad range of perishable products should This operational distribution planning task fits into the vehicle
capture their specific nature distinguishing, for example, between routing problem (VRP) class of problems. In specific, we are deal-
products with and without best before dates (Amorim, Meyr, Alme- ing with a VRP with time windows (VRPTW) that has to consider
der, & Almada-Lobo, 2011). Highly perishable products have an the perishable nature of the products delivered. This hard prob-
important role in the operational distribution process, particularly lem (Savelsbergh, 1985) is to be modelled using a multi-objective
in the vehicle routing planning task. Examples of this kind of framework in which distribution costs are minimized and the
products are fruits, vegetables and prepared meals. In this category freshness of the products delivered to the customers is
of goods, the quality changing imposed by the perishability phe- maximized simultaneously. The first objective reflects explicitly
nomenon is noticeable by the entity receiving the products during the need of reducing operation costs and the second one
the planning horizon. Hence, in this paper, the decreasing value expresses the intangible customer value stemming from product
that customers attribute to a decreasing freshness state is freshness, which the company wants to grasp when designing
acknowledged. their routes.
Let us focus on the prepared meals segment to exemplify the In order to investigate the impact of product perishability in the
particularities of this vehicle routing problem that delivers highly distribution process a set of empirical hypotheses are postulated,
perishable goods. Consider a company specialized in gourmet pre- relating distribution scenarios and the cost-freshness trade-off.
pared meals based on a very busy city that services daily hundreds The customers’ typology, time windows width and perishability
of customers. Moreover, this company runs an own fleet of refrig- intensity are varied through these distribution scenarios. To test
erated trucks to perform the distribution of the products. The these hypotheses an -constraint method is employed to solve ex-
actly the well-known Solomon instances (Solomon, 1983) with 25
q
customers. Afterwards, the findings are validated for the instances
This manuscritp was processed by Area Editor Iris F.A. Vis.
⇑ Corresponding author. Tel.: +351 22 508 1735. with 100 customers using a multi-objective evolutionary algo-
E-mail addresses: [email protected] (P. Amorim), [email protected]
rithm (MOEA) as the exact methods fail to generate feasible
(B. Almada-Lobo). solutions.

0360-8352/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.cie.2013.11.006
224 P. Amorim, B. Almada-Lobo / Computers & Industrial Engineering 67 (2014) 223–233

The remainder of this paper is as follows. In the next section a the demand through a probability density function. Afterwards,
brief literature review is performed, and the mathematical formu- the integrated model is solved in an iterative scheme in which
lation that models this problem is then presented in Section 3. Sec- the VRP part is solved by a constructive heuristic followed by an
tion 4 formulates the hypotheses that establish the possible improvement one.
influences of distribution scenarios in the cost-freshness trade- In Table 1 we compare our work against the closest papers in
off. In Section 5 the methodology used to test the hypothesis for the literature, in terms of modelling characteristics (type of VRP
small and large instances is described. Afterwards, in Section 6 considered, number of objectives, perishability behaviour and
the results obtained through the computational experiments are number of products), solution methods and instances tested.
shown. Finally, the Conclusions section resumes the main findings From the literature review, it is clear that incorporating the per-
of this work and gives some hints for future research. ishability factor explicitly in the formulations seems to be of great
advantage (Akkerman, Farahani, & Grunow, 2010) since the cus-
tomers’ point of view is also taken into account. In our work, a mul-
2. Literature review ti-objective framework is used to tackle this phenomenon and,
hence, to give to the decision maker a whole set of equally efficient
The VRP field of research is very vast and proficuous. The same solutions, evidencing the trade-off between supply chain optimiza-
applies for the VRPTW extension. In this review, the focus will be tion and customer service related to the freshness aspect. Further-
kept on papers dealing with the VRP for perishable goods. The more, in the experiment design, the goal is to provide new insights
readers are referred to Laporte (2007) or Eksioglu, Vural, and Reis- into the relation between distribution scenarios and the aforemen-
man (2009) for more general VRP problems. tioned trade-off.
Some literature focus on different distribution problems related
with perishable food products but without considering explicitly 3. Mathematical formulation
the degradation of quality (losing of freshness) during transporta-
tion. In fact, these models could be most of the times applied to This section aims to present a formal definition of the Multi-
products without a perishable nature. The work of Tarantilis and Objective Vehicle Routing Problem with Time Windows dealing
Kiranoudis (2001) concentrates on the distribution of fresh milk with Perishability (MO-VRPTW-P). The formulation and notation
and formulates the problem as a heterogeneous fixed fleet VRP. is based on the VRPTW formulation proposed in Cordeau, Desaul-
The same authors (Tarantilis & Kiranoudis, 2002) solve a real- niers, Desrosiers, Solomon, and Soumis (2001).
world distribution problem of fresh meat as a multi-depot VRP. The MO-VRPTW-P trades-off the optimal design of routes and
Faulin (2003) implements a hybrid algorithm procedure that uses the freshness state of the delivered products. A set K of identical
a combination of heuristics and exact algorithms to find a solution fixed capacity vehicles indexed by k ¼ 1; . . . ; m initially located at
to a VRP with constraints enforcing narrow time windows and a depot are available to deliver perishable food goods to a set N
strict delivery quantities. According to the author, these delivery of customers i; j ¼ 1; . . . ; n through a set of arcs A. The number of
scenarios are usually the case in the agribusiness industry. vehicles m is enough to always guarantee a feasible solution. The
Concerning the articles modelling perishability explicitly, Os- VRPTW structure can be defined on a direct graph G ¼ ðV; AÞ with
vald and Stirn (2008) extend a heuristic proposed by the first V ¼ N [ f0; n þ 1g, where the depot is simultaneously represented
author in a previous work to solve the problem of distributing fresh by the two vertices 0 and n þ 1 and, therefore, jVj ¼ n þ 2. Each
vegetables in which perishability represents a critical factor. The possible arc ði; jÞ has an associated time and cost that is related
problem is formulated as a VRPTW with time-dependent travel to the euclidean distance of the vertices that it connects to. Each
times. The objective function minimizes the distance and time customer has a demand that needs to be satisfied for a certain
travelled, the delay costs for servicing late a customer and the costs number of products. Without loss of generality, these products
related to perishability. In this model, the perishability costs are are of identical size, and they have different deterministic shelf-
calculated by multiplying the load transported in each arc by the lives. It is assumed that as soon as the vehicle departs the depot,
time needed to do it. Hsu, Hung, and Li (2007) consider the all products that it is carrying are at their maximum freshness.
randomness of the perishable food delivery process and present a Moreover, customers want their requests available within a strict
stochastic VRPTW model that is further extended to consider time window and they need a certain time to be served. From a
time-dependent travel times. The objective function of this work modelling point of view we will just account for the shelf-life of
is very similar to that of Osvald and Stirn (2008), but the calcula- the products within a customer order that deteriorate the most.
tion of costs due to perishability is done in a stochastic manner. Thus, by maximizing the freshness of the products delivered, it is
The authors attribute probability density functions to determine ensured that the worst case is tackled for each customer and,
the chances of having spoiled products due to the opening of the therefore, all the other products still have some remaining shelf-
vehicle door and to the travel time. The problem is solved by a heu- life when delivered.
ristic procedure. Chen, Hsueh, and Chang (2009) integrate produc- A feasible solution for this problem implies a collection of
tion scheduling with the VRPTW for perishable food products. In routes that correspond to paths starting at vertex 0 and ending
the distribution part, they consider a value decay on the products at vertex n þ 1. These routes have to ensure that each customer
distributed that will influence the price paid by the retailer to is visited exactly once satisfying simultaneously its demand and
the transporter. This model has a stochastic nature by defining time window. Furthermore, it is not admissible that any of the

Table 1
Comparison of the VRP related works dealing with perishability.

Author VRP-type Number of Obj. Perishability behaviour Products Solution method Instances
Chen et al. (2009) STW 1 Random Multiple Heuristic Solomon (1983)
Osvald and Stirn (2008) HTW; TD 1 Deterministic Single Heuristic Solomon (1983)
Hsu et al. (2007) STW; TD 1 Random Single Heuristic Tailor-Made
This paper HTW 2 Deterministic Multiple MOEA Solomon (1983)

STW: soft TW, HTW: hard TW, TD: time dependent travel time.
 P. Amorim, B. Almada-Lobo / Computers & Industrial Engineering 67 (2014) 223–233 225

products delivered to a customer is spoiled and the cumulative de-

mand that each vehicle serves cannot exceed its capacity.
Consider the following indices, parameters, and decision
variables.

Indices
k vehicles
i; j vertices
ði; jÞ arcs

Parameters
C vehicle capacity Fig. 1. Representative scheme of the freshness objective function.
si service time of customer i
The objective function (1) minimizes the total routing cost, which is
cij ðtij Þ travel cost (time) from customer i to customer j
the usual objective in the VRP related problems. The objective func-
ai starting time of time window of customer i
tion (2) maximizes the average freshness of all requests by stan-
bi finishing time of time window of customer i
dardizing the remaining shelf-life of the most deteriorating
di demand of customer i
product of each customer (see expression (10)). The value of this
sli shelf-life of the most perishable product contained in
objective varies between 0% and 100%, where 100% corresponds
the request of customer i
to the maximum possible freshness. We assume that when a vehicle
Decision variables leaves the depot the freshness of the products inside is at its max-
xkij equals 1 if arc ði; jÞ is crossed by vehicle k, 0 otherwise imum. Naturally, we are assuming that all the upstream supply
time at which vehicle k starts servicing vertex i, 0 if chain planning is based on the distribution plans and follows the
wki
vertex i is not visited by vehicle k Just-in-Time (JIT) philosophy. Thus, in case customer i is serviced
fr i freshness level for customer i request upon delivery by vehicle k, the freshness of the product delivered to the customer
is obtained by subtracting to the maximum shelf-life, sli , the time it
takes to start servicing customer i, given by the term wki  wk0 . Divid-
ing this value by sli the freshness is converted to a value that is be-
In the preprocessing of the arcs, besides eliminating arcs ði; jÞ
tween 0% and 100%. In the second case, when the customer i is not
due to temporal reasons (in case ai þ si þ t ij > bj ), and to capacity
serviced by vehicle k, constraints (10) are loose (in these constraints
reasons (in case di þ dj > C), arcs are also eliminated due to perish-
we could use the big number P equal to b0  a0 ). In Fig. 1, a visual
ability reasons (in case si þ t ij > slj ).
representation of this objective function is given.
Let dþ ðiÞ ¼ fj : ði; jÞ 2 Ag and d ðjÞ ¼ fi : ði; jÞ 2 Ag denote the set
Constraints (3) state that each customer is visited exactly once,
of successors and predecessors of i and j, respectively. The mathe-
while constraints (4)–(6) ensure that each vehicle is used and that
matical formulation for the MO-VRPTW-P can be stated as follows:
flow conservation is satisfied at each customer vertex. The consis-
MO-VRPTW-P tency of the time variables wki is ensured through constraints (7),
XX while the time windows are imposed by (8). Again, these last con-
minf 1 ¼ cij xkij ð1Þ straints also enforce wki to be 0 in case customer i is not visited by
k2K ði;jÞ2A
vehicle k. Regarding the vehicle capacity, constraint (9) enforces it
1X to be respected. Finally, constraints (11) ensure that customer re-
max f 2 ¼ fr ð2Þ
N i2N i quests are satisfied with products that are not spoiled (i.e. that still
have some degree of freshness).
subject to :
XX
xkij ¼ 1 8i 2 N ð3Þ Remark 1. Formulation (1)–(12) is nonlinear because of con-
k2K j2dþ ðiÞ straints (7) and (11). The first constraint can be linearized as
X follows:
xk0j ¼ 1 8k 2 K ð4Þ
j2dþ ð0Þ  
X X wkj P wki þ si þ t ij  Mij 1  xkij 8k 2 K; ði; jÞ 2 A; ð13Þ
xkij  xkji ¼ 0 8k 2 K; j 2 N ð5Þ
i2d ðjÞ i2dþ ðjÞ
where Mij ¼ maxf0; bi þ si þ t ij g is a constant. We have changed the
X calculation of this parameter from the original formulation of Cor-
xki;nþ1 ¼ 1 8k 2 K ð6Þ deau et al. (2001). In that formulation M ij is equal to
i2d ðnþ1Þ maxf0; bi þ si þ tij  aj g. However, for the case where a vehicle does
  not pass neither at i nor at j then this parameter would force aj 6 bi
xkij wki þ si þ t ij  wkj 6 0 8k 2 K; ði; jÞ 2 A ð7Þ
since both wki and wkj would be 0 due to constraints (8). Clearly, this
X X equation does not hold for every instance and so the overall model
ai xkij 6 wki 6 bi xkij 8k 2 K; i 2 V ð8Þ
j2dþ ðiÞ j2dþ ðiÞ
becomes ill-defined.
XX The linearization of constraints (11) is as follows:
di xkij 6 C 8k 2 K ð9Þ
þ
wk0 þ sli  wki P 0 8k 2 K; i 2 N: ð14Þ
i2N j2d ðiÞ
 P 
wk0 þ sli  wki þ P 1  j2dþ ðiÞ xkij
fri 6 8k 2 K; i 2 N ð10Þ 4. Relations between distribution scenarios and the cost-
sli
X   freshness trade-off
xkij wk0 þ sli  wki P 0 8k 2 K; i 2 N ð11Þ
j2dþ ðiÞ From the modelling of the MO-VRPTW-P we predict that to ob-
tain an increase in the freshness of the products delivered, a higher
xkij 2 f0; 1g; fr i ; wki P 0: ð12Þ
distribution cost needs to be incurred. Moreover, the behaviour of
226 P. Amorim, B. Almada-Lobo / Computers & Industrial Engineering 67 (2014) 223–233

the relation between freshness and distribution cost probably de- the relative cost increase for each freshness level when increasing
pends on various factors, such as customers typologies, time win- perishability intensity.
dows width that each customer require to be served and the To understand how this method works let us first define P1 ð2 Þ
perishability intensity of the food products delivered. as:
In order to organize and to guide the computational experi-
min f 1
ments needed to understand the relation between the cost-fresh-
ness trade-off and distribution scenarios, a set of intuitive subject to :
hypotheses were raised covering the aforementioned factors influ- ð3Þ—ð6Þ
encing the problem under study. Fig. 2 guides the following ð8Þ—ð10Þ
hypotheses. ð13Þ and ð14Þ
The first two more general hypotheses are as follows:
f 2 P 2 ð15Þ
H 1. For a given scenario, the distribution cost incurred to satisfy
an increasing freshness level increases in a non-decreasing fashion. xki;j 2 f0; 1g; fri ; wki P 0; ð16Þ

where f 1 and f 2 denote the expressions of the objective functions

H 2. For a given scenario, the number of vehicles used to satisfy an (1) and (2), respectively.
increasing freshness level increases in a non-decreasing fashion. Second, we shall define z2u and z2l as the theoretical highest and
The third hypothesis is related to the general influence of the lowest possible values for the freshness objective (f 2 ). z2u is attained
perishability intensity: in case a vehicle is allocated to each customer and there is no wait-
ing time at the customer site:
H 3. For each freshness level, the distribution cost incurred in a
1 X sli  t 0i
scenario where products are subject to Low perishability is less or z2u ¼ : ð17Þ
equal to a scenario where products are subject to high perishability N i2N sli
(everything else held constant).
As for z2l , we assume that the vehicle departs as soon as possible
The next hypothesis focuses on the influence of time windows and serves each client at the latest possible point in time:
regarding the relation between distribution costs and freshness:
1 X maxð0; a0 þ sli  bi Þ
z2l ¼ : ð18Þ
H 4. On average, when increasing the perishability intensity from N i2N sli
one scenario to another (everything else held constant), the
Algorithm 1 shows the procedure to get the Pareto front F for
distribution cost to satisfy the same freshness level will be higher
the MO-VRPTW-P using the -constraint method.
in a setting with customers having narrow time windows com-
pared with customers having large time windows.
Algorithm 1. Pseudo-code to find the Pareto front with the -
Finally, the last hypothesis reflects our intuition about the influ- constraint method
ence of customers spatial organization on the studied interaction.

H 5. On average, when increasing the perishability intensity from

2 ¼ z2u
one scenario to another (everything else held constant), the while 2 P z2l
distribution cost to satisfy the same freshness level will be higher Solve P1 ð2 Þ with exact solver
in a setting with customers randomly dispersed compared with F SolutionofP1 ð2 Þ
customers organized in clusters. 2 ¼ 2  0:05
end while
return F
5. Methodology

Before looking into the details of the algorithms used to solve

this problem, we emphasize that the main goal of this paper is to 5.2. Large size instances
understand the relationship between distribution cost and fresh-
ness for different scenarios in this operational supply chain plan- To solve the MO-VRPTW-P for larger instances we have adapted
ning problem. Nevertheless, to conduct the study it is mandatory an approach based on that presented in Amorim, Günther, and
to solve with efficacy and a reasonable efficiency the underlying Almada-Lobo (2012), which proved to be successful in solving a
hard problem. lot-sizing and scheduling problem also concerned with perishable
In order to test the hypotheses we need to be able to ensure that products. This approach hybridizes a MOEA with a mixed integer
each solution found is very close to the Pareto optimal front. To linear solver. The underlying MOEA is the Non-dominated Sorting
achieve that, an exact method is employed to solve this problem Genetic Algorithm (NSGA-II), which has been successfully used in a
to optimality for small size instances. Afterwards, the results are panoply of different problems and with competitive results com-
generalized for large instances of a size closer to real-world prob- pared to other MOEAs (Khare, Yao, & Deb, 2003). The readers are
lems, by employing an approximate method. referred to Deb, Agrawal, Pratap, and Meyarivan (2000) for the spe-
cific details of NSGA-II. In Fig. 3, the outline of the used algorithm is
5.1. Small size instances depicted.
The operations performed in the main building blocks of this
In order to solve the small size instances, an -constraint meth- MOEA are based on the work of Ombuki, Ross, and Hanshar
od is used by adapting the algorithm proposed by Berube, Gend- (2006). In this work the authors were able to develop a MOEA to
reau, and Potvin (2009) to solve bi-objective combinatorial solve the VRPTW that proved to be very competitive in terms of
problems with integer solutions. The properties of this exact meth- solution quality when comparing to the results of other works.
od will be important in the computational experiments to assess Hence, for the sake of completeness, we will briefly describe our
 P. Amorim, B. Almada-Lobo / Computers & Industrial Engineering 67 (2014) 223–233 227

step 1. We note that perishability concerns are not taken into

account yet. In Fig. 4 these steps and encoding scheme are illus-
trated for an instance with 9 customers.

(b) Initializing a Population The initial population is created in

two different ways. 90% of the chromosomes are generated
based on random permutations of the total customer nodes.
The remaining 10% are generated with the nearest customer
insertion scheme.
(c) Genetic Operators The first two genetic operators are
derived from the work of Ombuki et al. (2006) and the last
one comes from the NSGA-II framework.

 Crossover The Best Cost Route Crossover is implemented. This

operator consists of two distinct steps. Firstly, a route from each
Fig. 2. Hypotheses framework for studying the influence of perishability scenarios
in the cost-freshness trade-off. parent is randomly chosen and the respective customers are
removed from the opposite parent. Secondly, the best possible
insertion of the removed customers is performed while ensur-
adaptation of the algorithm. However, for a complete explanation ing feasibility. The order of the missing customers to insert is
the readers are referred to the cited references (Amorim et al., chosen randomly. If there is no feasible insertion, then a new
2012; Ombuki et al., 2006). route is created. Fig. 5 shows an example of this operator for
an instance with 9 customers.
(a) Representation of a Chromosome and Routing Scheme In  Mutation The Constraint Route Inversion Mutation operator is
this algorithm each chromosome has a length equivalent to the designed to help the search escaping from local optima with
number of customers in a given instance. Each gene contains a the minimum possible number of infeasibilities created in the
customer and the gene sequencing represents the order in procedure. Fig. 6 shows how this operator works. After choosing
which they are visited. So a chromosome is a collection of randomly a route from a chromosome, two consecutive custom-
routes containing all clients. The only decision left is the assign- ers are selected and inverted. If this inversion results in a feasi-
ment of a sequence of genes to a given route. For that end, a ble chromosome, then it is maintained, otherwise it is rejected.
two-step routing scheme translates the input chromosome into  Selection/Reproduction After the offspring population is created,
a cluster of routes. the whole population is sorted according to nondominance
ranking. The new parent population is formed by adding solu-
The first step is responsible to assign each customer to a route tions from the better fronts until the number of chromosomes
that appears in the chromosome by the order that they are, ensur- exceeds the size of the population. The solutions of the last
ing feasibility regarding the vehicle capacity and the customers’ accepted front are sorted according to the crowding operator,
time windows. In the second step, a removal and insertion opera- in order to have a well-dispersed set of solutions, which are
tor sweeps the solution found in step 1 and tries to reallocate the added until the population size is reached (see Fig. 3). The
last customer visited in each route as the first customer visited in crowding operator ranks the chromosomes of the last front
the following one. If this movement proves to be better in terms of based on their relative solution distance through a measure
cost then it is performed, otherwise the chromosome is kept as in estimating a solution perimeter around its neighbours.

Fig. 3. Outline of the hybrid multi-objective GA to solve the MO-VRPTW-P.

228 P. Amorim, B. Almada-Lobo / Computers & Industrial Engineering 67 (2014) 223–233

Sub MO-VRPTW-P (X)

1X
max fr ð19Þ
N i2N i
subject to :
c
wki þ si þ tij  wkj 6 M ij ð1  X kij Þ 8k 2 K; ði; jÞ 2 A ð20Þ
 
P c
wk0 þ sli  wki þ P 1  j2dþ ðiÞ X kij
Fig. 4. Representation of a chromosome and routing scheme. fr i 6
sli
8k 2 K; i 2 N ð21Þ
ð14Þ
fr i ; wki P 0: ð22Þ
This approach seems to be worthwhile in case the model is to be
applied into a real-world scenario where more unpredictable con-
straints may appear, for example, having compartments for differ-
ent types of products in each vehicle. Thus, by coding the related
auxiliary constraints in the LP model a whole set of related prob-
lems can be solved.
When the chromosome is infeasible, constraints (14) are re-
laxed by introducing slack variables v i and the respective new
P
Fig. 5. Best Cost Route Crossover example.
model is solved attempting to minimize f 3 ¼ i v i (c.f. Fig. 3). This
violation v i represents the amount of time that spoiled products
passed their shelf-lives in the current solution. By preferring chro-
mosomes with better freshness conditions (even if with some
spoiled products) we expect that in the following generations the
(d) Fitness of a Chromosome and Infeasible Chromosomes operators guide the search into the freshness feasible domain.
The fitness of each chromosome in our solution method is
dependent on two distinct states: feasible or infeasible. Infeasi-  Alternative objective function In order to avoid to solve recur-
bility may only occur regarding the spoilage state of the prod- sively an LP model each time a chromosome is evaluated, this
ucts delivered, since our representation always entails a section presents a less flexible, but potentially faster procedure
chromosome satisfying the vehicles capacity and the custom- to evaluate the freshness-related objective.
ers’ time windows. Hence, if the chromosome is feasible we
should be able to evaluate its fitness by means of the two objec- Recall that a chromosome already fixes routes that comply with
tive functions: cost and freshness. The cost function can be cal- the customer time windows, capacity and demand. Nevertheless,
culated directly from the representation just by using the even for a chromosome that is able to deliver every product with-
routing scheme already explained to cluster the genes in routes. out spoilage, it is rather hard to compute directly the freshness
As for the freshness objective, we propose and study two alter- objective because there are several degenerated solutions for the
natives that influence the way unfeasibility is handled. A com- same value of freshness in a given set of routes. In order to over-
mon characteristic is that the amount of infeasibility is defined come this situation, we take advantage of the fact that for a fixed
as an additional objective (f 3 ) that must be minimized. The two set of routes, the controllable loss in product freshness comes from
alternatives are presented below: the waiting time to serve a customer.
 LP solver In this first approach we use the information coded in Let us define model MINWAIT (X) that tries to minimize the
the chromosomes to solve a LP model that takes into account waiting time of the vehicle, setting simultaneously the departure
the time windows and capacity feasibility of the input routes. of the vehicle from the depot as late as possible.
Let the parameters X kij equals 1 in case arc ði; jÞ is crossed by MINWAIT (X)
vehicle k, and zero otherwise. These parameters are fixed for XX X n o X
min max 0; ðai  ðwkj þ sj þ t ji Þ X kji  wk0 ð23Þ
each chromosome and they flow straightly from the values of
k2K i2N j2d ðiÞ k2K
the decision variables xkij . Moreover, we take advantage of the
chromosome information to set the correct bounds for the wki
variables. This means that if customer i is not visited by vehicle
k, then wki is set to zero. Otherwise, its bounds are limited
according to the time window of the corresponding customer
(i.e. ai 6 wki 6 bi ). The following LP is solved for each chromo-
some in order to find the maximum freshness for a given set
of fixed routes.

Fig. 6. Constraint Route Inversion Mutation example. Fig. 7. Pareto front for instance C101 with 50% of perishability and vehicles used.
 P. Amorim, B. Almada-Lobo / Computers & Industrial Engineering 67 (2014) 223–233 229

subject to : wkp0 bp0

ð20Þ and ð21Þ for u ¼ p  1 ! 0 do
ð24Þ
ð14Þ wku0 wkðuþ1Þ0  su0  tu0 ;ðuþ1Þ0
wki P 0: end for
end if
end if
end for
Proposition 1. For a given input set of routes (given by X),
for p ¼ 1 ! NumberOfCustomersOfRouteðkÞ
maximizing freshness is equivalent to simultaneously minimizing  
waiting time and maximizing the departure time at which each vehicle CustFreshp wk0 þ slp0  wkp0 =slp0
leaves the depot. if CustFreshp < 0 then
f3 f 3 þ CustFreshp
Proof 1. We show the proposition by proving that models Sub FlagViolation ¼ TRUE
MO-VRPTW-P (X) and MINWAIT (X) lead to the same optimal solu- else
tion. Since the solution domains of both models are equivalent, one Freshness Freshness þ CustFreshp
just needs to be able to transform the objective function of Sub end if
MO-VRPTW-P (X) into that of MINWAIT (X). end for
First, we make use of parameter X kij to rewrite the objective end for
function (19) as follows: if FlagViolation ¼ FALSE then
 k P return Freshness=NumberOfCustomers
k
1 XX w0 þ sli  wi
k
j2dþ ðiÞ X ij else
max ð25Þ
N k2K i2N sli return f 3
end if
By removing some scalars and constants, and taking into account
the balance of the flows, (25) can be further reduced into:
!
XX  X 6. Computational experiments
min wk0 þ wki X kji : ð26Þ
k2K i2N j2d ðiÞ In order to perform the computational study about the influ-
Rewriting wki
as a function of its predecessor, allows us to replace ence of the distribution scenarios in the MO-VRPTW-P, we have
(26) by the expression: adapted the instances developed in Solomon (1983) for the
VRPTW. Four different instances types are considered: C1, C2, R1
XX X   
min wk0 þ maxfai ; wkj þ sj þ t ji g X kji : ð27Þ and R2. In these instances, the travel cost and time of each arc is
k2K i2N j2d ðiÞ equal to its length. Type C problem type refers to clustered custom-
ers whose time windows were generated based on a known solu-
It is easy to see that (27) is equivalent to (23). h tion, whereas type R to customers located randomly generated
The theoretical insight of Proposition 1 is important for the per- over a square. Then, sets of type 1 contain narrow time windows
formance of the overall solution method because MINWAIT (X) can and small vehicle capacity. Sets of type 2 have large time windows
be solved by a simple heuristic (see Algorithm 2) that finds the and large vehicle capacity. Therefore, the solutions of type 2 prob-
appropriate wki by minimizing waiting time and, thus, maximizing lems usually yield very few routes and significantly more custom-
freshness. In the description of the algorithm p0 ¼ gðpÞ denotes a ers per route.
function that returns customer i that is in position p of route k. The extension of these sets is performed in order to grasp the
Basically, this algorithm defines, sequentially, for each route the changes related to different perishability intensity scenarios,
wki values that imply a service to the customers as late as possible attributing to each customer order a shelf-life. Three different per-
in order to avoid waiting times. In case a entering customer needs ishability intensity scenarios are studied corresponding to Low,
to be serviced earlier than the solution is pointing to, a feedback Medium and High perishability. Hence, to each original instance,
mechanism synchronizes the customers upstream to be sure that there will be three corresponding new instances with a new
the customer inserted is serviced on time. Algorithm 2 returns parameter sli . The reference point used to set the different scenar-
the freshness value in case a chromosome is feasible, and the ios is the end of the time window at the depot in the original in-
amount of infeasibility incurred, otherwise. stance that corresponds to the end of the planning horizon.
Hence, the Low scenario is set in such a way that the product with
Algorithm 2. Pseudo-code to evaluate freshness in Alternative the lowest shelf-life is able to last throughout the planning hori-
objective function zon; the Medium means that the shelf-life of the products lasts
about 75% of the planning horizon and the High corresponds to
for k ¼ 1 ! NumberOfRoutes do 50%. Therefore, if in the original instance the depot closes at time
wk10 ¼ b10 1000, then for the Low scenario shelf-lives will be around 1000,
wk0 ¼ wk10  t 010 for the Medium scenario around 750 and for the Low scenario
around 500.
for p ¼ 2 ! NumberOfCustomersOfRouteðkÞ do
First we report results for the small size instances (25 custom-
wkp0 ¼ wkðp1Þ0 þ sðp1Þ0 þ t ðp1Þ0 ;p0
ers) and, afterwards, for the large size instances (100 customers).
if wkp0 6 ap0 then
wkp0 ap0 6.1. Small size instances
else
We have run Algorithm 1 on the 40 instances from Solomon for
if wkp0 P bp0 then
the two extreme perishability scenarios (Low and High), resulting
in 80 instances of 25 customers. The -constraint method was
230 P. Amorim, B. Almada-Lobo / Computers & Industrial Engineering 67 (2014) 223–233

Table 3
Results for H 4–H 5.

Test Average difference (%) p-Value

H4 C1-C2 43 0.00
R1-R2 34 0.00
H5 R1-C1 0 0.49
R2-C2 9 0.11

Table 4
Parameters of the MOEA.

Fig. 8. Pareto fronts for instances C101 with 50% and 100% of perishability Population size 300
intensity. Number of generations 350
Crossover rate 0.9
Mutation rate 0.1
applied to each instance with a maximum number of 20 iterations,
changing iteratively the freshness level with a 5% step (0%, 5%, . . .,
95%, 100%). Each MIP had a 10 min time limit and was solved on a
Cluster IBM eServer 1350 equipped with CPLEX 12.1. Out of the 80
instances we were able to obtain 34 instances with an optimality
gap bellow 10% for the Low scenarios and the lowest freshness le-
vel (the extensive results are presented in Appendix A). These are
the instances used to confirm the hypothesis and on which the sta-
tistical tests were performed.
Concerning the first hypothesis, H 1, which relates the conflict-
ing nature of the objectives, Fig. 7 shows the Pareto front for one of
the instances that is representative of the behaviour observed for
the other instances. Clearly, in order to achieve a higher freshness
level, higher distribution costs are incurred, confirming the
hypothesis (see detailed results for this instance in Appendix C).
Fig. 9. Pareto fronts of C1 instances for the 3 levels of perishability intensity.
In H 2 we intended to assess the impact of the number of vehicles
on achieving better freshness conditions of the products delivered.
Hence, for every step increase of the freshness level, the number of
vehicles used was calculated. Again, Fig. 7 is representative of the ity intensity was about 50% compared with 100%. Fig. 8 shows an
behaviour of all instances and it is noticeable the increasing usage example of the dominance of the Low perishability scenario over
of vehicles to satisfy a higher freshness level. In fact, to increase the High one for a representative instance. In Table 2 the average
from 60% to 90% the freshness level, it is necessary to multiply relative cost increase for all instances considered when shifting
by four the number of vehicles. from Low to High perishability is presented. This measure is crucial
Derived from the third hypothesis, the Pareto front from the to test the last two hypotheses (see Table 3).
Low perishability scenario should weakly dominate the one from To test H 4–H 5 we statistically assess the respective average
the High perishability scenario. In testing H 3 it was observed that differences calculated a priori. Table 5 shows these results. In order
within the 283 pairs of solutions analysed (a pair is found in the to understand the validity of H 4 the difference between instances
same instance comparing the cost for the same freshness level with narrow time windows (type 1) and large time windows (type
for perishability intensities of 100% and 50%), the cost for achieving 2) was calculated for the two typologies (C and R). The results
the same freshness level is systematically higher when perishabil- indicate a higher cost increase for type 1 customers than for type
2, when the perishability intensity augments. This difference is
statistically significant for both types of topologies. However,
Table 2 regarding H 5 no conclusions can be drawn. For both types of time
Percentage average cost increase for shifting from Low to the High perishability
intensity.
windows width no significant different is found across the
customer typologies.
Instance % Average cost increase
C101 58
C102 74
C105 59
C106 61
C107 62
C201 17
C202 43
C205 11
C206 18
C207 12
R101 16
R102 32
R105 25
R106 99
R111 123
R201 12
R202 46
Fig. 10. Pareto fronts of R1 instances for the 3 levels of perishability intensity.
 P. Amorim, B. Almada-Lobo / Computers & Industrial Engineering 67 (2014) 223–233 231

Fig. 11. Pareto fronts of C2 instances for the 3 levels of perishability intensity. Fig. 14. Pareto fronts of C2 and R2 instances having 100% and 50% levels of
perishability.

extensive results for the lowest freshness level are presented in

Appendix B). Based on preliminary tests and on the literature rec-
ommendations, we have defined the setting of the MOEA parame-
ters shown in Table 4. The computation time required for a
complete run of an instance is about 10 min on a PC with 1 GB of
RAM and an Intel Dual Core with 1.66 GHz running on Windows
XP. All the implementations were done in C++.
In Figs. 9–12 the aggregate results (solutions for all instances)
are illustrated for the C1, R1 and C2, R2 instances, respectively.
Space restrictions prevent us from showing all disaggregated re-
Fig. 12. Pareto fronts of R2 instances for the 3 levels of perishability intensity. sults, but for the instance C101 the complete results can be ob-
served in Appendix C. These graphs show unequivocally the
trade-off between delivering fresher products and the routing cost
involved in those solutions. Products subject to High perishability
Table 5 (50%) emphasize this trade-off and to deliver fresher products
Parameters of the MOEA.
more costs have to be incurred (validating hypothesis H 1). Fur-
Test Comparison interval (%) Average difference (%) p-Value thermore, as it is intuitive, products with larger shelf-lives relief
H4 C1–C2 100–75 16 0.00 the cost increase needed to achieve better freshness standards (H
75–50 3 0.32 3).
R1–R2 100–75 27 0.00 To test hypothesis H 4 and H 5 the same reasoning as for the
75–50 40 0.00
small instances is applied after rounding freshness levels. The only
H5 R1–C1 100–75 2 0.38 difference is that for the large instances three perishability inten-
75–50 29 0.01
sity scenarios are tested and, therefore, the statistical tests were
R2–C2 100–75 20 0.12
75–50 24 0.02 performed for convenient pairs of intensities (Between High and
Medium intensities and between Medium and Low intensities). Ta-
ble 5 summarizes the results that support the conclusions found
for the small instances. Once again no conclusions can be taken
regarding the difference between serving customers located ran-
6.2. Large size instances domly or in clusters. Concerning time windows width, the previous
conclusions were validated despite one of the tests giving statically
In this section the computational experiments are extended to insignificant (C1–C2 for Medium and High perishability intensity).
instances with 100 customers. Furthermore, in this section consid- Finally, Figs. 13 and 14 show the aggregate results, comparing
ers the three perishability scenarios (Low, Medium and High). now customer typologies. These results point out that the fresh-
All the instances are solved by the MOEA, which incorporates ness level attained when customers are organized in clusters is
Algorithm 2 to compute the freshness of each chromosome (the considerably higher than when geographically dispersed, despite
the fact that the average cost increase when subject to a higher
perishability intensity does not differ from random to clustered
customers (H 5).

7. Conclusions

In this paper, a new formulation for a vehicle routing problem

dealing with perishability issues was proposed based on a
multi-objective framework. The model distinguished two separate
objectives of the decision maker: the minimization of operational
tangible costs and the maximization of the freshness intangible va-
lue. The formulation of such vehicle routing problem is suitable in
Fig. 13. Pareto fronts of C1 and R1 instances having 100% and 50% levels of a setting where the perishable nature of the products imposes a
perishability. high pressure to the planner.
232 P. Amorim, B. Almada-Lobo / Computers & Industrial Engineering 67 (2014) 223–233

The main objective of the paper was to understand the relation- Acknowledgement
ship between different distribution scenarios and the cost-fresh-
ness trade-off. Therefore, guided by this aim, a set of hypotheses The first author appreciate the support of the FCT Project PTDC/
were formulated and tested for small and large size instances EGE-GES/104443/2008 and the FCT Grant SFRH/BD/68808/2010.
adapted from the work of Solomon (1983). The small instances This work was also partially financed by the ERDF – European
gave the statistical validity of the experiments as they were solv- Regional Development Fund through the COMPETE Programme
able close to optimality and the large instances explored the poten- (operational programme for competitiveness) and by National
tial of applicability of this approach in a real-world scenario. Four Funds through the FCT – Fundação para a Ciência e a Tecnologia
out of the five hypotheses that were formulated proved to be true (Portuguese Foundation for Science and Technology) within
testing the 25 customer instances. The main conclusions point out project «FCOMP-01-0124-FEDER-037281»; and by the Project
that, first, there is an evident trade-off between the mentioned «NORTE-07-0124-FEDER-000057», supported by the North Portu-
objectives; second, time windows have a strong impact on the gal Regional Operational Programme (ON.2 – O Novo Norte), under
freshness levels of products delivered, hence, large time windows the National Strategic Reference Framework (NSRF), through the
lead to less spoilage; finally, regarding customer typology no con- ERDF, and by national funds, through FCT.
clusions could be taken. For the large instances these conclusions
were partially validated. Appendix A. Results of the small size instances for the lowest
We use an exact method – the -constraint method – for the freshness level
small size instances and we propose two versions of a MOEA to
solve instances with 100 customers. While the first alternative Table A.6 presents the results of the small size instances for the
hybridizes the algorithm with a LP solver, the second alternative lowest freshness level.
speeds up the MOEA performance by making use of a tailored algo-
rithm to obtain the freshness value. Appendix B. Results of the large size instances for the lowest
Future work shall be devoted to test the proposed model on freshness level
real-world instances. Moreover, it could be interesting to under-
stand the effect of dropping the JIT assumption and integrate this Table A.7 presents the results of the large size instances for the
model with production scheduling decisions. lowest freshness level.

Table A.7
Results of the large size instances for the lowest freshness level.
Table A.6
Original Low Medium High Optimal value
Results of the small size instances for the lowest freshness level.
Instance perishability perishability perishability (without
Original Low High Optimal value (without perishability
instance perishability perishability perishability considerations) considerations)

C101 191.815 285.652 191.3 C101 1120.144 1466.272 1584.202 828.94

C102 190.739 296.522 190.3 C102 1391.02 1453.281 1521.605 828.94
C103 272.225 647.403 190.3 C103 1172.629 1304.301 1524.071 828.06
C104 265.324 749.027 186.9 C104 1146.617 1171.012 1373.171 824.78
C105 191.815 288.415 191.3 C105 1478.612 1489.434 1762.882 828.94
C106 191.815 285.652 191.3 C106 1379.509 1437.115 1471.821 828.94
C107 191.815 291.755 191.3 C107 1432.283 1553.687 1789.568 828.94
C108 227.457 317.889 191.3 C108 1114.458 1207.651 1462.652 828.94
C109 226.754 298.807 191.3 C109 1211.279 1256.342 1819.665 828.94
C201 215.542 247.286 214.7 C201 925.873 1153.568 1242.076 591.56
C202 215.542 334.461 214.7 C202 1410.592 1454.465 1550.613 591.56
C203 263.546 371.286 214.7 C203 1239.964 1291.021 1373.401 591.17
C204 297.409 327.227 213.1 C204 1186.162 1223.131 1223.39 590.6
C205 215.542 247.286 214.7 C205 1075.396 1094.072 1342.755 588.88
C206 226.566 272.473 214.7 C206 1100.845 1178.511 1828.105 588.49
C207 224.651 266.179 214.5 C207 1013.923 1055.17 1716.533 588.29
C208 247.286 284.928 214.5 C208 1048.461 1143.27 1177.117 588.32
R101 618.328 644.422 617.1 R101 1921.313 1937.303 1937.303 1645.79
R102 580.487 697.079 547.1 R102 1809.696 1880.717 2095.431 1486.12
R103 570.239 994.788 454.6 R103 1619.626 1688.332 2157.102 1292.68
R104 527.669 599.336 416.9 R104 1314.428 1362.315 1478.51 1007.24
R105 531.537 574.565 530.5 R105 1686.94 1827.805 1852.745 1377.11
R106 492.674 661.553 465.4 R106 1644.892 1731.886 1799.042 1251.98
R107 500.181 998.588 424.3 R107 1528.261 1653.655 1729.492 1104.66
R108 466.016 1140.49 397.3 R108 1187.583 1362.771 1745.908 960.88
R109 498.001 600.993 441.3 R109 1625.115 1635.282 1813.623 1194.73
R110 529.015 960.817 444.1 R110 1511.963 1554.342 1618.41 1118.59
R111 431.114 1006.338 428.8 R111 1415.812 1488.047 1503.684 1096.72
R112 506.604 634.598 393 R112 1190.576 1309.047 1529.147 982.14
R201 465.908 501.496 463.3 R201 1737.092 1793.26 1798.748 1252.37
R202 450.454 476.803 410.5 R202 1638.221 1680.157 1751.453 1191.7
R203 498.747 500.383 391.4 R203 1416.241 1511.296 1675.031 939.5
R204 297.409 498.102 – R204 1063.558 1126.592 1222.298 825.52
R205 215.542 215.542 393 R205 1577.541 1598.556 1628.979 994.42
R206 465.535 488.8 374.4 R206 1395.153 1402.674 1493.319 906.14
R207 439.776 459.136 361.6 R207 1292.192 1353.112 1400.031 890.61
R208 441.416 452.493 330.9 R208 1002.999 1019.859 1027.652 726.75
R209 442.906 450.945 370.7 R209 1457.2 1463.534 1500.708 909.16
R210 468.101 490.979 404.6 R210 1463.612 1513.986 1657.885 939.34
R211 414.248 452.669 350.9 R211 1160.76 1174.64 1179.269 885.71
 P. Amorim, B. Almada-Lobo / Computers & Industrial Engineering 67 (2014) 223–233 233

Table A.8 References

Results for instance C101 with 25 customers.
Akkerman, R., Farahani, P., & Grunow, M. (2010). Quality, safety and sustainability
Freshness Low perishability High perishability in food distribution: A review of quantitative operations management
approaches and challenges. OR Spectrum, 32, 863–904.
0.05 191.815 285.652
Amorim, P., Günther, H.-O., & Almada-Lobo, B. (2012). Multi-objective integrated
0.1 191.815 285.652 production and distribution planning of perishable products. International
0.15 191.815 285.652 Journal of Production Economics, 138, 89–101.
0.2 191.815 285.652 Amorim, P., Meyr, H., Almeder, C., & Almada-Lobo, B. (2011). Managing perishability
0.25 191.815 285.652 in production-distribution planning: A discussion and review. Flexible Services
0.3 191.815 285.652 and Manufacturing Journal, 1–25. http://dx.doi.org/10.1007/s10696-011-9122-
0.35 191.815 285.652 3.
0.4 191.815 285.652 Berube, J., Gendreau, M., & Potvin, J. (2009). An exact -constraint method for bi-
0.45 191.815 285.652 objective combinatorial optimization problems: Application to the traveling
0.5 191.815 285.652 salesman problem with profits. European Journal of Operational Research, 194,
0.55 191.815 285.652 39–50.
0.6 191.815 288.034 Chen, H.-K., Hsueh, C.-F., & Chang, M.-S. (2009). Production scheduling and vehicle
routing with time windows for perishable food products. Computers &
0.65 191.815 312.63
Operations Research, 36, 2311–2319.
0.7 213.981 322.397
Cordeau, J.-F., Desaulniers, G., Desrosiers, J., Solomon, M. M., & Soumis, F. (2001). In
0.75 244.106 623.799 VRP with time windows, the vehicle routing problem (pp. 157–193). Philadelphia,
0.8 315.52 725.177 PA, USA: Society for Industrial and Applied Mathematics.
0.85 558.493 774.665 Deb, K., Agrawal, S., Pratap, A., & Meyarivan, T. (2000). A fast elitist non-dominated
0.9 848.844 1030.55 sorting genetic algorithm for multi-objective optimisation: NSGA-II. London, UK:
Springer-Verlag.
Eksioglu, B., Vural, A. V., & Reisman, A. (2009). The vehicle routing problem: A
taxonomic review. Computers & Industrial Engineering, 57, 1472–1483.
Table A.9
Faulin, J. (2003). Applying MIXALG procedure in a routing problem to optimize food
Results for instance C101 with 100 customers.
product delivery. Omega, 31, 387–395.
Low perishability Medium perishability High perishability Hsu, C.-I., Hung, S., & Li, H.-C. (2007). Vehicle routing problem with time-windows
for perishable food delivery. Journal of Food Engineering, 80, 465–475.
Freshness Costs Freshness Costs Freshness Costs Khare, V., Yao, X., & Deb, K. (2003). Performance scaling of multi-objective
evolutionary algorithms. Berlin, Heidelberg: Springer.
0.891679 3090.261 3000.251 0.842288 0.721263 2393.622
Laporte, G. (2007). What you should know about the vehicle routing problem. Naval
0.892416 3385.463 3522.99 0.860898 0.636355 1796.341
Research Logistics (NRL), 54, 811–819.
0.832786 1724.243 1466.272 0.718217 0.713246 2161.125 Millard, D. (1960). Industrial inventory models as applied to the problem of
0.767717 1270.418 2994.42 0.841893 0.751483 2988.76 inventorying whole blood. In Industrial engineering analyses of a hospital blood
0.870628 2171.058 1768.467 0.776236 0.722531 2430.322 laboratory, engineering experiment station bulletin (Vol. 180). Ohio State
0.714888 1123.219 2909.771 0.839 0.637325 1799.676 University Columbus.
0.887905 2928.625 1998.149 0.801192 0.172254 1584.202 Ombuki, B., Ross, B., & Hanshar, F. (2006). Multi-objective genetic algorithms for
0.892509 3464.16 2130.266 0.816493 0.633405 1786.48 vehicle routing problem with time windows. Applied Intelligence, 24, 17–30.
0.878204 2725.69 2988.355 0.841626 0.710234 2100.389 Osvald, A., & Stirn, L. (2008). A vehicle routing algorithm for the distribution of fresh
0.875642 2416.213 2044.128 0.809999 0.701717 2057.35 vegetables and similar perishable food. Journal of Food Engineering, 85, 285–295.
0.880936 2763.288 3564.895 0.861648 0.693199 2029.395 Savelsbergh, M. (1985). Local search in routing problems with time windows.
0.832648 1720.367 1966.366 0.796436 0.641947 1803.283 Annals of Operations Research, 4, 285–305.
0.882927 2781.978 2064.492 0.812025 0.627686 1745.37 Solomon, M. M. (1983). Vehicle routing and scheduling with time windows
constraints: Models and algorithms. Ph.D. Dissertation. Department of Decision
0.885894 2851.714 3430.322 0.859763 0.718711 2332.54
Sciences, University of Pennsylvania, USA.
0.891844 3327.948 2966.418 0.841011 0.636237 1793.422
Tarantilis, C., & Kiranoudis, C. (2001). A meta-heuristic algorithm for the efficient
0.873662 2265.313 3070.223 0.84333 0.684099 1980.946
distribution of perishable foods. Journal of Food Engineering, 50, 1–9.
0.750275 1195.515 2897.874 0.838338 0.638953 1801.097 Tarantilis, C., & Kiranoudis, C. (2002). Distribution of fresh meat. Journal of Food
0.89218 3340.963 1711.491 0.769944 0.745506 2889.155 Engineering, 51, 85–91.
0.748091 1183.636 1725.927 0.770531 0.640963 1801.294
0.782951 1374.945 2104.077 0.814233 0.644259 1845.114
0.864845 2072.407 3548.852 0.861056 0.69365 2044.661
0.827882 1678.38 3041.362 0.84326 0.702811 2087.664
0.836353 1756.027 2163.777 0.818194 0.678187 1976.222
0.892467 3463.943 2875.937 0.837723
0.838115 1873.25 3320.34 0.854017
0.855154 2010.288 2896.223 0.837818
0.706549 1120.144
0.866974 2122.647
0.875202 2400.373
0.730586 1127.571
0.765485 1237.423
0.883213 2807.645
0.833953 1727.26
0.89244 3390.059

Appendix C. Results for instance C101 with 25 and 100

customers

Tables A.8 and A.9 present the results for instance C101 with 25
and 100 customers, respectively.