A Stochastic Inventory Routing Problem for Infectious

Medical Waste Collection

Pamela C. Nolz
WU Vienna University of Economics and Business, Institute for Production Management, 1090 Vienna, Austria

Ecole Nationale Supérieure des Mines de Saint-Etienne, CMP Georges Charpak, F-13541 Gardanne, France

Nabil Absi and Dominique Feillet

Ecole Nationale Supérieure des Mines de Saint-Etienne, CMP Georges Charpak, F-13541 Gardanne, France

We consider the problem of designing a logistic sys- 1. INTRODUCTION

tem to organize adequate collection of infectious medical
waste. Waste materials are produced by patients in self- This study is part of a project called Trace-de-TIC, funded
treatment, stored at pharmacies and picked up by local by the French Environment and Energy Management Agency
authorities for disposal. The problem is formulated as (ADEME). It is also connected to the PAC ID DASRI project,
a collector-managed inventory routing problem, encom-
passing stochastic aspects, which are common in such funded by the Provence-Alpes-Côte d’Azur (PACA) region in
problems. Social objectives, specifically the satisfaction France and the European Union. These two projects notably
of pharmacists and local authorities, as well as the min- investigate the use of radio frequency identification (RFID)
imization of public health risks, are considered for the technologies for infectious waste management.
real-world-motivated inventory routing problem. To opti- Infectious waste materials considered in these projects
mize the planning process for a predefined time horizon,
we take advantage of radio frequency identification tech- relate to waste materials disposed of by home health and
nology, which allows an improvement of the planning medical care services, as well as by self-attending patients.
process for pharmacists as well as local authorities. We Most of them, especially sharp objects such as contaminated
develop two solution approaches to optimize the deter- needles or syringes, need to be handled with great care to
mination of visit dates and corresponding vehicle routes.
We design a sampling method and an approach based on
prevent the spread of diseases. In view of the development
an adaptive large neighborhood search algorithm to treat of home care services and the aging of the population, one
the stochastic problem. The suggested approaches are can observe a fast increase in these types of waste [22].
tested on real-world data from the region of Provence- There is therefore a growing importance in securing the com-
Alpes-Côte d’Azur, in France. To evaluate the perfor- plete disposal process: the unsafe disposal of health-care
mance of the proposed solution methods, the results
for different waste collection scenarios are analyzed waste indeed poses major public health risks, considering that
and compared. © 2013 Wiley Periodicals, Inc. NETWORKS, health workers, waste handlers, or pedestrians might come
Vol. 63(1), 82–95 2014 into contact with contaminated waste if it is dumped into
areas without restricted access. As a matter of fact, the world
Keywords: medical waste collection; stochastic inventory rout-
ing; adaptive large neighborhood search; sustainable logistics
health organization recently established strict guidelines for
the management of infectious waste materials [33].
The purpose of the DASRI project was to improve the
traceability in the treatment of infectious waste, by devel-
oping automated RFID containers (Fig. 1b). With these
containers, located in pharmacies, dedicated boxes contain-
Received November, 2011; accepted April, 2012 ing infectious waste (Fig. 1a) can be identified and traced
Correspondence to: P. C. Nolz; e-mail: [email protected] until the final stage of the treatment. Within the Trace-de-TIC
Contract grant sponsor: French Environment and Energy Management project, and in this article, we investigate how new informa-
Agency (ADEME); Contract grant number: #10 02C 0003. tion provided by RFID can also help improving the planning
DOI 10.1002/net.21523
Published online 1 October 2013 in Wiley Online Library
of the disposals.
(wileyonlinelibrary.com). In the south-east region of France, the disposal of medical
© 2013 Wiley Periodicals, Inc. waste accumulated by patients in self-treatment is commonly

NETWORKS—2014—DOI 10.1002/net
 available space. Third, no clear vision on future collection
tasks is given to the collector. As this activity typically only
represents a limited part of his/her daily work, managing new
collection requests complicates the rest of his/her work.
Being equipped with RFID tags (Fig. 2b), boxes can auto-
matically be registered when deposited at a pharmacy without
disturbing the pharmacist. This provides two valuable pieces
of information to the collector. First, with RFID, the exact
number of boxes in containers can be accessed at any time.
Second, statistics on full-box arrival rates can be recorded.
Based on this information, a new collection policy can be
carried out, with waste collection tours planned a priori at
FIG. 1. Infectious medical waste containers. [Color figure can be viewed the beginning of the planning horizon for the whole period,
in the online issue, which is available at wileyonlinelibray.com.] depending on the actual number of boxes, the forecasted
arrival rates and the distances between pharmacies. We call
this policy the RFID scenario. Basically, this policy should
organized as follows (Fig. 2a). At pharmacies, patients help limiting the three inconveniences of the current pro-
receive empty boxes which are dedicated to the safe stor- cess. The flexibility in fixing the visit dates of pharmacies
age of their medical waste. Full boxes are brought back by makes the simultaneous optimization of collection days and
the patients to the pharmacies, where they are stored. When a routing possible. Having statistics on arrival rates and entire
certain storage capacity at a pharmacy is reached, a pharma- freedom to plan collections should permit avoiding most of
cist makes a call to the local authority, which organizes the the stock excesses. Possessing information on when collec-
pick-up of the waste material. Full boxes are then collected by tions of waste material will be carried out during the whole
a vehicle and deposited at collection centers for disposal. An planning period, collectors can adjust their time schedule and
external company periodically visits and depletes collection better organize their daily activities. Nevertheless, the whole
centers. process becomes less reactive as the planning is based on
Currently, a pharmacist makes a call to the local author- information available at the beginning of the time horizon.
ity when the maximal storage capacity (a container for full However, real-time information on inventory levels at each
boxes) at the pharmacy is reached. The pharmacy is then vis- pharmacy can be used each day to revise the planning of
ited in one of the two following days. The current system collection tours. This allows another collection policy to be
has three main inconveniences. First, transportation costs are carried out by revising solutions found by the RFID scenario
unduly high: as the number of boxes at pharmacies is only each day based on actual arrivals of data. We call this pol-
known by the collector when a call is received and as phar- icy Rolling horizon scenario. From a methodological point of
macists do not expect any visit before calling the collector, view, the Rolling horizon scenario does not add any complex-
no anticipation is possible. Hence, no optimization based on ity to the problem, as it is solved with the tools which have
geographical clustering can be introduced when designing been developed for the static case (RFID scenario). There-
collection routes. Second, the service at pharmacies is bad, fore, in the course of the article, we focus on the static RFID
not only for pharmacists but also for patients, and might raise scenario. Details concerning the Rolling horizon scenario
some risks: if a container is full, waiting to be emptied by the follow when computational experiments are presented.
collector, patients must have recourse to the pharmacist for Two conflicting aspects have to be considered when
finding another storage place, causing some disturbance for determining visit dates for pharmacies. On the one hand,
both of them and possibly some danger depending on the pharmacies should not be visited too late, as their storage

FIG. 2. Medical waste collection process. [Color figure can be viewed in the online issue, which is available at
wileyonlinelibray.com.]

NETWORKS—2014—DOI 10.1002/net 83
capacity will be exceeded, thus rising some risk and distur- of a customer is uncertain. Stockout costs at a customer are
bance as explained before. On the other hand, pharmacies modeled with a penalty function that contains fixed and vari-
should not be visited too early (when only few boxes are able components. Campbell and Hardin [8] study the problem
stored), as each visit disturbs people present in the pharmacy of minimizing the number of vehicles for periodic deliveries,
(including the pharmacists). which is motivated by IRPs considering customer demands
Our study contains several innovative and new aspects. at a steady rate. Campbell and Savelsbergh [9] present a
First of all, the problem formulation combines characteristics two-phase approach for an IRP. In the first phase, a delivery
of inventory routing and stochasticity in an applied research schedule is determined by solving an integer programming
study. Second, we consider a mix of four economic and model and in the second phase a set of delivery routes is
social performance indicators for the collection of infectious constructed by applying heuristics. Yu et al. [34] study an
medical waste, namely minimization of public health risks, IRP with split deliveries, allowing the delivery to each cus-
satisfaction of pharmacists, satisfaction of the local author- tomer in each period to be performed by multiple vehicles.
ity, and monetary cost of collection tours. Then, we provide Le Blanc et al. [19] present a collector-managed IRP for
a mathematical formulation of the waste collection problem reverse logistics, where the collection company is respon-
described above, addressed as a stochastic inventory routing sible for the inventory to be picked up and recycled. Solyali
problem (IRP). Finally, we develop several heuristic solution et al. [29] present a robust IRP with uncertain demands. The
methods, that we compare, and we analyze the performance authors determine the delivery quantities and delivery times,
of the new policies against the actual one on real-world data as well as the vehicle routes, and they allow backlogging
from the PACA region. of demand at customers. They propose mixed integer pro-
In Nolz et al. [24], the infectious medical waste collec- gramming formulations of the problem and its deterministic
tion problem was presented and a tabu search algorithm with variant.
preliminary results was detailed. In this article, we develop
two additional solution approaches, based on sampling and
2.2. Real-World Waste Collection Problems
on adaptive large neighborhood search (ALNS), respectively,
and present extensive computational experiments. Real-world waste collection problems have received a spe-
The remainder of the article is organized as follows. In cial interest in the last decade. Baptista et al. [3] present a
the next section, a literature review is presented. In section 3, periodic vehicle routing problem (PVRP) case study consid-
the problem is described in detail and a mathematical for- ering the collection of recycling paper containers in Portugal.
mulation is provided. In section 4, solution approaches are The number of visits to containers within a period is included
introduced and explained. Different computational experi- in the problem formulation as a decision variable. Nuortio et
ments are presented in section 5 to show the effectiveness of al. [25] address a stochastic PVRP with time windows for
our solution approaches and discuss managerial implications the collection of municipal solid waste in Eastern Finland,
of RFID. Concluding remarks are provided in section 6. which is solved with a metaheuristic. Kulcar [17] presents a
case study on waste collection in Brussels, evaluating sev-
eral means of collection, such as vehicle, rail, and canal.
2. LITERATURE REVIEW
Mourao and Almeida [23] examine a capacitated arc rout-
The infectious medical waste collection problem under ing problem for a refuse collection VRP. They present lower
investigation contains characteristics of inventory routing bounding techniques and a heuristic to solve a real-world
and stochasticity. It is furthermore embedded in an applied problem for the collection of household waste in Lisbon.
research project. This is why we partition the literature Teixeira et al. [30] provide a case studied on the collection of
review into three sections: IRP, real-world waste collection recyclable urban waste. Periodic vehicle routes are planned
problems, and stochastic vehicle routing problems. for the collection of three different waste types, regarding a
relatively long-time horizon. Tung and Pinnoi [32] studied
a vehicle routing and scheduling problem for waste collec-
2.1. Inventory Routing Problems
tion in Hannoi, Vietnam. The problem is divided into several
The IRP deals with the distribution (or collection) of phases, where waste has to be picked up and delivered at
a product to (from) several facilities over a given plan- different stages, while time windows have to be respected.
ning horizon (Bertazzi et al. [4]). In general, three different Alegre et al. [1] describe a PVRP for the pick-up of auto parts
decisions are addressed: when to visit each pharmacy, how in northern Spain regarding a long-time horizon. The authors
much to collect each time a pharmacy is served and how propose a metaheuristic solution approach based on scatter
to route the vehicle(s). The general objective of an IRP is search. Campbell et al. [7] present a case study for a real-
to minimize routing and inventory costs. For an overview world IRP. They develop a two-phase solution approach. In
on IRPs regarding different formulations and extensions, the first phase, they determine the customers to be served on a
the reader is referred to Bertazzi et al. [4]. Federgruen and specific day and the amount to be delivered, and in the second
Simchi-Levi [13] provide a mathematical formulation of the phase they determine the delivery routes and schedules. Shih
single-depot single-period IRP. Campbell et al. [6] study the and Chang [28] develop a computer system for the collection
IRP as well as the stochastic IRP, where the future demand of infectious hospital waste. Their solution approach for the

84 NETWORKS—2014—DOI 10.1002/net
considered PVRP consists of two phases. First, the authors in the PACA region, we consider a single vehicle for collec-
develop vehicle routes for all customers without considering tion tours. The vehicle is located at the depot, where it starts
allowable day combinations. In the second phase, they apply and ends each of its tours. At most, one tour is allowed every
a mixed integer programming method to assign the routes to day. The capacity of the vehicle is not considered, as being
days while respecting all constraints. Faccio et al. [12] intro- assumed to be sufficient to visit any possible subset of phar-
duce a multiobjective vehicle routing model considering real macies. Arrival of new boxes in pharmacies is modeled using
time traceability data for the collection of solid waste in Italy. random variables, dependent on the pharmacy and the time
period (day). The initial inventory level for each pharmacy
is exactly known; no constraint is imposed on the inventory
2.3. Stochastic Vehicle Routing Problems level at the end of the planning period.
In the following, the determination of visit dates and
Gendreau et al. [14] provide a review paper on stochas-
collection tours is performed once at the beginning of the
tic VRPs, referring to stochastic customers, stochastic travel
planning horizon. This problem corresponds to the RFID
times and stochastic demands. The same authors develop
scenario described in section 1.
a Tabu search heuristic called Tabustoch [15] for the VRP
Being equipped with RFID tags, boxes can automatically
with stochastic demands and customers. Laporte and Lou-
be registered when deposited at a pharmacy without dis-
veaux [18] present a two-index formulation of the SVRP
turbing the pharmacist. This provides two valuable pieces
and explain the integer L-shaped method for continuous
of information to the collector. First, with RFID, the exact
stochastic programming, which is based on a branch-and-cut
number of boxes in containers can be accessed at any time.
approach. Several papers focus on the solution of a stochastic
Second, statistics on full-box arrival rates can be recorded.
program with recourse. In the first stage, an a priori solution
Based on this information, a new collection policy can be
is determined, where the realizations of the random variables
carried out, with waste collection tours planned a priori at
are not known. In the second stage, a recourse or corrective
the beginning of the planning horizon for the whole period,
action can be taken, if for example the vehicle capacity is
depending on the actual number of boxes, the forecasted
exceeded. The aim is to determine a solution of minimum
arrival rates, and the distances between pharmacies.
expected total costs. Tricoire et al. [31], for example, formu-
The objective function is composed of two types of costs:
late a biobjective two-stage stochastic program with recourse
distribution and inventory. Distribution costs comprise rout-
for a covering tour problem with stochastic demand. They
ing costs and fixed costs incurred each time a tour is planned;
apply a branch-and-cut technique to a sample-average ver-
the latter relate to fixed extra-times involved by each tour for
sion of the problem within an epsilon constraint algorithm.
the collector. Inventory costs reflect the aim of targeting a
Hjorring and Holt [16] use the L-shaped method to solve a
certain range in the number of boxes actually picked up by
single VRP. A relaxed IP is solved and additional constraints
the vehicle when visiting a pharmacy. Unitary excess inven-
(optimality cuts) are added for a tighter approximation of
tory costs are counted when the container maximum storage
route failure costs. Rei et al. [27] propose a hybrid heuris-
capacity (number of full boxes) at a pharmacy is exceeded;
tic combining Monte Carlo sampling and local branching
these costs reflect the perturbance and risk that every box in
to solve a single VRP with stochastic demands. Mendoza et
excess provokes. A penalty cost is incurred when the number
al. [21] propose a memetic algorithm for the solution of a mul-
of boxes at a pharmacy is lower than a predefined threshold
ticompartment VRP with stochastic demands. The authors
inventory level; this cost reflects the difficulty for a pharma-
apply genetic operators in combination with local search
cist to accept a visit as long as a minimum storage level has
procedures to solve the stochastic program with recourse.
not been reached, because the visit perturbates the normal
Ak and Erera [2] present a paired-vehicle recourse strat-
functioning of the pharmacy. Regarding the final inventory
egy, considering the coordination of vehicles in pairs. A
level, unitary excess inventory costs are also considered, in
Tabu search heuristic is applied to real-world instances from
the same fashion as when a visit is planned, so as to avoid
Istanbul, Turkey. Christiansen and Lysgaard [10] introduce
delaying collection costs to future periods.
a branch-and-price algorithm for the capacitated VRP with
Note that the way inventory costs are modeled and the
stochastic demands. The authors formulate a set partitioning
values given to the different parameters in these costs (for the
problem and develop a dynamic programming approach for
PACA region) were defined in agreement with pharmacists,
the solution of the column generation subproblem.
through inquiries conducted during the project. Note also
that excess inventory costs are proportional to the number
of boxes above the container capacity but do not depend on
3. PROBLEM DESCRIPTION
the number of periods these boxes are kept in the pharmacy
We coin our problem as the Stochastic collector managed before being collected. This way, we privileged disturbance
IRP (SCMIRP). Two types of decisions have to be taken: (1) happening at the moment when the box is deposited against
when to visit each pharmacy, (2) how to sequence vehicle risk and disturbance that occur all along the time period when
routes. the box is stored out of the container. Every model and method
The planning period is fixed to a given number of days presented subsequently could, however, easily be adapted to
(typically a few weeks). Based on the situation encountered this alternative modeling if needed.

NETWORKS—2014—DOI 10.1002/net 85
 We assume that, at a given period, when a visit is planned, TABLE 1. Notation.
all boxes arrive before the vehicle visits the pharmacies. This
assumption relies on the hypothesis that the vehicle visits Parameters
n Number of pharmacies
pharmacies at relatively regular schedules. Hence, random T Number of time periods
variables can be defined according to that schedule. ξit Number of boxes disposed at pharmacy i at time t
The problem can be formulated as a two-stage stochastic (random variable with Poisson distribution of mean λit )
optimization problem with recourse. In the first stage, deci- Ii0 Initial inventory at pharmacy i
sions have to be made on the visit dates for each pharmacy Iimax Maximal accepted inventory at pharmacy i
Iithrd Threshold inventory at pharmacy i
and the tours by which the vehicle visits the pharmacies. The cij Travel cost between locations i and j
actual number of boxes at a pharmacy becomes known when cf Fixed vehicle cost
the pharmacy is visited. Therefore, the number of boxes to be cithrd Penalty cost at pharmacy i
collected at each pharmacy is determined in the second stage. cimax Unitary excess inventory cost at pharmacy i
Note that for our model, it makes no difference whether the Variables
wit 1 if pharmacy i is visited at time t,
number of boxes at the pharmacies becomes known at once 0 otherwise
(e.g., assessed via RFID technology), or gradually, going zit Number of boxes picked up at pharmacy i at time t
from one pharmacy to the next. This particularity is explained xijt 1 if pharmacy j is visited immediately after pharmacy i at time t,
by the fact that in the second stage the whole inventory at each 0 otherwise
pharmacy is collected. Iit Inventory at pharmacy i at time t
vit Number of boxes exceeding Iimax at time t
Let G = (V , A) be a complete graph and {1, . . . , T } a set of t
yi 1 if Iithrd is not reached when visiting pharmacy i at time t,
discrete time periods. Set V = {0, 1, . . . , n} is the vertex set 0 otherwise
and A is the arc set. Vertex 0 stands for the vehicle depot. Ver-
tices i = 1, . . . , n represent the pharmacies. The travel cost on
arc (i, j) ∈ A is denoted cij ; the travel cost matrix is assumed
to satisfy the triangle inequality. The fixed cost incurred when 
n

carrying out a tour is denoted cf . Random variables ξit rep- xijt = wjt ∀j = 1, . . . , n; t = 1, . . . , T (3)
i=0, i=j
resent the number of boxes disposed at pharmacy i at time
period t. The initial inventory level at pharmacy i is Ii0 . Cost 
n

cithrd denotes the penalty cost incurred by a visit to pharmacy xijt = wit ∀i = 1, . . . , n; t = 1, . . . , T (4)
j=0, j=i
i, if the threshold inventory level Iithrd is not reached. Cost

cimax represents the unitary excess inventory cost incurred by xijt ≤ |S| − 1 ∀S ⊆ {1, . . . , n} ; t = 1, . . . , T (5)
a visit at pharmacy i, if the maximal accepted inventory level i,j∈S
Iimax is exceeded.
The following decision variables are introduced. Binary xijt ∈ {0, 1} ∀i = 0, . . . , n; j = 0, . . . , n; t = 1, . . . , T (6)
variable wit is equal to 1 if pharmacy i is visited at time period wit ∈ {0, 1} ∀i = 1, . . . , n; t = 1, . . . , T (7)
t, 0 otherwise; zit is the number of boxes picked up during the
visit. Flow variable xijt is a binary variable set to 1 if pharmacy
j is visited immediately after pharmacy i at time period t, 0 Second stage:
otherwise. Variables Iit are introduced to compute the inven-
tory level at pharmacy i after arrival of new boxes and before 
n 
T 
n 
T
collection of stored boxes at time period t. The nonnegative Q(x, w, ξ ) = min cimax vit + cithrd yit
number of boxes exceeding the maximal accepted inventory i=1 t=1 i=1 t=1
is denoted vit ; binary variable yit indicates if Iithrd is reached 
n
or not. + cimax IiT +1 s.t. (8)
Variables xijt and wit determine the first-stage decisions, i=1
whereas variables zit determine the second-stage decision. 
t 
t−1
Table 1 summarizes the general notation for the SCMIRP. Iit = Ii0 + ξis − zis ∀i = 1, . . . , n; t = 1, . . . , T
Using this notation, the SCMIRP can be formulated as s=1 s=1
follows. (9)
First stage: IiT +1 ≥ IiT − ziT − Iimax ∀i = 1, . . . , n (10)

n 
n 
T 
n 
T
zit ≤ + M(1 − wit ) ∀i = 1, . . . , n; t = 1, . . . , T
Iit (11)
min cij xijt + t
cf x0j + E(Q(x, w, ξ )) s.t.
i=0 j=0 t=1 j=1 t=1 zit ≥ − M(1 − wit ) ∀i = 1, . . . , n; t = 1, . . . , T
Iit (12)
(1) zit ≤ Mwit ∀i = 1, . . . , n; t = 1, . . . , T (13)

n
vit ≥ zit − Iimax ∀i = 1, . . . , n; t = 1, . . . , T (14)
t
x0j ≤ 1 ∀t = 1, . . . , T (2)
j=1 zit ≥ Iithrd (wit − yit ) ∀i = 1, . . . , n; t = 1, . . . , T (15)

86 NETWORKS—2014—DOI 10.1002/net
 Iit ≥ 0 ∀i = 1, . . . , n; t = 1, . . . , T + 1 (16) random scenarios (Birge and Louveaux [5]). In this approach,
the given distribution of the random vector ξ is approximated
zit ≥ 0 ∀i = 1, . . . , n; t = 1, . . . , T (17)
by an empirical sample distribution, obtained by drawing N
vit ≥0 ∀i = 1, . . . , n; t = 1, . . . , T (18) scenarios ξ ν (ν = 1, …, N)) from the original distribution
yit ∈ {0, 1} ∀i = 1, . . . , n; t = 1, . . . , T (19) of ξ . The original distribution is then replaced by a discrete
distribution, each of the N scenarios possessing the same
The objective function of the first-stage problem is given probability 1/N. The expected unitary excess inventory costs
by (1). It expresses the minimization of the total cost, which and the expected costs of a visit at a pharmacy are then
comprises travel costs, fixed costs, and expected inventory approximated by the sample average:
costs. The expectation is taken with respect to the distribution
1 
N
of the vector ξ of random variables ξit . The actual inven-
tory costs Q(x, w, ξ ) result as the solution of the second-stage Q(x, w, ξ (ν) ) (20)
N
problem (8)–(19). In the terminology of stochastic program- ν=1
ming, Q(x, w, ξ ) is called the recourse function. It expresses
the cost of the decisions taken after the realization of ξ . By defining Iit(ν) , zit(ν) , vit(ν) , and yit(ν) as the instance of the
Constraints (2) state that at most one tour can be carried second-stage variables corresponding to scenario ν, we can
out per discrete time period. Constraints (3) and (4) guarantee combine the first-stage model (1)–(7) and the second-stage
that a pharmacy is only visited when an arc leads to that phar- model (8)–(19) into a single mixed integer linear program as
macy; they also ensure that if the vehicle visits a pharmacy follows:
at a discrete time period, it also leaves the pharmacy at the
same time period. Constraints (5) are the subtour elimination 
n 
n 
T 
n 
T
min cij xijt + t
cf x0j
constraints. Constraints (6) and (7) define the variables of the
i=0 j=0 t=1 j=1 t=1
first-stage problem.

1    max (ν)   thrd (ν)
In the second-stage problem, Equation (8) expresses the N n T n T
objective function as the sum of unitary excess inventory + ci vit + ci yit
N
costs and penalty costs (variables IiT +1 represent inventory ν=1 i=1 t=1 i=1 t=1

in excess at the end of the horizon). Constraints (9) and (10) n
max (ν)
define the level of inventory of pharmacy i at time period t + ci Ii,T +1 (21)
and the inventory in excess at the end of the planning hori- i=1
zon, respectively. Constraints (11) and (12) ensure that the
whole inventory is collected when pharmacy i is visited. Con- s.t.
straints (13) force the number of collected boxes to zero

n
if a pharmacy is not visited. Constraints (14), with Con- t
x0j ≤1 ∀t = 1, . . . , T (22)
straints (18), determine the nonnegative number of boxes that j=1
exceed the maximal accepted inventory at pharmacy i. Con-
straints (15) define yit as equal to 1 if the number of boxes 
n
xijt = wjt ∀j = 1, . . . , n; t = 1, . . . , T (23)
picked up at pharmacy i at time t is lower than the threshold
i=0, i=j
inventory, and 0 otherwise. Constraints (16) to (19) define the
variables of the second-stage problem. 
n
xijt = wit ∀i = 1, . . . , n; t = 1, . . . , T (24)
j=0, j=i
4. SOLUTION APPROACHES 
xijt ≤ |S| − 1 ∀S ⊆ {1, . . . , n} ; t = 1, . . . , T (25)
Due to the complexity of the problem, an exact solu- i,j∈S
tion method cannot be envisaged to solve realistic problem
instances. Therefore, we propose two heuristic methods to 
t 
t−1

solve the SCMIRP for infectious medical waste. The first Iit(ν) = Ii0
(ν)
+ ξis(ν) − zis(ν)
solution approach is a sampling method. This method is based s=1 s=1

on solving a mixed integer program in which we consider that ∀i = 1, . . . , n; t = 1, . . . , T ; ν = 1, . . . , N (26)

the probability distribution of ξ has a finite number of real- (ν) (ν) (ν)
+1 ≥ IiT − ziT − Ii ∀i = 1, . . . , n; ν = 1, . . . , N
max
izations called scenarios. The second solution approach is Ii,T
based on an ALNS algorithm and a collection pattern based (27)
operator is developed to handle stochasticity.
zit(ν) ≤ Iit(ν) + M(1 − wit )
∀i = 1, . . . , n; t = 1, . . . , T ; ν = 1, . . . , N (28)
4.1. Sampling Method
zit(ν) ≥ Iit(ν) − M(1 − wit )
The aspect of stochasticity in our problem is treated here
by performing optimization based on a fixed sample of ∀i = 1, . . . , n; t = 1, . . . , T ; ν = 1, . . . , N (29)

NETWORKS—2014—DOI 10.1002/net 87
 zit(ν) ≤ Mwit ∀i = 1, . . . , n; t = 1, . . . , T ; ν = 1, . . . , N Algorithm 1 ALNS algorithm in a multi-start procedure
(30) 1: f (s∗ ) ← +∞
2: repeat
vit(ν) ≥ zit(ν) − Iimax 3: Initialize probabilities for perturbation
∀i = 1, . . . , n; t = 1, . . . , T ; ν = 1, . . . , N (31) 4: s ← initial solution
5: s ← local search(s)
zit(ν) ≥ Iithrd (wit − yit ) 6: repeat
7: STOP = false
∀i = 1, . . . , n; t = 1, . . . , T ; ν = 1, . . . , N (32) 8: repeat
xijt ∈ {0, 1} ∀i = 0, . . . , n; j = 0, . . . , n; t = 1, . . . , T 9: Select destroy mechanism
10: s ← perturbation(s) (destroy mechanism)
(33)
11: s ← local search(s ) (repair mechanism)
wit ∈ {0, 1} ∀i = 1, . . . , n; t = 1, . . . , T (34) 12: until Improvement or fail
13: if f (s ) < f (s) then
Iit(ν) ≥0 ∀i = 1, . . . , n; t = 1, . . . , T + 1; ν = 1, . . . , N 14: s←s
(35) 15: Update probabilities for perturbation
16: else
zit(ν) ≥ 0 ∀i = 1, . . . , n; t = 1, . . . , T ; ν = 1, . . . , N (36) 17: STOP = true
18: end if
vit(ν) ≥0 ∀i = 1, . . . , n; t = 1, . . . , T ; ν = 1, . . . , N (37)
19: until STOP
yit(ν) ∈ {0, 1} ∀i = 1, . . . , n; t = 1, . . . , T ; ν = 1, . . . , N 20: if f (s) < f (s∗ ) then
21: s∗ ← s
(38)
22: end if
23: until Time limit

4.2. ALNS Algorithm

ALNS (Pisinger and Ropke [26]) is a large neighborhood The ALNS algorithm relies on four different destroy
search based algorithm, where an initial solution is gradually operators and a single-large neighborhood local search mech-
improved by alternately destroying and repairing the solution. anism. These operators and the way they are combined is
detailed below.
4.2.1. Basic Scheme As an extension of large neighbor- 4.2.2. Destroy and Repair Mechanisms We explore four
hood search, ALNS allows multiple destroy and repair mech- different diversification mechanisms in the ALNS, which are
anisms (multiple neighborhoods) to be used within the same selected by a roulette wheel method based on their success
search. Each of the destroy and repair mechanisms is assigned during the search. At the beginning, each destroy mechanism
a weight representing its performance during the search, has the same selection probability, which is increased by 1.5%
which controls the selection of a particular mechanism. The if the particular operator leads to an improvement of the cur-
weights are adjusted dynamically during the search; there- rent solution (probabilities of other operators are decreased
fore, mechanisms that perform well have a higher chance of by 0.5%). Each perturbation is followed by local search. The
being selected. perturbated solution s , improved with local search, is only
The following pseudocode presents the basic scheme of kept when beating current solution s. Otherwise, another
our ALNS algorithm, which is embedded in a multistart destroy operator is tried. When all destroy operators fail,
procedure. ALNS is stopped.
Solution s∗ represents the best solution found during the The four operators are listed below:
execution of the algorithm. Initially, s∗ refers to a fictitious
solution whose value f (s∗ ) is artificially set to +∞. Each iter- • The first operator removes all visits on a random day d. For
ation of loop 2–23 corresponds to a new multistart iteration: each pharmacy that was visited on day d, either day d-1
a random solution s is constructed (line 4), then improved or d+1 is randomly selected, and the pharmacy is added
with ALNS (lines 5–19); once improved, s is compared to s∗ with a best insertion policy (giving the possibility that a
and kept if better (lines 20 to 22). Multi-start is repeated until new route is created).
a given time limit is exceeded. Details on the way solutions • The second destroy operator determines the day d, at which
are evaluated, especially with regard to inventory costs, can least often a tour has been planned. All pharmacies are then
inserted on day d.
be found in section 4.2.2.
• The third operator temporarily multiplies inventory costs
Initial solutions are generated by randomly assigning col- by a factor 3 and applies a local search iteration (see below).
lection days to pharmacies. Then, for each day, a traveling • The fourth destroy operator selects the pharmacy with
salesman problem (TSP) is solved with the Lin–Kernighan highest inventory costs and reassigns its visit dates such
heuristic (Lin and Kernighan [20]). The Lin–Kernighan that expected inventory costs are minimized; this is
heuristic is a generalization of 2-opt and 3-opt, swapping achieved using the local search operator (see below) and
pairs of subtours to create a new tour. fixing transportation costs to zero.

88 NETWORKS—2014—DOI 10.1002/net
 FIG. 3. Shortest path problem for pharmacy i.

The intensification mechanism of our ALNS is inspired reached or because the maximal accepted inventory level is
by periodic TSPs and is based on collection patterns. A col- exceeded. Given a pharmacy i and two successive visit dates
lection pattern is defined as a set of visit dates for a pharmacy. t and t , the estimated inventory costs can be calculated as
Neighbor solutions are obtained from the current solution by follows:
selecting a pharmacy, exchanging the collection pattern for ∞

this pharmacy, and reoptimizing vehicle routes with the Lin– ditt = cimax P(Iitt = I) × (I − Iimax )
Kernighan heuristic, thus defining 2T neighbor solutions per I=Iimax +1
pharmacy. Local search then consists in selecting the best
neighbor solution for pharmacies in turn, until a complete + cithrd P(Iitt ≤ Iithrd − 1) (39)
round of the pharmacies shows that current collection patterns

t
are the best ones for all pharmacies. Iitt = ξir (40)
Figure 3 illustrates the calculation of the best neighbor for r=t+1
a pharmacy i, which is achieved by solving a shortest path
problem in the following graph. Each node of the graph rep- Equation (39) represents inventory costs, which are com-
resents a discrete time period. Additionally, a fictitious start posed of unitary excess inventory costs and penalty costs
node and a fictitious end node complete the graph. Visiting at pharmacy i. To take account of the stochastic data, we cal-
a node means that pharmacy i is visited at the corresponding culate the probability of reaching a certain inventory level
time period (day) t. An arc leads from each node to all of its when pharmacy i is visited. Equation (40) determines the
successors. Paths in the graph then exactly match collection inventory at pharmacy i at time t , if a collection is done at
patterns. time t.
The distance (costs) for the shortest path problem is com- Regarding inventory costs dist for outgoing arcs of ficti-
posed of two series of costs. With the traversal of each node tious node s, Equation (40) is replaced with:
is associated a cost t , which represents the routing cost
induced by the visit of pharmacy i on that day. With each 
t
Iist = Ii0 + ξir (41)
arc is associated an inventory cost ditt , which computes the
r=1
expected inventory cost occasioned when planning the next
visit of the pharmacy after day t on day t . computing the accumulated inventory, from the initial inven-
To evaluate quickly costs t , t is defined as the phar- tory Ii0 , at period t . For the costs dite , leading to the fictitious
macy insertion cost (when pharmacy i is not present in the end node, only unitary excess inventory costs are considered.
vehicle tour on day t) or removal saving (when pharmacy i Equation (39) is changed to:
is present). Note that once the best neighbor is found, actual ∞
routing costs for the collection pattern are computed using 
dite = cimax P(IitT = I) × (I − Iimax ) (42)
the Lin–Kernighan heuristic, as stated above.
I=Iimax +1
Stochasticity of the SCMIRP is captured in arc costs
ditt . In the SCMIRP, the number of boxes disposed at a The shortest path is found by applying Dijkstra’s algorithm
pharmacy cannot be known exactly a priori. Therefore, the [11] and provides the best collection pattern for pharmacy i,
number of boxes disposed at pharmacy i by unit of time is in view of collection patterns for other pharmacies. If the col-
expressed by the random variable ξit . We calculate the esti- lection pattern represented by the shortest path differs from
mated inventory costs ditt as the estimated costs of being the current pattern, the solution is updated. Note that at this
penalized either because the threshold inventory level is not point, as explained above, Lin–Kernighan heuristic is applied

NETWORKS—2014—DOI 10.1002/net 89
to improve routing costs at the days where vehicle routes have 5.2. Evaluation of the Methods
changed. Local search then continues for the next pharmacy.
In a first part of our experiments, we evaluate the pro-
5. COMPUTATIONAL EXPERIMENTS posed solution approaches. We proceed with two series of
experiments.
All experiments are conducted on a PC equipped with an First, to assess the performance of our ALNS algo-
Intel Core2 processor paced at 2.5 GHz and with 3.5 Gb of rithm, solutions determined by ALNS based on deterministic
RAM. Mixed integer linear programs are solved using IBM box arrival data (ALNS deterministic) are compared to the
ILOG CPLEX 11.2 and Concert Technology. optimal solutions determined with CPLEX (CPLEX real-
izations). The ALNS algorithm can simply be adapted to a
5.1. Test Instances deterministic setting: costs ditt assigned to arc traversals in
We tested the proposed solution approaches on several the collection pattern based local search operator are replaced
realistic instances from the region of PACA. In this region, with real inventory costs. The mixed integer linear program
pharmacies equipped with RFID technology participate in (21)–(38) presented in section 4.1 provides the optimal solu-
the project “PAC ID DASRI”, with the aim of improving tion for a given scenario by considering only this scenario
the logistics of infectious medical waste. Information on the in the sample. It can be solved with a cutting-plane method
number of boxes arriving at each pharmacy and the spec- as follows: formulation (21)–(38) is solved without subtour
ification of the threshold inventory level and the maximal elimination constraints (25), which are added when needed
accepted inventory level (is) gathered from the participating until a feasible solution is found.
pharmacies. This series of experiments are carried out on the set of
To evaluate the proposed solution approaches, numerical small instances as the exact solution is out of reach for
experiments are carried out for five small instances and five real-world instances. CPLEX and ALNS are applied to 10
real-world instances. Concerning the small instances, five (randomly generated) realizations of box arrivals for each
pharmacies are considered on a time period of 10 days, each instance. Table 2 presents the average costs and computa-
discrete time instant representing 1 day. The fixed vehicle tion times for the 10 realizations for each of the five small
costs are set to 1000, penalty costs (if boxes are collected instances. Time limit for ALNS is set to 30 s.
when the threshold inventory level has not been reached) The results show that the optimal solution is found by the
are 1000, and excess inventory costs are 1000 for each box ALNS algorithm for all 50 test instances, indicating that our
exceeding the maximal accepted inventory level. The real- algorithm performs very well on small instances.
world instances contain 30 pharmacies each, with a time Then, we compare the quality of the two solution methods
horizon of 20 days. For these instances, the fixed vehicle introduced in section 4 with stochastic data. Therefore, we
costs are 5000, penalty costs are 1000 and excess inventory oppose the ALNS algorithm (ALNS average) to the sampling
costs are 1000 for each box exceeding the maximal accepted method (CPLEX sampling). We complete these comparisons
inventory level. with a third heuristic that computes the optimal solution when
We assume that full boxes of infectious medical waste box arrival data are set to their average value (CPLEX aver-
arrive independently at pharmacies according to a Poisson age). The latter is solved using the sampling formulation with
distribution with mean λti . The estimated inventory costs for a single-scenario set to average box arrival values. Note that a
pharmacy i are then calculated as follows from Equations (39) direct comparison between these approaches is meaningless.
and (40). Indeed, only ALNS average provides the value of objective
function (1). For the sake of the comparison, we evaluate the

t
solution costs a posteriori for the two other methods using
λ= λri (43)
Equation (44), which allows computing average inventory
r=t+1
costs once routes are known. Table 3 reports these results and
∞
 tt
−λ λ Ii the computation times. Again these experiments could only
ditt = cimax e (I tt − Iimax ) be performed on small instances. For the sampling method,
Iitt ! i
Iitt =Iimax +1 10 randomly generated realizations are used. Mixed integer
Iithrd −1 linear programs are solved with a cutting-plane method as
 tt
λI i
−λ described above. The stopping criterion for ALNS is set to
+ cithrd e (44)
Iitt ! 30 s.
Iitt =0
Table 3 shows that our ALNS algorithm can handle
Iimax −1
 λI i
tt
stochasticity and determine solutions with a quality signif-
= cimax (λ − Iimax + (Iimax − λ) e−λ icantly better than other methods. CPLEX average is not
Iitt !
Iitt =0 able to provide solutions that are robust regarding other
 realizations of the stochastic box arrivals, with gaps up to
Iithrd −1
λ Ii
max
 tt
λI i 43.5%. CPLEX sampling generates solutions that are more
+ Iimax e−λ + cithrd e −λ
Iimax ! Iitt ! robust in terms of unknown realizations of box arrivals but is
Iitt =0 still incapable of providing robust solutions with respect to

90 NETWORKS—2014—DOI 10.1002/net
 TABLE 2. Computational results for realizations of small instances (deterministic).

Routing costs Inventory costs Total costs Computing time

CPLEX realizations
Instance 1 2863.2 2100.0 4963.2 18 min
Instance 2 2983.8 4200.0 7183.8 45 min
Instance 3 2346.9 700.0 3046.9 10 min
Instance 4 2147.1 1100.0 3247.1 7 min
instance 5 2966.9 2700.0 5666.9 27 min
ALNS deterministic
Instance 1 2863.2 2100.0 4963.2 30 sec
Instance 2 2983.8 4200.0 7183.8 30 sec
Instance 3 2346.9 700.0 3046.9 30 sec
Instance 4 2147.1 1100.0 3247.1 30 sec
Instance 5 2966.9 2700.0 5666.9 30 sec

every possible realization of box arrivals. Computing times can be explained by the fact that the stochasticity impacts
for ALNS are also several order of magnitudes faster than only inventory costs which play a small role when routing
for other methods. It should, however, be noted that the costs are very high.
cutting-plane method used for solving the mixed integer
linear programs is quite naive and could probably be signif-
icantly accelerated using addditional valid inequalities from 5.3. Managerial Implications of RFID
the TSP polytope.
We now evaluate the performance of different policies
In what follows, we conduct a sensitivity analysis on rout-
for the collection of infectious medical waste. For these
ing and inventory costs to show the behavior of the three
experiments, the set of five real-world instances is used. We
methods. Initial inventory and routing costs are multiplied
compare the new policy investigated in this article with three
by different factors to give more or less importance to each
other policies. Remind that the new policy, subsequently
cost component. Table 4 summarizes average costs, average
called the RFID scenario, refers to the planning of collec-
computation times, and average gaps for each cost factors
tion tours on the whole time horizon at once, based on stock
configuration and for each method. Inv. Factor and Rout. Fac-
levels which are known in real time, and statistics on full-box
tor represent factors for, respectively, the inventory costs and
arrival rates. Other policies are:
the routing costs. The stopping criterion for ALNS is set to
30 s.
From Table 4, we can easily notice that computation times Current scenario 85%: Pharmacies make a phone call to
increase considerably for CPLEX average and CPLEX sam- the local authority when reaching 85% of the maximal
pling when routing costs are high. We can also notice that gaps accepted inventory. Boxes at a pharmacy are collected the
to ALNS become smaller when routing costs are high. This day after having called the local authority. This policy

TABLE 3. Computational results for small instances (stochastic).

Routing costs Inventory costs Total costs Computing time Gap to ALNS (%)

CPLEX average
Instance 1 2497.5 6365.1 8862.6 5 min 17.3
Instance 2 3687.7 10, 061.7 13, 749.4 30 min 25.8
Instance 3 1269.6 4200.8 5470.4 15 min 0.0
Instance 4 1188.5 9600.4 10, 788.9 2 min 43.5
Instance 5 2455.5 7644.9 10, 100.4 35 min 3.6
CPLEX sampling
Instance 1 3637.2 5344.5 8981.7 32 h 18.8
Instance 2 2551.9 8371.0 10, 922.9 46 h 0.0
Instance 3 1269.6 4200.8 5470.4 11 h 0.0
Instance 4 2426.7 5945.6 8372.3 2h 11.4
Instance 5 2455.5 7644.9 10, 100.4 72 h 3.6
ALNS average
Instance 1 2497.7 5059.7 7557.4 30 sec
Instance 2 2551.9 8371.0 10, 922.9 30 sec
Instance 3 1269.6 4200.8 5470.4 30 sec
Instance 4 2427.8 5088.7 7516.5 30 sec
Instance 5 2455.5 7290.9 9746.4 30 sec

NETWORKS—2014—DOI 10.1002/net 91
 TABLE 4. Cost analysis for small instances (stochastic).

Method Rout. Inv. Total Routing Inventory Computing Gap to

Factor Factor costs costs costs time (sec.) ALNS

ALNS average 1 10 4236.01 52, 401.68 56, 637.66 30 0.00%

CPLEX average 1 10 3351.52 75, 385.50 78, 737.08 8 44.44%
CPLEX sampling 1 10 4002.17 58, 629.92 62, 632.10 1749 11.77%
ALNS average 1 5 3804.21 26, 467.18 30, 271.36 30 0.00%
CPLEX average 1 5 3351.52 37, 208.96 40, 560.48 9 37.68%
CPLEX sampling 1 5 3787.52 29, 722.70 33, 510.18 2489 10.83%
ALNS average 5 1 10, 014.27 6408.94 16, 423.20 30 0.00%
CPLEX average 5 1 9962.86 7352.36 17, 315.22 11483 5.71%
CPLEX sampling 5 1 10, 049.91 6948.21 16, 998.10 15, 7452 3.99%
ALNS average 10 1 15, 143.84 9400.23 24, 544.08 30 0.00%
CPLEX average 10 1 15, 017.28 12, 844.78 27, 862.04 24, 384 12.98%
CPLEX sampling 10 1 12, 589.62 12, 172.13 24, 761.76 132, 457 0.98%

approximately corresponds to the existing one in the determines the collection route for the following day, which
PACA region. allows, once box arrivals for the next day are known, to deter-
mine new inventory levels. Time limit for ALNS is set to 60
Current scenario threshold: Pharmacies call the local
s for the first day and to 10 s for the remaining days (where
authority when the threshold inventory level is reached.
the algorithm is warm-started with the current solution).
Collection of boxes is performed the same day. This pol-
It can easily be observed that the total costs for the current
icy is optimized as against the previous one, by assuming
scenario 85% are much higher than the costs for the RFID
a high reactivity of local authorities and a strong intu-
scenario. The advantage of the RFID scenario is that the col-
ition (without RFID) or reactivity (with RFID) of the
lector can flexibly combine pharmacies in collection tours to
pharmacists.
decrease total costs. This is performed by a trade-off between
Rolling horizon scenario: Solutions found by the RFID routing costs and inventory costs. For example, it might be
scenario are revised every day, according to the actual preferable to visit a pharmacy which is close-by, even if the
arrivals of data. threshold inventory level has not been reached. Remember
that penalty costs incur when a pharmacy is visited too early
All these policies cannot easily be compared on average. (with respect to the threshold inventory) and unitary excess
Indeed, decisions taken on a given day depend on the actual inventory costs incur when a pharmacy is visited too late
arrival of boxes on previous days. We thus propose to com- (with respect to the maximal accepted inventory). In the cur-
pare them using a set of 100 realizations of data. To evaluate rent scenario 85%, the local authority is not able to prevent
whether it provides a good approximation of the average visits above the excess inventory level, because a visit is only
behavior of these policies, we first apply this methodology to carried out when a pharmacist makes a call. For this reason,
the RFID scenario. inventory costs are rather high.
In Table 5, we compare two measures for ALNS solutions In the current scenario threshold, a tour is carried out
on the five instances. ALNS average refers to the expected whenever the local authority receives a call, even if only one
costs based on average box arrival rates and corresponds to pharmacy has to be visited. As the routing of the vehicle can
the solution value returned by the algorithm. ALNS approx- only be optimized on a daily basis, routing costs are very
imated represents the values determined by calculating the elevated. Due to frequent collections just after the threshold
actual inventory costs for the 100 realizations based on the inventory is reached, inventory costs can, however, be kept
solution (collection routes) provided by ALNS. The pre- very low, which leads to total costs comparable to those of
sented value for each instance includes the mean of the the RFID scenario. However, the current scenario threshold
inventory costs for the 100 realizations plus routing costs, requires a strong implication of pharmacists, assumes a high
which remain unchanged.
Table 5 shows very limited gaps between the two mea-
sures. It thus tends to indicate that evaluating the cost of a
solution by computing its average cost on 100 realizations TABLE 5. Computational results for real-world instances (ALNS).
gives a good approximation of the expected costs. ALNS approximated ALNS average gap in %
Table 6 then compares the four different policies. For all
policies, costs are the average values of costs for 100 identi- Instance 1 141, 148.4 139, 569.0 1.1
cal realizations. In the RFID scenario, the time limit is set to Instance 2 155, 121.7 154, 095.0 0.7
300 s. The solution framework for the Rolling horizon sce- Instance 3 150, 839.2 149, 121.0 1.1
Instance 4 131, 879.4 131, 130.0 0.3
nario is based on successive calls to the ALNS algorithm,
Instance 5 150, 764.0 149, 130.0 1.1
with updated initial inventory levels: every call to ALNS

92 NETWORKS—2014—DOI 10.1002/net
 TABLE 6. Computational results for real-world instances.

Routing costs Inventory costs Total costs

Current scenario 85%

Instance 1 98, 060.0 156, 870.0 254, 930.0
Instance 2 97, 535.4 177, 960.0 275, 495.4
Instance 3 98, 779.4 185, 810.0 284, 589.4
Instance 4 96, 575.0 140, 610.0 237, 185.0
Instance 5 98, 475.8 179, 700.0 278, 175.8
Mean value 97, 885.1 168, 190.0 266, 075.1
Current scenario threshold
Instance 1 107, 120.4 31, 830.0 138, 950.4
Instance 2 107, 355.0 39, 390.0 146, 745.0
Instance 3 107, 721.4 37, 510.0 145, 231.4
Instance 4 106, 346.7 25, 810.0 132, 156.7
Instance 5 107, 296.3 38, 270.0 145, 566.3
Mean value 107, 167.9 34, 562.0 141, 729.9
RFID scenario
Instance 1 47, 868.8 91, 700.0 141, 148.4
Instance 2 53, 114.8 100, 980.0 155, 121.7
Instance 3 42, 700.7 106, 420.0 150, 839.2
Instance 4 42, 630.6 88, 860.0 131, 879.4
Instance 5 47, 939.6 101, 190.0 150, 764.0
Mean value 46, 850.9 97, 830.1 144, 681.0
Rolling horizon scenario
Instance 1 53, 383.4 71, 650.0 125, 033.4
Instance 2 53, 069.2 83, 540.0 136, 609.2
Instance 3 54, 102.3 81, 920.0 136, 022.3
Instance 4 50, 074.3 68, 310.0 118, 384.3
Instance 5 52, 191.8 82, 080.0 134, 271.8
Mean value 52, 564.2 77, 500.0 130, 064.2

reactivity of the collector and prevents from any long-term along the rolling horizon and on the other hand only taking
planning of his/her activity. into consideration whether or not a tour has been planned or
Basically, the current scenario threshold should rather be canceled for the following day. The number of times the col-
compared with the rolling horizon scenario, where informa- lector or the pharmacists are disturbed because of a change of
tion on actual arrival of boxes is also exploited to reoptimize planning (which is quite subjective) certainly stands some-
collection routes. Results then show that a better balance where between these two extreme measures. Table 7 provides
between routing and inventory costs can be obtained, thus the values of these two indicators for the rolling horizon
allowing a large improvement of the solution costs. scenario.
The conclusion that can be drawn from these results is that Table 7 exhibits consistent values. On average, 3.1 times
the RFID technology can significantly help reducing routing within the 20 days a tour is scheduled or canceled for the
and inventory costs, and that the two aforementioned policies following day. Thus, the driver for the local authority has
deserve to be considered: the RFID scenario, based on a static to reorganize the planned work load for the next day, caus-
planning, and the rolling horizon scenario where updates of ing major inconvenience for him/herself as well as for the
the planning are accepted every day.
So as to better compare these two policies, we assessed
two additional indicators. These two indicators concern the TABLE 7. Average number of changes for the Rolling horizon scenario.
rolling horizon scenario. The first one states the average num-
ber of modifications of the planning over the whole time Rolling horizon scenario
horizon: one change is counted every time the solution pro-
vided by the daily optimization modifies the whole planning. Modifications of the Strong modifications for
whole planning the following day
Note that not all of these modifications will be realized in
practice as they might concern distant days. The second indi- Instance 1 16.7 3.5
cator represents the number of times the daily optimization Instance 2 16.6 2.8
has a strong impact on the following day, where strong means Instance 3 16.7 3.1
that the fact that a route is scheduled or not changes. There- Instance 4 16.5 3.1
Instance 5 16.5 2.8
fore, the two indicators reflect two extreme cases, on the one
Mean value 16.6 3.1
hand accounting for each pharmacy that has been rescheduled

NETWORKS—2014—DOI 10.1002/net 93
concerned pharmacists. Over the whole time horizon, the process can be ameliorated for both, local authority and
initial collection schedule is changed on average 16.6 times. pharmacists.
The present solution approaches are to be embedded into a
decision support system. Together with the decision makers
5.4. Conclusions (pharmacists, local authority), the weight for the inventory
costs and the threshold and maximal accepted inventory
The presented computational results enable the following
levels can be determined. These values can be chosen inde-
conclusions and implications.
pendently for each pharmacy, to reflect the inconvenience
First, the comparison between the optimal solutions for
caused by a visit or excess of inventory.
five small instances and the solutions determined by our
In the basic model, we do not consider vehicle capacity
ALNS algorithm illustrates that our solution approach per-
or tour length constraints, as they were not relevant for our
forms very well and can be applied to generate good
case-study. For the sake of more generality, these constraints
approximations of the optima for real-world size problems.
could be included in the future. In addition, a further devel-
Second, further numerical results on larger instances show
oped version might also consider the inventory level at the
that our ALNS algorithm is able to handle stochasticity by
collection centers to optimize the distribution of the boxes
computing a reliable expectation of inventory costs.
among the collections centers.
Third, comparisons with mathematical programming
based approaches show that a simple implementation of these
methods is not competitive. Quickened solution of the sam- Acknowledgment
pling method with larger samples is required to hope beating The authors gratefully acknowledge the support of the
ALNS. French Environment and Energy Management Agency
Fourth, computational results for five real-world instances (ADEME).
provide an insight into the different policies which could be
pursued by the local authority and the pharmacists. Focusing
on the planning for a whole time horizon, the RFID scenario REFERENCES
is recommended. It is based on box arrival rate statistics and [1] J. Alegre, M. Laguna, and J. Pacheco, Optimizing the periodic
initial inventory data. A schedule of visit dates and routes pick-up of raw material for a manufacturer of auto parts, Eur
for the whole time horizon is then computed, which allows J Oper Res 179 (2007), 736–746.
the local authority as well as pharmacists to plan their daily [2] A. Ak and A.L. Erera, A paired-vehicle recourse strategy for
activities and optimize their work schedule. Furthermore, as the vehicle-routing problem with stochastic demands, Transp
the routing is optimized by visiting adjacent pharmacies on Sci 41 (2007), 222–237.
the same tour, distances and emissions can be lowered. As the [3] S. Baptista, R.C. Oliveira, and E. Zuquete, A period vehicle
planning is performed in advance and based on data gathered routing case study, Eur J Oper Res 139 (2002), 220–229.
via RFID technology, the process can be termed automated [4] L. Bertazzi, M. Savelsbergh, and M.G. Speranza, “Inven-
and does not require administrative effort of pharmacists or tory routing,” The Vehicle Routing Problem: Latest Advances
and New Challenges, B. Golden, S. Raghvan, and E. Wasil
the local authority. This results in less disturbance for all
(Editors), Springer, New York, 2008, 49–72.
involved parties.
[5] J.R. Birge and F. Louveaux, Introduction to stochastic pro-
The embedding of the RFID scenario in a rolling horizon
gramming, Springer, New York, 1997.
framework allows further improvements of all cost compo-
[6] A. Campbell, L. Clarke, A. Kleywegt, and M. Savelsbergh,
nents, as the data registered via RFID is used every day to The inventory routing problem, Fleet Manag Log, Kluwer
optimize the time schedule for the remaining time horizon. Academic Publishers, Massachusetts (1998), 95–113.
However, this implies that changes in planning are likely to [7] A. Campbell, L. Clarke, and M. Savelsbergh, “Inventory
occur and the involved parties cannot rely on the provided routing in practice,” The Vehicle Routing Problem, P. Toth
schedules. Therefore, flexibility and the potential to adapt to and D. Vigo (Editors), Society for industrial and applied
new situations are required, rendering the process of work mathematics, Philadelphia, 2002, 309–330.
scheduling less convenient for the pharmacists and the local [8] A.M. Campbell and J.R. Hardin, Vehicle minimization for
authority. periodic deliveries, Eur J Oper Res 165 (2005), 668–684.
[9] A.M. Campbell and M.W.P. Savelsbergh, A decomposition
approach for the inventory-routing problem, Transp Sci 38
6. SUMMARY AND PERSPECTIVES (2004), 488–502.
[10] C.H. Christiansen and J. Lysgaard, A branch-and-price algo-
We presented several solution approaches to optimize rithm for the capacitated vehicle routing problem with
the determination of visit dates and the corresponding vehi- stochastic demands, Oper Res Lett 35 (2007), 773–781.
cle routes for infectious medical waste collection based on [11] E.W. Dijkstra, A note on two problems in connexion with
stochastic input data. We consider social objectives, such graphs, Numer Math 1 (1959), 269–271.
as the satisfaction of pharmacists and local authority and [12] M. Faccio, A. Persona, and G. Zanin, Waste collection multi
the minimization of public health risks. Experimental results objective model with real time traceability data, Waste Manag
show, that by integrating an RFID technology, the planning 31 (2011), 2391–2405.

94 NETWORKS—2014—DOI 10.1002/net
[13] A. Federgruen and D. Simchi-Levi, Analysis of vehicle [24] P.C. Nolz, N. Absi, and D. Feillet, “Optimization of infec-
routing and inventory-routing problems, Handbooks in Oper- tious medical waste collection using RFID,” Proceedings
ations Research and Management Science 8 (1995), 297– of the Second International Conference on Computational
373. Logistics, ICCL 2011, J.W. Böse, H. Hu, C. Jahn, X. Shi, R.
[14] M. Gendreau, Laporte G., and R. Séguin, Stochastic vehicle Stahlbock, and S. Voß (Editors), Springer, Berlin Heidelberg,
routing, Eur J Oper Res 88 (1996), 3–12. 2011, pp. 86–100.
[15] M. Gendreau, G. Laporte, and R. Séguin, A tabu search [25] T. Nuortio, J. Kytöjoki, H. Niska, and O. Bräysy, Improved
heuristic for the vehicle routing problem with stochastic route planning and scheduling of waste collection and trans-
demands and customers, Oper Res 44 (1996), 469–477. port, Expert Syst Appl 30 (2006), 223–232.
[16] C. Hjorring and J. Holt, New optimality cuts for a single- [26] D. Pisinger and S. Ropke, “Large neighborhood search,”
vehicle stochastic routing problem, Ann Oper Res 86 (1999), Handbook of Metaheuristics, M. Gendreau and J.Y. Potvin
569–584. (Editors), Springer, New York, 2010, 399–419.
[17] T. Kulcar, Optimizing solid waste collection in Brussels, Eur [27] W. Rei, M. Gendreau, and P. Soriano, A hybrid Monte
J Oper Res 90 (1996), 71–77. Carlo local branching algorithm for the single vehicle rout-
[18] G. Laporte and F. Louveaux, “Solving stochastic routing ing problem with stochastic demands, Transp Sci 44 (2010),
problems with the integer L-shaped method,” Fleet Manage- 136–146.
ment and Logistics, T.G. Crainic and G. Laporte (Editors), [28] L. Shih and H. Chang, A routing and scheduling system for
Kluwer Academic Pub, Massachusetts, 1998, 159–167. infectious waste collection, Environ Model Assess 6 (2001),
[19] H.M. Le Blanc, M.G.C. Van Krieken, H. Fleuren, and H.R. 261–269.
Krikke, Collector managed inventory, a proactive planning [29] O. Solyali, J.F. Cordeau, and G. Laporte, Robust inventory
approach to the collection of liquids coming from end-of-life routing under demand uncertainty, Technical Report, 2010.
vehicles, Technical Report, Tilburg University, 2004. [30] J. Teixeira, A.P. Antunes, and J.P. de Sousa, Recyclable waste
[20] S. Lin and B.W. Kernighan, An effective heuristic algorithm collection planning–a case study, Eur J Oper Res 158 (2004),
for the traveling-salesman problem, Oper Res 21 (1973), 543–554.
498–516. [31] F. Tricoire, A. Graf, and W.J. Gutjahr, The bi-objective
[21] J.E. Mendoza, B. Castanier, C. Guéret, A.L. Medaglia, and stochastic covering tour problem, Comput Oper Res 39
N. Velasco, A memetic algorithm for the multi-compartment (2012), 1582–1592.
vehicle routing problem with stochastic demands, Comput [32] D.V. Tung and A. Pinnoi, Vehicle routing-scheduling for
Oper Res 37 (2010), 1886–1898. waste collection in Hanoi, Eur J Oper Res 125 (2000),
[22] M. Miyazaki, T. Imatoh, and H. Une, The treatment of 449–468.
infectious waste arising from home health and medical care [33] World Health Organisation, Health-care waste manage-
services: Present situation in Japan, Waste Manag 27 (2007), ment (October 7, 2010). available at: http://www.who.int/
130–134. mediacentre/factsheets/en/.
[23] M. Mourao and M.T. Almeida, Lower-bounding and heuristic [34] Y. Yu, H. Chen, and F. Chu, A new model and hybrid approach
methods for a refuse collection vehicle routing problem, Eur for large scale inventory routing problems, Eur J Oper Res
J Oper Res 121 (2000), 420–434. 189 (2008), 1022–1040.

NETWORKS—2014—DOI 10.1002/net 95