Robust Optimization For A Maritime Inventory Routing Problem2018Transportation Science

This article was downloaded by: [157.193.8.
33] On: 11 January 2021, At: 02:50

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA
Transportation Science
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org
Robust Optimization for a Maritime Inventory Routing

Problem
Agostinho Agra, Marielle Christiansen, Lars Magnus Hvattum, Filipe Rodrigues
To cite this article:

Agostinho Agra, Marielle Christiansen, Lars Magnus Hvattum, Filipe Rodrigues (2018) Robust Optimization for a Maritime
Inventory Routing Problem. Transportation Science 52(3):509-525. https://doi.org/10.1287/trsc.2017.0814
Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-

Conditions
This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.
The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.
Copyright © 2018, INFORMS
Please scroll down for article—it is on subsequent pages
With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.)
and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual
professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to
transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
TRANSPORTATION SCIENCE
Vol. 52, No. 3, May–June 2018, pp. 509–525
http://pubsonline.informs.org/journal/trsc/ ISSN 0041-1655 (print), ISSN 1526-5447 (online)
Robust Optimization for a Maritime Inventory Routing Problem

Agostinho Agra,a Marielle Christiansen,b Lars Magnus Hvattum,c Filipe Rodriguesa
a Department of Mathematics and Center for Research and Development in Mathematics and Applications, University of Aveiro, 3810-193
Aveiro, Portugal; b Department of Industrial Economics and Technology Management, Norwegian University of Science and Technology,
7491 Trondheim, Norway; c Molde University College, 6402 Molde, Norway
Contact: aagra@ua.pt, http://orcid.org/0000-0002-4672-6099 (AA); mc@ntnu.no, http://orcid.org/0000-0003-1862-6282 (MC);
hvattum@himolde.no, http://orcid.org/0000-0003-0490-9978 (LMH); fmgrodrigues@ua.pt, http://orcid.org/0000-0002-3770-5456 (FR)
Received: September 27, 2016 Abstract. We consider a single product maritime inventory routing problem in which
Revised: April 6, 2017; October 6, 2017; the production and consumption rates are constant over the planning horizon. The prob-
November 15, 2017 lem involves a heterogeneous fleet and multiple production and consumption ports with
Accepted: November 29, 2017 limited storage capacity. Maritime transportation is characterized by high levels of uncer-
Published Online in Articles in Advance: tainty, and sailing times can be severely influenced by varying and unpredictable weather
April 2, 2018
conditions. To deal with the uncertainty, this paper investigates the use of adaptable robust
https://doi.org/10.1287/trsc.2017.0814 optimization where the sailing times are assumed to belong to the well-known budget
polytope uncertainty set. In the recourse model, the routing, the order of port visits, and
Copyright: © 2018 INFORMS the quantities to load and unload are fixed before the uncertainty is revealed, while the
visit time to ports and the stock levels can be adjusted to the scenario. We propose a
decomposition algorithm that iterates between a master problem that considers a subset of
scenarios and an adversarial separation problem that searches for scenarios that make the
solution from the master problem infeasible. Several improvement strategies are proposed
aiming at reducing the running time of the master problem and reducing the number of
iterations of the decomposition algorithm. An iterated local search heuristic is also intro-
duced to improve the decomposition algorithm. A computational study is reported based
on a set of real instances.
Funding: The work of the first author was funded by FCT (Fundação para a Ciência e a Tecnologia)
and CIDMA (Centro de Investigação e Desenvolvimento em Matemática e Aplicações) within
[project UID/MAT/04106/2013]. The second and third authors were supported financially from
the Research Council of Norway through the AXIOM-project. The work of the fourth author was
funded by CIDMA and FCT under [Grant PD/BD/114185/2016].
Keywords: robust optimization • uncertainty • travel time • decomposition • matheuristic • inventory routing
1. Introduction research has formed the basis of several surveys: Papa-

In maritime transportation, large cargo quantities are georgiou et al. (2014), Christiansen et al. (2013), and
transported between ports. Often storages are placed Christiansen and Fagerholt (2009, 2014). In addition,
at or close to the ports at both ends of a sailing leg. The Coelho, Cordeaux, and Laporte (2014) and Andersson
transportation at sea, as well as the storages at ports, et al. (2010) surveyed both land-based and MIRPs.
is most often part of a supply chain from producers Maritime transportation is characterized by high
to end customers. When a single decision maker has levels of uncertainty, and one of the most prevalent
the responsibility for both the transportation of the car- sources of uncertainty is the sailing times that are af-
goes and the inventories at the ports, the routing and fected heavily by changing weather conditions. In prac-
scheduling of the ships and the inventory management tice, unpredictable delays may affect the execution of
can be planned simultaneously. The resulting problem an otherwise optimal deterministic plan. To compen-
is called a maritime inventory routing problem (MIRP). sate for such delays, it is possible for ships to speed
The shipping industry is capital intensive with high up when necessary. However, in practice it will most
investment and operating costs for the ships as well as often be beneficial to consider the uncertainty explic-
large and valuable cargoes, so a modest improvement in itly when finding the optimal plan.
the fleet utilization can imply a large increase in profit. Therefore, we consider an MIRP with uncertain sail-
Therefore, the MIRP is a very important and common ing or traveling times. A heterogeneous fleet of ships
problem in maritime shipping. These reasons, as well is transporting a single product between ports. There
as the difficulties to solve the problem because of high is one set of ports where the product is produced, and
degrees of freedom in the routing, scheduling, number another set of ports where the product is consumed.
of port visits, and the loaded and unloaded quantity, The production and consumption rates are assumed
have led to a solid amount of research on MIRPs. This constant over the planning horizon. In all ports, there
509
Agra et al.: Robust Optimization for a Maritime Inventory Routing Problem
510 Transportation Science, 2018, vol. 52, no. 3, pp. 509–525, © 2018 INFORMS
exists a storage for the product, and lower and upper Agra et al. (2016) to be able to solve larger instances of
inventory limits are given for each storage. Each port the MIRP. Additionally, the authors explain why using
can be visited once or several times during the plan- penalties for backlogged demands make the determin-
ning horizon depending on the size of the storage, istic problem much harder, which also motivates the
the production or consumption rate, and the quantity recourse to robust approaches for MIRP.
loaded or unloaded at each port visit. The MIRP with Zhang et al. (2018) (see also Zhang 2015) developed
uncertain traveling times consists of designing routes robust approaches for an annual delivery plan problem
and schedules for a fleet of ships that are robust against involving a single producer and multiple customers in
delay in traveling times to minimize the transportation the liquefied natural gas business. First, an MIRP with
and port costs, and to determine the quantities han- given time windows for deliveries with uncertain travel
dled at each port call without exceeding the storage disruptions is solved by use of a Lagrangian heuristic
capacities. scheme to obtain robust solutions. Second, a more gen-
Even though maritime transportation is heavily influ- eral robust MIRP with time windows is studied, where
enced by uncertainty, most of the research reported the length and placement of the time windows are also
in the literature on maritime routing and scheduling decision variables. The problem is formulated as a two-
considers static and deterministic problems. We review stage stochastic mixed-integer program (MIP), and the
some of the existing contributions within maritime authors propose a two-phase solution approach that
transportation considering uncertainties. considers a sample set of disruptions as well as their
For a ship routing and scheduling problem with recovery solutions.
predefined cargoes, Christiansen and Fagerholt (2002) Robust IRPs have also been considered in land trans-
design ship schedules that are less likely to result in portation, but the uncertainty is related to the demands.
ships staying idle at ports during weekends by impos- Solyali, Cordeau, and Laporte (2012) proposed a dual-
ing penalty costs for arrivals at risky times (i.e., close ization approach, while Agra et al. (2016) developed
to weekends). The resulting schedule needs to be more
a decomposition approach for inventory models that
robust with respect to delays from bad weather and
can be combined with routing and uncertain demands.
unpredictable time in port because of the restricted
More general robust inventory problems have been con-
operating hours each day and ports being closed dur-
sidered; see, for instance, Bertsimas and Thiele (2006).
ing weekends. Agra et al. (2013) solved a full-load
However the inventory problems usually assume the
ship routing and scheduling problem with uncer-
time is discretized into a finite set of time periods,
tain travel times using robust optimization. Further-
which contrasts with our problem where the time is
more, Halvorsen-Weare and Fagerholt (2011) analyzed
considered continuous. For recent overviews on robust
various heuristic strategies to achieve robust weekly
optimization see Ben-Tal, El Ghaoui, and Nemirovski
voyages and schedules for off-shore supply vessels
working under tough weather conditions. Heuristic (2009); Bertsimas, Brown, and Caramanis (2011); Goris-
strategies for obtaining robust solutions with uncertain sen, Yanikoğlu, and den Hertog (2015); and Gabrel,
sailing times and production rate were also discussed Murat, and Thiele (2014).
by Halvorsen-Weare, Fagerholt, and Rönnqvist (2013) Uncertainty has been considered for other related
for the delivery of liquefied natural gas. routing problems recently. Roldán, Basagoiti, and
For a crude oil transportation and inventory prob- Coelho (2016) present three new customer selection
lem, Cheng and Duran (2004) developed a decision methods for a dynamic and stochastic IRP. Aghezzaf
support system that takes into account uncertainty in (2008) considers a variant of the inventory routing opti-
sailing time and demand. The problem was formu- mization problem where customer demand rates and
lated as a discrete time Markov decision process and travel times are stochastic but stationary. He proposes
solved by using discrete event simulation and opti- an approach to generate optimal robust distribution
mal control theory. Rakke et al. (2015) and Sherali and plans. Li, Wang, and Chan (2016) consider an IRP under
Al-Yakoob (2006a, b) introduced penalty functions for replenishment lead time where the inventory levels
deviating from the customer contracts and the storage are uncertain. They propose an optimization approach
limits, respectively, for their MIRPs. Christiansen and based on a genetic algorithm. Bertsimas, Gupta, and
Nygreen (2005) used soft inventory levels to handle Tay (2016) introduce a scalable approach for solving a
uncertainties in sailing time and time in port, and these robust and adaptive mixed-integer formulation for an
levels were transformed into soft time windows for a IRP with demand uncertainty. Desaulniers, Rakke, and
single product MIRP. Agra et al. (2015) were the first to Coelho (2016) consider a new mathematical formula-
use stochastic programming to model uncertain sailing tion for the IRP and develop a state-of-the-art branch-
and port times for an MIRP with several products and price-and-cut algorithm for solving the problem. A sur-
inventory management at the consumption ports only. vey on the IRP with stochastic lead times and demands
Recently, a heuristic stochastic approach is presented in can be found in Roldán, Basagoiti, and Coelho (2017).
Transportation Science, 2018, vol. 52, no. 3, pp. 509–525, © 2018 INFORMS 511
Zhang et al. (2015) study a robust MIRP with time number of iterations of the decomposition algorithm.
windows and stochastic travel times, where the length Most of these improvements can be extended to other
and placement of the time windows are decision vari- related problems solved by robust optimization.
ables. The problem is modeled as a two-stage stochas- (iv) The paper presents a new iterated local search
tic mixed-integer program, and a two-phase heuristic heuristic. The heuristic provides a good-quality solu-
solution approach is proposed. tion and can also be extended to other robust optimiza-
Adulyasak and Jaillet (2016) consider the vehicle tion problems. The heuristic is used to improve the
routing problem with deadlines under travel time un- exact decomposition approach.
certainty, discussing both stochastic and robust prob- The rest of the paper is organized as follows: The
lem variants. Adulyasak, Cordeau, and Jans (2015) mathematical model of the deterministic problem is
solve the land production routing problem under de- presented in Section 2, while the robust optimization
mand uncertainty considering a stochastic setting by model and the decomposition algorithm are described
using a Benders decomposition approach with several in Section 3. Section 4 is devoted to improvement strate-
enhancements. Other heuristic algorithms for the pro- gies for the decomposition approach. An iterated local
duction routing problem were proposed by Solyali and search heuristic is presented in Section 5. Furthermore,
Süral (2017) and Russell (2017). A comparison between computational results are reported and discussed in
two scenario-based frameworks, a stochastic program- Section 6, followed by some concluding remarks in
ming and robust optimization approach, for supply Section 7.
planning under uncertainty is provided by Maggioni,
Potra, and Bertocchi (2017). 2. Mathematical Model for the
The objective of this paper is to present a gen-
Deterministic Problem
eral robust optimization procedure for solving a sin-
In this section we present a mathematical formulation
gle product MIRP with uncertain traveling times that
for a deterministic version of our MIRP.
results in robust solutions that are immune to some
sailing time delays and where the inventory limits are Routing Constraints. Let V denote the set of ships and
not violated because of the delays. In the robust model, N denote the set of ports. Each ship v ∈ V must depart
the traveling times belong to an uncertainty set, which from its initial position, which is either a port or a point
we assume to be the well-known budget constrained at sea. For each port, we consider an ordering of the
polytope introduced by Bertsimas and Sim (2003). The visits according to the time of the visit.
total deviation of the traveling times to the nominal The ship paths are defined on a network where the
values is controlled by a parameter. This set has the nodes are represented by a pair (i, m), where i indicates
advantage that it is easy to interpret from a practi- the port and m indicates the visit number to port i.
cal point of view, and its structure can be explored Direct ship sailings (arcs) from node (i, m) to node (j, n)
from a computational point of view when decomposi- are represented by (i, m, j, n). Figure 1 depicts two ship
tion techniques are employed (Agra et al. 2013, 2016; paths. Ship 1 leaves its origin, sails to Port 1 for the first
Bienstock and Özbay 2008). visit, then Port 2 is visited for the first time, and finally
In relation to the existing literature, this paper pro- ship 1 terminates its route servicing at Port 3. This is
vides the following contributions: the second visit to Port 3, because Ship 2 visited Port 3
(i) The paper introduces a robust model to an MIRP first on its route to Port 1.
to derive solutions that are immune to a certain num- We define S A as the set of possible nodes (i, m), SvA
ber of delays in relation to inventory level deviations. as the set of nodes that may be visited by ship v, and
This model assumes that the routing, number of port set SvX as the set of all possible sailings (i, m, j, n) of
visits, and the quantities to load and unload cannot ship v. For the routing, we define the following binary
be adjusted to the uncertain scenario, while the time
for start of service as well as the inventory levels are Figure 1. Example of Two Ship Routes
adjustable.
(ii) The paper develops a decomposition algorithm, 1, 1 1, 2 D2
where the problem is relaxed into a master problem
O1
and each robust constraint is written for a small subset
of scenarios only, and a separation subproblem that 2, 1
checks whether the solution is feasible for the omitted
robust constraints. O2
(iii) The paper introduces several improvement
3, 1 3, 2 D1
strategies for the decomposition algorithm. One set of
improvements aims to reduce the running time of each Note. The route of Ship 1 is represented by solid lines and the route
master problem, while the other intends to reduce the of Ship 2 is represented by dashed lines.
variables: ximjnv is 1 if ship v travels from node (i, m) O

f imv Q vO x imv
O
, ∀ v ∈ V, (i, m) ∈ SvA , (11)
O
directly to node (j, n), and 0 otherwise; x imv indicates fimjnv ≤ C v ximjnv , ∀ v ∈ V, (i, m, j, n) ∈ SvX , (12)
whether ship v travels directly from its initial position
to node (i, m) or not; w imv is 1 if ship v visits node (i, m), Q i w imv ≤ q imv ≤ min{C v , Q̄ i }w imv ,
¯
and 0 otherwise; z imv is equal to 1 if ship v ends its route ∀ v ∈ V, (i, m) ∈ SvA , (13)
at node (i, m), and 0 otherwise; z vO is equal to 1 if ship v fimjnv ≥ 0, ∀v ∈ V, (i, m, j, n) ∈ SvX , (14)
is not used and 0 otherwise; yim indicates whether a
ship is making the mth visit to port i, (i, m), or not. The
O
f imv , q imv ≥ 0, ∀ v ∈ V, (i, m) ∈ SvA . (15)
parameter µ i denotes the minimum number of visits at
port i and ¯the parameter µ̄ i denotes an upper bound Equations (10) are the flow conservation constraints at
on the number of visits at port i node (i, m). Equations (11) determine the quantity on
X ship v when it travels from its initial node to node
O
x imv + z vO 1, ∀ v ∈ V, (1) (i, m). Constraints (12) require that the ship capacity
(i, m)∈S vA
X is obeyed. Constraints (13) impose lower and upper
w imv − xjnimv 0, ∀ v ∈ V, (i, m) ∈ SvA , (2) limits on the loading and unloading quantities. Con-
(j, n)∈S vA
X straints (14)–(15) are the nonnegativity constraints.
w imv − ximjnv − z imv 0, ∀ v ∈ V, (i, m) ∈ SvA , (3)
(j, n)∈S vA
Time Constraints. We define the following parame-
Q
X ters: Ti is the time required to load or unload one unit
w imv yim , ∀ (i, m) ∈ S A , (4)
v∈V
of product at port i and Ti jv is the travel time between
yim 1, ∀ (i, m) ∈ S A : m ∈ {1, . . . , µ i }, (5) port i and j by ship v. The travel time also includes any
¯ setup time required to operate at port j. Parameter TivO
yi(m−1) − yim ≥ 0, ∀ (i, m) ∈ S : µ i + 1 < m ≤ µ̄ i ,
A
(6) indicates the traveling time required by ship v to travel
¯
ximjnv ∈ {0, 1}, ∀ v ∈ V, (i, m, j, n) ∈ SvX , (7) from its initial position to facility i. Parameter TiB is
O
x imv , w imv , z imv ∈ {0, 1}, ∀ v ∈ V, (i, m) ∈ SvA , (8) the minimum time between two consecutive visits to
port i. Parameter T is the length of the time horizon,
yim ∈ {0, 1}, ∀ (i, m) ∈ S . A
(9)
and A im and Bim are the time windows for starting the
Equations (1) ensure that each ship either departs from mth visit to port i. Such time windows are considered
its initial position and travels to another node or the only for generality, since here we will consider A im 0,
ship is not used. Equations (2) and (3) are the flow con- and Bim T. To ease the presentation, we also define,
servation constraints, ensuring that a ship arriving at for each node (i, m), the following upper bound for the
a node either leaves that node or ends its route. Con- end time of the visit: Tim0 Q
min{T, Bim + Ti Q̄ i }. Given
straints (4) ensure that a ship can visit node (i, m) only time variables t im that indicate the start time of each
if yim is equal to one. Equations (5) fix yim to 1 for the visit at each port, the time constraints can be written as
mandatory visits. Constraints (6) state that if port i is
visited m times, then it must also have been visited X
t im + 0
max{Tim +Ti jv −A jn ,0}ximjnv
m − 1 times. Constraints (7)–(9) define the variables as v∈V:(i, m, j, n)∈S vX
binary. X Q
+ 0
Ti ·q imv −t jn ≤Tim −A jn , ∀ (i, m),(j, n)∈S A , (16)
Loading and Unloading Constraints. Parameter Ji is 1 v∈V
if port i is a producer and −1 if it is a consumer. The X Q
t im −t i, m−1 − Ti q i, m−1, v −TiB yim ≥0,
quantity on ship v at the beginning of the planning v∈V
horizon is given by Q vO , and the capacity of ship v ∀ (i, m)∈SA : m >1, (17)
X
is denoted by C v . The minimum and maximum load- TivO x imv
O
≤ t im , ∀ (i, m)∈S , A
(18)
ing and unloading quantities at port i are given by Q i v∈V
and Q̄ i , respectively. ¯ A im ≤ t im ≤ Bim , ∀ (i, m)∈S A . (19)

To model the loading and unloading constraints,
we define the following continuous variables: q imv is Constraints (16) relate the start time associated with
the amount loaded or unloaded from ship v at node node (i, m) to the start time associated with node ( j, n)
(i, m); fimjnv denotes the amount that ship v transports when ship v travels directly from (i, m) to (j, n). Con-
O
from node (i, m) to node (j, n), and f imv gives the straints (17) impose a minimum interval between two
amount that ship v transports from its initial position consecutive visits at port i. Constraints (18) ensure that
to node (i, m). The loading and unloading constraints if ship v travels from its initial position to (i, m), then
are given by the start time associated with (i, m) is at least the trav-
X X
O
f imv + fjnimv + Ji q imv fimjnv , eling time between the initial position and port i. Time
(j, n)∈S vA (j, n)∈S vA windows for the start time of visits are given by con-
∀ v ∈ V, (i, m) ∈ SvA , (10) straints (19).
Inventory Constraints. The inventory constraints are 3.1. Mathematical Model for the
considered for each port. They ensure that the stock Robust Formulation
levels are within the corresponding limits and link the In the robust model the traveling times belong to an
stock levels to the loading or unloading quantities. For uncertainty set. The uncertainty set represents the situ-
each port i, the rate of consumption or production, R i , ation where there can be at most a number Γ of delays
the minimum Si , the maximum S̄i , and the initial Si0 in the ship paths. Instead of the traveling time Ti jv ,
stock levels are¯given. We define the nonnegative con- we might add a delay T̂i jv to a nominal traveling time
tinuous variables s im to represent the stock levels at the value T̄i jv . As the traveling times do not depend on
start of the mth visit to port i. The inventory constraints the visits, one would increase all of the traveling times
are as follows: between i and j if Ti jv is increased. This increase affects
s i1 Si0 + Ji R i t i1 , ∀ i ∈ N, (20) mostly those routes where a ship sails multiple times
X directly between the same two ports. To have full con-
s im s i, m−1 − Ji q i, m−1, v + Ji R i (t im − t i, m−1 ), trol on the number of delays, we replace Ti jv by ξimjnv
v∈V
∀ (i, m) ∈ S A : m > 1, (21) in constraints (16) and TivO by ξ imv
O
in constraints (18)
Q
X X
s im + q imv − R i Ti q imv ≤ S̄i , X
v∈V v∈V
t im + 0
max{Tim + ξimjnv −A jn , 0}x imjnv
∀ (i, m) ∈ S A : Ji −1, (22) v∈V | (i, m, j, n)∈S vX
Q
X X
q imv + R i Ti q imv ≥ Si , Q
X
s im − + 0
Ti q imv − t jn ≤ Tim −A jn , ∀ (i, m), (j, n) ∈ S A , (29)
v∈V v∈V ¯
∀ (i, m) ∈ S A : Ji 1, (23) Xv∈VO
X ξ imv x imv
O
≤ t im , ∀ (i, m) ∈ S A . (30)
s i µ̄i + q i, µ̄i , v − R i (T − t i µ̄i ) ≥ Si , ∀ i ∈ N: Ji −1, (24) v∈V
v∈V ¯
For ease of notation we will consider in the following
X
s i µ̄i − q i, µ̄i , v + R i (T − t i µ̄i ) ≤ S̄i , ∀ i ∈ N: Ji 1, (25) O
v∈V Timv as a particular case of Tjnimv where j is the initial
s im ≥ Si , ∀ (i, m) ∈ S A : Ji −1, (26) position of ship v, denoted as o(v), and n will be 1.
¯ The uncertainty set is now defined using the traveling
s im ≤ S̄i , ∀ (i, m) ∈ S A : Ji 1. (27)
times that depend on the visits, as follows:
Equations (20) calculate the stock level at the start time
of the first visit to a port, and Equations (21) relate Ξ ξ: ξimjnv T̄i jv + T̂i jv δ imjnv , 0 ≤ δinjmv ≤ 1,
Γ
the stock level at the start time of the mth visit to
the stock level at the start time of the previous visit.
X X
v ∈ V, (i, m, j, n) ∈ SvX , δ imjnv ≤ Γ .
Constraints (22) and (23) ensure that the stock lev- v∈V (i, m, j, n)∈S vX
els are within their limits at the end of each visit.
Constraints (24) impose a lower bound on the inven- This uncertainty set is the well-known budget polytope
tory level at time T for consumption ports, while con- introduced by Bertsimas and Sim (2004), where T̄i jv is
straints (25) impose an upper bound on the inventory the nominal value corresponding to the expected travel
level at time T for production ports. Constraints (26) time, T̂i jv is the maximum allowed deviation (delay),
and (27) ensure that the stock levels are within their δinjmv is the deviation of parameter Timjnv from its nom-
limits at the start of each visit. inal value, and Γ limits the number of deviations.
Objective Function. The objective is to minimize the The model introduced here is an adjustable robust
total routing costs, including traveling and operating program (Billionnet, Costa, and Poirion 2014; Ben-Tal
costs. The traveling cost of ship v from port i to port j is et al. 2004; Delage and Iancu 2015; Thiele, Terry, and
denoted by C Tijv and it includes the setup costs. Param- Epelman 2010), which features two levels of decisions:
eter C TO first-stage variables must be fixed before the uncer-
iv represents the traveling cost of ship v from
its initial position to port i. The objective function is tainty is revealed, while adjustable variables can react
defined as follows: to account for the uncertainty. Such a concept has
X X also been known as recoverable robustness (Liebchen
min C(X) C Tijv ximjnv et al. 2009).
v∈V (i, m, j, n)∈S vX
X X The first-stage variables ximjnv , z imv , w imv , yim , and q imv
+ iv x imv .
C TO O
(28) are the routing, the port visits sequence, and the load-
v∈V (i, m)∈S vA ing and unloading decisions. The adjustable variables
are those related to the time and the stock levels. These
3. Robust Optimization variables now depend on the uncertain parameters.
In this section we first present the robust optimiza- Hence, we define t im (ξ), and s im (ξ) as the time and the
tion model and then describe the solution method stock level of visit (i, m), respectively, when scenario ξ
proposed. (vector of travel times) is revealed.
The first-stage solution must ensure that, for each restricted to the set of extreme points ext(ΞΓ ) will be
possible vector of travel times in the budget polytope, denoted by R-IR(ext(ΞΓ )).
the stock level at each port i is within the inventory To frame our work within robust optimization, we
bounds Si and S̄i . For the robust model the time and make the following remarks:
¯ constraints are replaced by the following
inventory Remark 1. (a) The model R − IR(Θ) has fixed recourse
constraints: since the coefficients of the second-stage variables in
Time constraints: the objective function and constraints are determin-
X istic. Every variable is either nonadjustable or fully
t im (ξ)+ 0
max{Tim +ξimjnv −A jn ,0}x imjnv adjustable, and the uncertainty set can be considered
v∈V: (i, m, j, n)∈S vX as a scenario-generated uncertainty set.
Q
X
−t jn (ξ)+ 0
Ti q imv ≤Tim −A jn , (b) The deterministic problem is NP-hard, and it is
v∈V quite complex since it generalizes and combines NP-
∀ (i, m, j, n)∈S X , ξ ∈ΞΓ , (31)
hard problems such as the vehicle routing problem
Q
X
t im (ξ)−t i, m−1 (ξ)− Ti q i, m−1, v −TiB yim ≥0, with time windows (VRPTW) and inventory problems.
v∈V Current research is being conducted on that (determin-
∀ (i, m)∈SA : m >1, ξ ∈ΞΓ , (32) istic) problem, such as valid inequalities and extended
X
ξ o(v)1imv x imv
O
≤ t im (ξ), ∀ (i, m)∈S A , ξ ∈ΞΓ , (33) formulations. As the focus is on the robust approaches,
v∈V we will restrict ourselves to established results to the
A im ≤ t im (ξ)≤ Bim , ∀ (i, m)∈S A , ξ ∈ΞΓ . (34) deterministic case and focus only on small/medium
size instances that can be solved to optimality. Remarks
Inventory management constraints: will be made on how the results can be extended to
larger instances.
s i1 (ξ)Si0 + Ji R i t i1 (ξ), ∀ i ∈ N, ξ ∈ ΞΓ , (35) (c) The value of the second-stage (fully adjustable)
X
s im (ξ) s i, m−1 (ξ)− Ji q i, m−1, v + Ji R i (t im (ξ)− t i, m−1 (ξ)), variables do not directly affect the objective function
v∈V value. That is, for given routes and pickup and delivery
∀ (i, m) ∈ S A : m > 1, ξ ∈ ΞΓ , (36) quantities (first-stage decisions) all of the robust feasi-
ble solutions lead to the same objective function value.
X X Q
s im (ξ)+ q imv − R i Ti q imv ≤ S̄i ,
v∈V v∈V This is common for the robust VRPTW problem given
∀ (i, m) ∈ S A : Ji −1, ξ ∈ ΞΓ , (37) in Agra et al. (2013) where the second-stage variables
X X Q (the time of the visit to each node) do not affect the
s im (ξ)− q imv + R i Ti q imv ≥ Si ,
v∈V v∈V ¯ value of the solution. As a result, any policy for the
adjustable variables that finds a feasible solution for
∀ (i, m) ∈ S A : Ji 1, ξ ∈ ΞΓ , (38)
X each scenario when such a solution exists, is an optimal
s i µ̄i (ξ)+ q i, µ̄i , v − R i (T − t i µ̄i (ξ)) ≥ Si , policy.
v∈V ¯
∀ i ∈ N: Ji −1, ξ ∈ ΞΓ , (39) As we are dealing with fixed recourse, in theory the
second-stage variables, t im (ξ) and s im (ξ), can be elimi-
X
s i µ̄i (ξ)− q i, µ̄i , v + R i (T − t i µ̄i (ξ)) ≤ S̄i ,
v∈V nated. That is, the set of feasible solutions can be pro-
∀ i ∈ N: Ji 1, ξ ∈ ΞΓ , (40) jected onto the space of the first-stage variables. This
s im (ξ) ≥ Si , ∀ (i, m) ∈ S : Ji −1, ξ ∈ Ξ ,
A Γ
(41) approach was followed in Atamtürk and Zhang (2007)
¯ for a simpler two-stage robust network flow and design
s im (ξ) ≤ S̄i , ∀ (i, m) ∈ S A : Ji 1, ξ ∈ ΞΓ . (42)
problem, and in Agra et al. (2013) robust approaches
The robust model is defined by (1)–(15), (28), and were used for the two-dimensional spaces (with first-
(31)–(42). stage and with first- and second-stage variables) for the
robust VRPTW problem, albeit the projection has not
3.2. Model Analysis been done explicitly. In Atamtürk and Zhang (2007) it
The model has an infinite number of variables t im (ξ) was shown that even for simple graph structures the
and s im (ξ), as well as time constraints and inventory separation of the inequalities resulting from the projec-
constraints. However, as the recourse model (31)–(42) is tion is NP-hard. In Agra et al. (2013) a path formula-
a pure linear model with no binary variables, similar to tion was used for the formulation in the first-stage vari-
Lemma 1 in Agra et al. (2013), it is easy to show that the ables space and the corresponding approach was not
uncertainty set can be restricted to the extreme points considered preferable than the one working in the orig-
of the budget polytope. That is, ξ ∈ ext(ΞΓ ) in con- inal space. As in our case, the projection would become
straints (31)–(42), where ext(ΞΓ ) is the set of extreme much more complex, so we opted to eliminate only the
points of ΞΓ . stock s im (ξ) variables. Moreover, keeping the time vari-
For a given set Θ of scenarios the robust model ables t im (ξ) allows us to use implicitly the policy of
will be denoted by R-IR(Θ). Hence, the finite model assigning to t im (ξ) the earliest possible time of the visit.
Next we eliminate the s im (ξ) variables. Equa- as clear as in the case where time periods are consid-
tions (35) and (36) allow us to write stock variables as ered. A reasonable possibility would be to write t im (ξ)
functions of the time visits as an affine function of the traveling times
X X X
m−1
t im (ξ) ηkl
im
jn ξ kl jnv + ηo(v)1im
im
ξ o(v)1imv + η0im ,
XX
s im (ξ) Si0 + Ji R i t im (ξ) − Ji q inv , v∈V (k, l, j, n)∈S vX v∈V
v∈V n1
∀ (i, m) ∈ S A , ξ ∈ ΞΓ . (43) where the new (nonadjustable) variables are ηkl im
jn
and ηo(v)1im
im
.
This approximation has two main drawbacks: (i) it
Using Equations (43), constraints (39) and (40) are con- does not take into account other factors that are rele-
verted into the following constraints: vant to define the time of the visits, such as the inven-
tory levels and the time between consecutive visits;
X X thus it may restrict the solution space and deteriorate
Si0 + q imv ≥ R i T + Si , ∀ i ∈ N: Ji −1, (44)
v∈V m: (i, m)∈S A ¯ the quality of the solution. (ii) As we do not have time
X X periods, all possible ship paths must be considered
Si0 + R i T ≤ q imv + S̄i , ∀ i ∈ N: Ji 1. (45) leading to a large number of variables. Comparing to
v∈V m: (i, m)∈S A
the (exact) row-column decomposition approach dis-
cussed in Section 3.3 we can see that, in general, the
These constraints depend only on the first-stage deci- number of variables used in the approximation would
sions (decisions with no recourse). be larger.
For consumption ports, constraints (37) and (41) Overall, from Remark 1(c), optimal policies can be
imply, respectively, easily derived for the adjustable variables that can be
efficiently used from a computational point of view
m as discussed below. Thus, the approximation decision
Q
XX X
R i t im (ξ) ≥ q inv − R i Ti q imv + Si0 − S̄i , rules do not seem to help in our case as they may dete-
v∈V n1 v∈V riorate the quality of the solution and do not seem to
∀ (i, m) ∈ S A : Ji −1, ξ ∈ Ξ∗ , (46) simplify the problem much.
m−1
3.3. Solution Method
XX
R i t im (ξ) ≤ q inv + Si0 − Si ,
v∈V n1 ¯ Though finite, the model R − IR(ext(ΞΓ )) tends to be
∀ (i, m) ∈ S A : Ji −1, ξ ∈ Ξ∗ . (47)
very large because the number of extreme points of
the uncertainty polytope tends to grow rapidly with
For loading ports, constraints (38) and (42) imply the problem size. However, as it happens with many
MIP models that are defined through a large number
m of constraints but where only a few of them need to be
Q
XX X
R i t im (ξ) ≥ q inv − R i Ti q imv − Si0 + Si , included in the model, an efficient solution method can
v∈V n1 v∈V ¯ be designed that can be seen as a variant of the Benders’
∀ (i, m) ∈ S A : Ji 1, ξ ∈ Ξ∗ , (48) decomposition approach. The main difference is that
m−1
XX here both columns and variables are added. Within
R i t im (ξ) ≤ q inv + S̄i − Si0 , robust optimization such approaches became popular
v∈V n1 recently; see Agra et al. (2013, 2016), Bienstock and
∀ (i, m) ∈ S A : Ji 1, ξ ∈ Ξ∗ . (49)
Özbay (2008), Thiele, Terry, and Epelman (2010), and
Zeng and Zhao (2013). This procedure is also known
In both cases ξ ∈ Ξ∗ ΞΓ . as the adversarial approach (Gorissen, Yanikoğlu, and
Henceforward we consider the robust model as den Hertog 2015).
the model defined by the objective (28) and con- The decomposition approach works as follows: the
straints (1)–(15) and (44)–(49). For a given set of scenar- problem is relaxed into a master problem (MP), where
each robust constraint is written only for a small subset
ios Ξ∗ , it will be denoted by R − IR(ext(Ξ∗ )).
Ξ∗ ⊂ ext(ΞΓ ). Given a feasible solution to the MP, we
check whether the solution is feasible for the omitted
Remark 2. A common approach to handle adjustable robust constraints by solving an adversarial separation
robust problems is to employ approximation tech- problem (ASP). If a scenario leading to infeasibility is
niques such as the well-known affine decision rules; see found, we expand Ξ∗ and the corresponding columns
Ben-Tal et al. (2004); and Bertsimas, Iancu, and Parrilo and rows are added to the MP and the augmented MP
(2011). In our problem, the approximation to use is not is solved again.
The MP is defined by the objective (28), and con- For each port visit, we compute the earliest time of
straints (1)–(15) and (44)–(48) defined for a subset Ξ∗ ⊂ the visit and then check whether this time is below
ext(ΞΓ ). Given a first-stage solution, the ASP amounts the upper time window limit. A recursive approach
to check whether constraints (46)–(48) are satisfied is followed. Let α((i, m), γ) be the earliest arrival
for all ξ ∈ ext(ΞΓ ), and if not, add the corresponding time at (i, m) when γ arcs are using their maximum
violated constraints. The decomposition procedure is travel time. For γ 0, . . . , Γ, the value of α((i, m), γ) is
summarized in Algorithm 1. given by
Algorithm 1 (A column-and-row generation approach for 
 A im ,
the robust problem)

im ,

Ā


1: Initialize the subset Ξ∗ ⊆ ext(ΞΓ ) ← {ξ 0 } where

 X Q
α((i, m − 1), γ) +

 Ti q̄ i, m−1, v
ξ 0 is the scenario with no delays




 v∈V
2: Solve the restricted R − IR(Ξ∗ ) problem α((i, m), γ) max + Ti , m > 1,
B
3: while There is a scenario ξ ∗ in ext(ΞΓ )\Ξ∗ , leading 
Q
α((j, n), γ) + T j q̄ jnv + T̄ jiv : x̄jnimv 1,



to a violation of a constraint (46)–(48) do 

 α((j, n), γ − 1) + T jQ q̄ jnv

4: Add ξ ∗ to Ξ∗ (and add the corresponding



variables and rows to the model)  + T̄ jiv + T̂i jv : x̄jnimv 1,


5: Solve the new restricted R − IR(Ξ∗ ) problem
6: end while. where α((i, m), γ) −∞ if γ < 0. A im is the lower time
window limit to the visit, and Ā im is the lower limit
3.4. Adversarial Separation Problem forced by the stock level, as derived above. The third
Here we discuss the ASP considered in step 3 of Algo- expression accounts for the case where the earliest visit
rithm 1. Given a solution that considers the scenarios time results from the constraint imposing a time limit
in Ξ∗ , the ASP checks whether this solution is feasible between consecutive visits to the same port (forced
for the remaining scenarios, that is, when the first-stage by constraints (32)). The fourth and fifth expressions
decisions are fixed, we check if there is a set of at most account for the case where the service time results from
Γ delays that leads to a violation of an inventory level the visit to another port j (implied by constraints (31)).
(lower or upper bound) constraint. The fourth expression is for the case where there is no
Assume that the values of the first-stage variables delay in the traveling time, and the fifth expression is
are given by x̄imjnv , z̄ imv , w̄ imv , ȳim , and q̄ imv . We can see for the case where a delay occurs.
that when variables q imv are set to q̄ imv , the robust con- The values α((i, m), γ) are computed following the
straints (46)–(48) define time windows Ā im ≤ t im ≤ B̄im order given by the times t im observed in the solution
for each visit (i, m). of the MP. This is enough to ensure that α((i, m), γ)
Inequalities (46)–(47) imply the following time win- is computed after the value α((i, m − 1), γ) and values
dows for t im , with (i, m) ∈ S A | Ji −1, ξ ∈ ΞΓ : α((j, n), γ), α((j, n), γ − 1), when x̄jnimv 1 in the MP
solution.
q̄ inv + Si0 − S̄i The earliest starting time of visit (i, m) is ᾱ(i, m)
P Pm
Q v∈V n1
X
− Ti q̄ imv + max{α((i, m), γ): γ ∈ {1, . . . , Γ}}. After computing
Ri
|
v∈V
{z } ᾱ(i, m) for each port visit (i, m), we need to check
Ā im whether ᾱ(i, m) ≤ min{Bim , B̄im }. If there is a violated
inequality, that means there is a scenario leading to
q̄ inv + Si0 − Si
P Pm−1
v∈V n1
≤ t im (ξ) ≤ ¯ ; an inventory limit violation, and the constraints and
Ri variables corresponding to that scenario are added.
| {z }
Note that this separation algorithm generalizes the one
B̄ im
given in Agra et al. (2013) where only time windows
and, inequalities (48)–(49) imply the following time are considered.
windows for Ji 1, ξ ∈ ΞΓ : Example 1. Consider the example with T 20, V
{1, 2}, and N {1, 2, 3}. Port 1 is a supply port
+ Si
Pm 0
n1 q̄ inv − S i
P
Q v∈V
X
− Ti q̄ imv + ¯ and Ports 2 and 3 are demand ports. The produc-
v∈V Ri tion/demand rates and the initial stock levels are given
by R 1 5, R 2 1, R 3 2 and S10 22, S20 10, S30 10,
| {z }
Ā im
respectively. Consider Si 0 and S̄i 50, for all i
P
v∈V
Pm−1
n1 q̄ inv + S̄i − Si0 ¯
1, 2, 3. The quantities loaded/unloaded in the solution
≤ t im (ξ) ≤ .
Ri are q111 37, q 211 10, q 321 22, q312 8, and q122 45.
| {z } The initial load onboard the ships are Q10 0 and
B̄ im Q 20 8, and assume the operation times are negligible
Figure 2. Example of Computation of α((i, m), γ) for Γ 2 add a subset of those inequalities that has proved to
for an Instance with Three Ports and Two Ships provide improvements on the integrality gap of related
((1, 1), 0) = 3 problems.
Let Di denote the total net demand of port i ∈ V
1, 1 1, 2 during the planning horizon. For i ∈ N with Ji −1,
2
Di T × R i − Si0 + Si . For i ∈ N with Ji 1, Di T × R i +
O1 3+1 Si0 − S̄i . Then, the ¯following integer set, for i ∈ N, is a
4 relaxation of the feasible set:

2, 1 |V |
X
Θi χ ∈ + : C v χv ≥ Di ,
((2, 1), 1) = 7 2+1 v∈V
O2
3 ((3, 2), 2) = 10
where X
χv w imv ,
3, 1 3, 2 m: (i, m)∈S vA
denotes the number of times vehicle v visits port i dur-

Q Q
ing the planning horizon T.
T1 T2 0. The routes are depicted in Figure 2. Valid inequalities for Θi , i ∈ N are valid for the set of
The values assigned to each arc represent the travel- feasible solutions. A particular case of these inequali-
ing times. We consider an uncertainty set with Γ 2 ties is the following integer rounding cut:
and T̂i jv 1 for all (i, j, v). The plus 1 next to the
values assigned to arcs (1, 1, 2, 1) and (2, 1, 3, 2) rep-
X X Cv Di
w imv ≥ , (50)
resent the delay of one unit. The values α((i, n), γ) v∈V m: (i, m)∈S vA Q Q
are given for the critical path leading to node (3, 2).
The value α((1, 1), 0) is given by Ā11 , which is com- where Q can be any positive number. We consider a
puted from the initial stock (S10 22), the production different inequality for each v ∈ V, taking Q Q v .
rate R 1 5, and the load quantity (q 111 37). The val-
Initial Primal Bound
ues of α((2, 1), 1) and α((3, 2), 2) are computed from
Q A common upper bound can be obtained by solv-
the expression α((i, m), γ) α((j, n), γ − 1) + T j q̄ jnv + ing the well-known box-constrained problem by con-
T̄ jiv + T̂i jv for the routing arcs (1, 1, 2, 1) and (2, 1, 3, 2). sidering
We can see that the solution presented is feasible for
ΞB ξ: ξimjnv T̄i jv + T̂i jv δ imjnv ,

the deterministic problem with the nominal traveling
times but is infeasible for the robust problem since the 0 ≤ δimjnv ≤ 1, ∀ (i, m, j, n) ∈ S X , v ∈ V .

earliest arrival time to node (3, 2) is 10 leading to a
stockout since B̄32 9. The worst case occurs when all of the traveling times
take their maximum value, that is, when we consider
the deterministic traveling times Timjnv T̄i jv + T̂i jv . Any
4. Improvement Strategies for the feasible solution to this deterministic problem is feasi-
Decomposition Algorithm ble for the robust problem, and therefore, its objective
To improve the performance of the decomposition ap- function value provides an upper bound, which can be
proach, we test two types of improvements. The first embedded in the decomposition procedure as a cut-
aims to reduce the running time of each master prob- off value to prune the search tree when solving each
lem, while the second aims to reduce the number of optimization problem in step 2 of Algorithm 1. The
iterations of the decomposition method. approach followed in this paper is to solve the box-
constrained problem until a desired optimality gap is
4.1. Solving the Master Problem attained.
To improve the MP solution time we propose two
improvements: the inclusion of valid inequalities and 4.2. Minimizing the Number of Iterations of the
the inclusion of a cutoff value. In both cases, these Decomposition Approach
improvements can lead to a significant improvement The running time of each instance depends greatly on
in the solution times for some difficult instances, the number of MPs solved. Here we propose three
while having a negligible influence when solving easy improvement strategies to reduce the number of iter-
instances. ations of the decomposition approach. The first one
is aggregation of scenarios. The second improvement
Valid Inequalities strategy facilitates a warmer start considering an ini-
Inequalities from knapsack relaxations have previously tial scenario that leads to an initial solution that
been used for MIRP; see, for instance, Agra et al. (2013); incorporates some degree of immunity against delays.
and Agra, Christiansen, and Delgado (2013). Here we Thus the corresponding solution is likely to provide a
better lower bound to the value of the optimal solu- This means that the following constraints must be sa-
tion than the one obtained with the deterministic case tisfied:
where no delays are considered. Finally, we adjust our t im ≤ Bim (q), ∀ (i, m) ∈ S A . (51)
solution approach to choose values on the first-stage
variables that maximize a given criterion of robustness. Let
dim Bim (q) − t im , ∀ (i, m) ∈ S A (52)
Scenarios Aggregation
denote the slack in the upper time window constraint
An improvement introduced in Agra et al. (2013) is to
that measures the delay that can occur on the time visit
add, in each iteration, a scenario that results from the
(i, m) without having a violation. Intuitively, a feasi-
aggregation of several scenarios. Instead of adding a
ble solution with large values for slack variables dim
delay on arc (i, m, j, n) for a given ship, one can either
will be more immune to delays than a solution with
add a delay to all arcs (k, l, j, n) that enter node ( j, n)
small slack values. To choose a first-stage solution, two
or add a delay to all arcs (i, m, k, l) that leave node
options are considered. The first option (denoted as
(i, m). This follows from the fact that at most one such
Option 1) is to maximize the sum of the slack vari-
arc can be selected in each feasible solution. Hence,
ables z (i, m)∈SA dim . The second option (denoted as
P
many scenarios can be aggregated into one augmented
scenario. Option 2) is to maximize the smallest slack dim value,
that is, to maximize z d, where
¯
Warm Start
A warmer start can be obtained by adding delays, that d ≤ dim , ∀ (i, m) ∈ S A . (53)
¯
is, consider Timjnv T̄i jv + T̂i jv , for some arcs. These
delays need to be carefully added since one needs to For completeness, Algorithm 2 details the steps that
guarantee that no more than Γ delays are added to any replace steps 2 and 4 (solve the restricted R-IR(Ξ∗ ) prob-
feasible solution. An easy way to ensure that no more lem) in Algorithm 1.
than Γ delays are added is either to add delays to all Algorithm 2 (A two-stage approach for the master problem
leaving arcs from port visit (i, m), or add delays to all R-IR(Ξ∗ ))
those arcs entering port visit (j, n), for Γ selected port 1: Solve the restricted R-IR(Ξ∗ ) problem
visits. Preferably, the selected visits should be among 2: Fix the variables x imjnv , z imv , w imv , yim to
those that are made in all feasible solutions. Therefore, their solution values
a natural selection starts by choosing the origins of each 3: Add nonnegative variables dim (and d for
ship and then consider the first visit to each port, for Option 2), and constraints (51) and¯ (52)
up to Γ ships.
(and constraints (53) for Option 2)
Choice of First-Stage Solution Maximizing a 4: Solve the resulting linear problem by maximizing
Robustness Criterion function z.
The first-stage variables include both the binary vari- Example 2. In Example 1 we can see that for the same
ables x̄imjnv , z̄ imv , w̄ imv , and ȳim , and the continuous routing decisions there are alternative values of load
variables q̄ imv . In general, for each vector of binary vari- and unload quantities. For instance, by decreasing q 111
ables, one can find alternative values for the continu- to 32 and increasing q 122 to 50, the earliest time for visit
ous variables q imv leading to the same objective func- (1, 1) is now set to 2, (α(1, 1), 0) 2, which leads to a
tion value. Hence, given a binary vector representing situation where no shortage occurs at node (3, 2).
the binary decisions, it makes sense to choose among
all alternative continuous solutions a solution that Remark 3. One can see that the first-stage solution can-
maximizes a given robustness criterion. This results in not be complemented with a second-stage solution if
a pure linear problem when the binary variables are there is a delay in the route to node (i, m) greater
fixed as we explain next. than dim . A lower bound for the worst-case delay can
The ASP checks whether the time windows are con- be computed as the sum of the min{k, Γ} highest val-
sistent by assigning to t im the earliest time of visit ues of T̂i jv in the route to node (i, m), where k is the
value and comparing it with an upper bound resulting number of arcs in that route.
from the inventory level. Here, since the load/unload
quantities q imv are variables, the upper time-windows
5. Iterated Local Search
limit depends on these variables as follows:
As the robust MIRP considered is hard to solve, we
m−1 present an iterated local search heuristic based on local

XX
q + S̄i − Si R i , Ji 1,

 0
branching (Fischetti and Lodi 2003). The idea is to con-

 v∈V n1 inv



Bim (q) sider as the starting solution the most conservative one,
m−1

 XX the solution obtained for the box uncertainty set. This
+ 0
Ri , Ji −1.


 q inv S i − S i

 v∈V n1 ¯ choice has three advantages: (i) it guarantees that a
feasible solution is found; (ii) the starting solution is 4: while There is a scenario ξ ∗ in ext(ΞΓ )\Ξ∗ , leading
obtained solving a deterministic problem and can also to a violation of a constraint (46)–(49) do
be obtained heuristically; (iii) for many instances the 5: Add ξ ∗ to Ξ∗ (and add the corresponding
value of this solution is not far from the value of the variables and rows to the model)
optimal robust solution. Therefore, any improvement 6: Solve the new restricted R-IR(Ξ∗ ) problem
obtained will generate good upper bounds for those using Algorithm 3
cases. 7: Perform steps 2–4 of Algorithm 2
For the local search, following the local branching 8: end while
idea of Fischetti and Lodi (2003), we define the neigh- 9: Save the current best solution as incumbent
borhood of a solution as the set of solutions that can 10: Solve the restricted R-IR(Ξ∗ ) to optimality using
differ in at most Λ of the w imv variables from the cur- the incumbent solution
rent solution. The local search can be done by adding 11: Perform steps 2–4 of Algorithm 2
the following inequality to the model: 12: while There is a scenario ξ ∗ in ext(ΞΓ )\Ξ∗ , leading
X X to a violation of a constraint (46)–(49) do
w imv + (1 − w imv ) ≤ Λ. 13: Add ξ ∗ to Ξ∗ (and add the corresponding
(i, m)∈S vA , v∈V | w̄ imv 0 (i, m)∈S vA , v∈V | w̄ imv 1 variables and rows to the model)
(54) 14: Solve the new restricted R-IR(Ξ∗ ) using the
The iterated local search is given in Algorithm 3. incumbent solution
Algorithm 3 (An MIP iterated local search-based approach 15: end while.
for a given set of scenarios Ξ∗ ) Note that the solution obtained through steps 1 to 8
1: Solve the deterministic problem corresponding to is feasible for the model R-IR(Ω). Hence, the enhanced
the box uncertainty set for α seconds algorithm discussed in Section 4 corresponds to steps
2: repeat 9–15 of Algorithm 4 where the warm start is replaced
3: Add constraint (54) to the model R − IR(Ξ∗ ) by the best robust solution obtained (resulting from
4: Solve the model for β seconds using the warm steps 1–8).
start enhancement
5: Update the solution w̄
6: until No improvement in the objective function
6. Computational Tests
This section reports the computational experiments
is observed.
carried out to test the solution approaches for a set
The decomposition approach followed in Algo- of instances of an MIRP with possible delays when
rithm 1 decomposes the problem into an MP and an traveling between ports. All tests were run on a com-
ASP. While the ASP can be solved efficiently, the MP puter with an Intel Core i5-2410M processor, having
is NP-hard. Solving the MP to optimality in each iter- a 2.3 GHz CPU and 8 GB RAM, using the optimiza-
ation of the decomposition method may be too time tion software Xpress Optimizer Version 21.01.00 with
consuming. Moreover, to prove optimality, only the MP Xpress Mosel Version 3.2.0.
considered at the last iteration must be solved to opti-
mality. To use this fact we propose to use Algorithm 3 6.1. Instances
to solve the MP in each iteration of the decomposition The instances are based on those presented in Agra
algorithm. Only when no new scenario is found do we et al. (2013). There are two main differences: one is the
follow Algorithm 1 to prove optimality. The drawback computation of the traveling times, which we discuss
is that we do not know in advance which is the last in detail below, and the other is the production and
iteration and whether a new better solution for the MP consumption that we assume here to be constant. The
with the same set of scenarios will be violated by a number of ports and ships of each instance is given in
new scenario. In that case the iterative process must be the second column of Table 1. Operating and waiting
resumed. The new enhanced column-and-row genera- costs are time invariant. The size of the deterministic
tion approach is given in Algorithm 4. model is given in the following three columns. A time
horizon of 30 days is considered. The set of 21 instances
Algorithm 4 (The enhanced column-and-row generation includes both easy instances and difficult instances for
approach) which the deterministic problem is already difficult to
1: Initialize the subset Ξ∗ ⊆ ext(ΞΓ ) ← {ξ ∗ } where solve.
ξ ∗ is a scenario with Γ delays (see Warm start The three instances in each group differ from each
in Section 4) other in the initial inventory levels. The nominal trav-
2: Solve the restricted R-IR(Ξ∗ ) problem using eling times T̄i jv are taken as the traveling times of the
Algorithm 3 instances given in Agra et al. (2013). To define the max-
3: Perform steps 2–4 of Algorithm 2 imum delays allowed, T̂i jv , we chose a constant delay α
Table 1. Summary Statistics for the 21 Instances and Option 1 of minimizing the number of iterations
improvement is used in Algorithm 2 (columns Improve
Deterministic model Opt1); and (3) the case where all of the improvements
Inst. (|N |, |V |) No. of rows No. of col. No. of int. var. are included and Option 2 is used (columns Improve
Opt2). When Γ 0, the improved versions are not run,
A1, A2, A3 (4, 1) 765 545 273
B1, B2, B3 (3, 2) 767 590 302
and “—” is used in the corresponding columns. The
C1, C2, C3 (4, 2) 1,214 1,042 530 case Γ 0 corresponds to the deterministic case, and
D1, D2, D3 (5, 2) 1,757 1,622 822 the corresponding cost is therefore a lower bound for
E1, E2, E3 (5, 2) 1,757 1,622 822 the instances with other values of Γ.
F1, F2, F3 (4, 3) 1,663 1,539 787
Note that the instance G2 with Γ 4 could not be
G1, G2, G3 (6, 5) 4,991 5,717 2,909
solved to optimality without improvements, within
one week.
for each instance. This delay was found by increasing Similar results were obtained for the remaining in-
the value of ∆ using steps of 0.1 until the determinis- stances; however, to simplify the text, those results
tic problem with Ti jv (ξ) T̄i jv + ∆ for all (i, j) becomes were aggregated and are presented in Figures 3 and 4.
infeasible. The last ∆ leading to a feasible instance was In these two figures we display the aggregated time
selected. This deterministic case corresponds to the box and the total number of iterations required by each
uncertainty set. strategy of the decomposition algorithm to solve all of
the instances in terms of the different levels of protec-
6.2. Summary of the Computational Results tion Γ, respectively.
In this section we report the computational results car- The results show a clear benefit of using the improve-
ried out to test and compare the performance of both ments in the decomposition algorithm since, when the
the iterated local search (ILS) heuristic and the decom- improvements are employed, the running times can be
position algorithm. much lower than in the case where no improvements
6.2.1. Decomposition Algorithm. We start by evaluat- are used. Instances G2 and G3 are the ones in which
ing the benefits of the use of the improvement strate- more significant improvements can be observed. Note
gies in the decomposition algorithm. Table 2 displays that the benefit of the use of the improvement strategies
the results for the instances G1, G2, and G3, which tends to increase when the level of protection increases.
are the most difficult instances. The first column gives The number of iterations is lower when the improve-
the instance, the second column gives the value ∆, and ments are included and, in several cases, mainly when
the third column indicates the maximum number of Γ is small, no iterations are required. When compar-
delays, Γ, allowed. The values of Γ range from 0 (corre- ing both improvement options to choose a robust solu-
sponding to the deterministic model) to the number of tion among alternative optimal solutions to the mas-
delays leading to the same solution as the one obtained ter problem, we can see that the number of iterations
for the box uncertainty set. The optimal objective func- is lower using Option 2 than Option 1. In terms of
tion value is given in column Cost. The last three pairs time, there is no clear evidence indicating which is the
of columns give the running time in seconds (columns best improvement option. However, since for the most
Seconds), and the number of iterations (columns Iter) difficult instance, instance G2, optimal solutions are
of the decomposition algorithm for (1) the case where obtained faster by using the improvement Option 2, in
no improvements are used (columns No improvements); what follows we will only consider the decomposition
(2) the case where all of the improvements are included algorithm enhanced by this option.
Table 2. Computational Results (Time and Number of Iterations) for Instances G1, G2, and G3
No improvements Improve Opt1 Improve Opt2
Instance ∆ Γ Cost Seconds Iter Seconds Iter Seconds Iter
G1 0.2 0 645.8 975 0 — — — —

0 560.0 536 0 — — — —
1 560.0 1,791 2 1,325 2 1,375 2
G2 1.6 2 589.9 13,303 6 2,818 6 1,960 3
3 589.9 14,881 9 4,728 5 3,010 5
4 631.4 >604,800 >22 123,981 13 63,734 13
0 550.7 712 0 — — — —
G3 0.4 1 576.2 20,999 7 1,907 1 3,650 4
Figure 3. Aggregated Computational Time for All Instances, Figure 5. Comparison Between the Aggregated
Except for Instances G1, G2, and G3, for the Different Values Computational Time Required by the Decomposition
of Γ Algorithm Improved by Option 2 and by the ILS Heuristic to
20,000
Obtain the Robust Solution for All Instances, Except for
18,000 No improvements Instances G1, G2, and G3, for the Different Values of Γ
16,000 Improve Opt1 3,500
Time (seconds)
14,000 Improve Opt2 Decomposition algorithm

12,000 3,000
ILS
10,000
Time (seconds)
2,500
8,000
6,000 2,000
4,000 1,500
2,000
0 1,000
1 2 3 4 500
Γ 0
1 2 3 4
Γ
6.2.2. Comparison Between the Decomposition Algo-
rithm and the ILS Heuristic. For the ILS heuristic, nine
different strategies were tested: the time limit in each second columns, respectively. Columns “Max,” “T̄,”
subproblem β was set to 50, 100, and 200 seconds, and and “Min” display, respectively, the maximum, the
the number of variables w imv that are allowed to flip average, and the minimum computational time used
their value from the value taken in the current solution, to obtain the robust solution over the nine different
parameter Λ, is set to 2, 4, and 5. parameter settings. Column “DT” displays the compu-
For the 21 instances, the solutions obtained using the tational time required by the decomposition algorithm
ILS heuristic and the decomposition algorithm have improved by Option 2 to get the robust solution and in
the same cost. Hence, the comparison of the solution column “dif” the difference between values of columns
methods is done only in terms of the running time. For “DT” and “T̄” is computed. Column “TO ” reports
all of the instances, except for instances G1, G2, and G3, the additional computational time used to prove the
the results obtained by the nine strategies tested are optimality of the solution obtained by the ILS heuris-
very similar. This can be explained by the fact that most tic. Column “Ī” displays the average number of itera-
of the subproblems are solved to optimality within the tions required to find the robust solution and column
time limit. Hence, the results for these instances are “DI” reports the number of iterations required by the
aggregated and summarized in Figures 5 and 6. In Fig- decomposition algorithm improved by Option 2 to get
ure 5, the aggregated computational time required by the robust solution.
the decomposition algorithm improved with Option 2 Results not reported here show that for the easiest
to obtain the robust solution for all of the instances, in instances there is no advantage in using the ILS heuris-
terms of the values of Γ, is presented. In Figure 6, sim- tic since it is necessary to spend some time in solving
ilar results for the number of iterations are presented. the box-constrained problem. When the complexity of
For the instances G1, G2, and G3, the computa- the instances increase, the ILS heuristic becomes more
tional times required to find the robust solution vary efficient than the decomposition algorithm. By using
a lot from strategy to strategy. The complete results
for these three instances are displayed in Table 3. The Figure 6. Comparison Between the Aggregated Number of
instance and the Γ value are indicated in the first and Iterations Required by the Decomposition Algorithm
Improved by Option 2 and by the ILS Heuristic to Obtain the
Robust Solution for All Instances, Except for Instances G1,
Figure 4. Aggregated Number of Iterations for All
G2, and G3, for the Different Values of Γ
Instances, Except for Instances G1, G2, and G3, for the
Different Values of Γ 40
Decomposition algorithm
35
80 ILS
Number of iterations
70 No improvements 30
Number of iterations
60 Improve Opt1 25
50 Improve Opt2
20
40
15
30
10
20
10 5
0 0
1 2 3 4 1 2 3 4
Γ Γ
Table 3. Comparison Between the ILS Heuristic and the Table 6. Slack Time Values for Instance E3
Decomposition Algorithm for Instances G1, G2, and G3
Γ 0 1 2 3
Inst. Γ Max T̄ Min DT dif TO Ī DI
Slack time 0 2 4.1 6.95
G1 0 273 272 271 975 703 88 — — ∆ 0 0.5 0.8 1.8
0 305 249 176 536 287 87 — —
1 630 433 232 1,325 892 31 2 2
G2 2 931 938 689 1,960 1,022 471 5 3 that allow us to identify how sensitive an instance is to
3 4,153 1,831 1,498 3,010 1,179 602 12 5 the uncertainty set parameters.
4 16,948 7,654 6,666 63,734 56,080 6,001 16 13 For instance E3, Table 6 gives, for each value of Γ,
G3 0 201 195 174 712 517 58 — — the maximum slack time (row slack time) correspond-
1 829 604 466 1,907 1,303 527 2 4
ing to the maximum delay that can occur in a sin-
gle arc without leading to an inventory bound viola-
the ILS heuristic, in some iterations, not all subprob- tion. As expected, the slack increases as Γ increases.
lems are solved to optimality, leading to worse solu- Row ∆ gives the maximum delay that can be added
tions and consequently a greater number of iterations, to all traveling times while keeping the solution feasi-
as shown in Figure 6. ble. This value is lower than the value of the box con-
Instances G1, G2, and G3 are the most difficult in- straint (which is 2) since we have the routing decisions
stances, thus more computational time is required to fixed. As expected, this value also increases with the
obtain a solution in each iteration, and this time varies value of Γ.
a lot from strategy to strategy. However, these are the Figure 7 depicts the routing of an optimal solution
instances that most reflect the power of the ILS heuris- with Γ 0. The triple [t im , q imv , s im ] is shown next to
tic since, independently of the strategy used, the cor- each node. We can see that s41 0, meaning that when
responding computational time is always lower than the ship arrives at node 4 for the first visit the inven-
the one required by the decomposition algorithm. The tory level is zero, and therefore any delay in the critical
obtained results reveal a good performance of the ILS path to this node will imply a shortfall. Figures 8 and 9
heuristic, since the optimal solution for each instance give two alternative optimal solutions for the case Γ 1.
was always found with a small computational time. The routing is the same, but the quantities loaded and
unloaded and the time of visits are different. The solu-
6.3. Results for Instance E3 tion in Figure 8 maximizes the minimum slack time,
To better understand the influence of the levels of pro- which is determined by the inventory level at node
tection in a robust solution, we deeply analyze the com- (5, 2). The solution in Figure 8 is the one that maximizes
putational results obtained for instance E3. Detailed ∆ (the value considered when the traveling time in all
information for instance E3 is given in Table 4. arcs is increased by ∆), so the inventory level at nodes
In Table 5, for each level of protection Γ, the cost (5, 2) and (4, 1) is zero. We see from the figures that the
of the robust solution is displayed. As an example of ship traveling the dashed line route is cheaper to use
robustness, we can see that for this instance the optimal than the other ship (traveling the solid line route).
solution changes for all values of Γ. That is not the usual Finally, Figure 10 depicts the minimum, the aver-
behavior for the remaining instances where the same age, and the maximum percentage of increase in cost,
solution may be optimal for more than one value of Γ. among all instances, as a function of Γ. Such informa-
In fact we could not determine instance characteristics tion gives the price to obtain robust solutions. As we
can see, there is a wide range in the cost increase varia-
tion according to the instance, from a minimum of zero
Table 4. Detailed Information for Instance E3 to a maximum that almost doubles the cost when Γ 4.
i 1 2 3 4 5
Figure 7. Optimal Solution of Instance E3 with Γ 0 and
[Si , S̄i ] [0, 300] [0, 350] [0, 250] [0, 145] [0, 300]
Cost 226.5
R¯i 11.63 9.60 7.17 4.27 11.03
Ji 1 1 −1 −1 −1
1, 1 5, 1 4, 1
[0, 100, 150] [5.9, 100, 113.8] [23.4, 48, 0]
Table 5. Cost of the Robust Solution for Instance E3 in

2, 1 3, 1
Terms of Value Γ
[0, 129.2, 150] [6.7, 129.2, 51.8]
Γ 0 1 2 3
1, 2 5, 2
Cost 226.5 230.4 248.6 267.2 [14.4, 98.9, 218] [20.8, 50.9, 50.9]
Figure 8. Optimal Solution of Instance E3 with Γ 1 That Table 7. Computational Results for the Two Deterministic
Maximizes the Minimum Slack Time and Cost 230.4 Problems: Γ 0 and Box-Constrained
1, 1 5, 1 4, 1 Deterministic (Γ 0) Box-constrained
[0, 100, 150] [5.9, 66.9, 113.8] [8.7, 28, 63] Inst. Seconds Cost Seconds Cost
2, 1 3, 1 C4 9 529 23 595
[0, 115, 150] [6.5, 115, 53.2]
C5 22 546 215 675
D4 223 379 163 414
1, 2 5, 2 D5 16 289 40 375
E4 39 272 870 391
[14.1, 98.9, 213.6] [20.4, 84.1, 22.1]
E5 86 285 278 362
F4 25 387 17 486
F5 6 419 47 507
Figure 9. Routing of an Optimal Solution of Instance E3 G4 2,104 672 18,000∗ 766 (20%)
with Γ 1 That Maximizes ∆ G5 4,194 648 18,000∗ 775 (14%)
1, 1 5, 1 4, 1 Note. An asterisk means that the run time limit was reached.
[0.5, 100, 156] [6.9, 66.9, 102.8] [23.4, 28, 0]
the instances A4, A5, B4, and B5 are easy to solve, we
2, 1 3, 1 omit these four cases.
[0.5, 115, 154.8] [7.5, 115, 46] In Table 7, we display the results for both the deter-
ministic (Γ 0) and the box-constrained problems.
1, 2 5, 2 The instance is indicated in the first column. The last
[15.6, 98.9, 231] [22.4, 84.1, 0] two pairs of columns give the running time in sec-
onds (columns Seconds), and the cost of the solution
(columns Cost).
Figure 10. Percentage Increase in Cost as a Function of Γ
For the box-constrained problem, instances G4 and
90 G5 were not solved to optimality within a time limit of
80
Max. 18,000 seconds (five hours) and the final gap is indi-
Avg.
Min.
cated in parentheses. This fact reinforces that these
70 instances are difficult since both problems are deter-
ministic problems.
60
% increase in cost
In Table 8 we present the results obtained for the

50 robust problems with Γ 1 and Γ 2. The instance
is indicated in the first column. The last two pairs of
40
columns give, for both the decomposition algorithm
30 improved with Option 2 and the ILS heuristic, the run-
ning time in seconds (columns TimeDA and TimeILS ) and
20
the cost of the obtained solutions (columns CostDA and
10 CostILS ).
Instances from C4 to F5 were solved to optimality
0
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 using both approaches, thus the cost of the obtained
Γ solutions is the same. In columns CostILS , the value in
parentheses is the cost of the solution obtained at the
end of step 8 in Algorithm 4. This value is presented
On average, we can observe a steady increase in the only for those cases in which it differs from the opti-
cost for protection against inventory level deviations mal one. Recall that for all of the instances with a time
with the protection level Γ. horizon of 30 days, the solution obtained at the end of
step 8 was always the optimal robust solution.
6.4. Results for Large Size Instances Analyzing the computational time required by each
Although the original set of instances includes both approach, we can see that, in general, the robust solu-
easy and difficult instances, to evaluate the performance tions are obtained faster when the ILS heuristic is used,
of our approaches for larger instances, a set of 14 as happened in the case of the instances with a time
instances with a time horizon of 60 days, was con- horizon of 30 days.
sidered. These instances were derived from those pre- Instances G4 and G5 were not solved to optimality.
sented in Table 1 by extending the time horizon and Note that even the deterministic box-constrained prob-
keeping the maximum number of visits allowed to each lem cannot be solved to optimality within a reason-
port. Since we aim to test the most difficult cases, and able amount of computation time, as shown in Table 7.
Table 8. Computational Results for Both the Decomposition Algorithm Enhanced with
Option 2 and the ILS Heuristic, for Γ 1 and Γ 2
Γ1 Γ2
Inst. TimeDA CostDA TimeILS CostILS TimeDA CostDA TimeILS CostILS
C4 114 564 62 564 507 595 233 595

C5 838 602 483 602 838 602 486 602
D4 617 379 159 379 1,522 414 1,159 414
D5 137 327 97 327 141 327 147 327 (375)
E4 286 289 108 289 7,702 369 6,512 369
E5 412 289 112 289 (305) 3,180 331 1,674 331 (362)
F4 55 387 96 387 (486) 624 486 865 486
F5 1,010 490 250 490 882 507 887 507
G4 2,404 771∗ 2,571 740∗ (787) 3,604 761∗ 10,800 761∗ (787)
G5 4,800 785∗ 4,032 736∗ (811) 10,800 771∗ 12,244 688∗ (820)
Note. An asterisk means that the obtained solution may not be the optimal solution.
Hence, for both approaches, a time limit of 600 seconds insight on the robustness of solutions is obtained by
was imposed to solve the restricted master problem in comparing robust solutions against the deterministic
each iteration. This means that the obtained solutions solution, and shows that robustness is closely related
can be suboptimal. For these two instances, better solu- to the slack time available for each time window that
tions were obtained using the ILS heuristic. However, can be established from the inventory levels for each
the computational time required to obtain those solu- port visit.
tions is greater than the one required by the decompo-
sition algorithm, since more iterations were executed. Acknowledgments
The authors thank the two anonymous reviewers for their
valuable comments and helpful suggestions.
7. Conclusions
We consider a maritime inventory routing problem References
where the travel times are uncertain. To handle prac- Adulyasak Y, Jaillet P (2016) Models and algorithms for stochastic
tical cases where the decision maker wishes to avoid and robust vehicle routing with deadlines. Transportation Sci.
inventory shortfalls at the consumers and exceed 50(2):608–626.
inventory capacities at producers, a robust model with Adulyasak Y, Cordeau JF, Jans R (2015) Benders decomposition
for production routing under demand uncertainty. Oper. Res.
recourse is proposed. The routing decisions and the
63(4):851–867.
quantities to load and unload are assumed to be fixed Aghezzaf EH (2008) Robust distribution planning for supplier-
and the time of visit to ports and the inventory lev- managed inventory agreements when demand rates and travel
els are adjustable. A two-stage decomposition proce- times are stationary. J. Oper. Res. Soc. 59(8):1055–1065.
dure that considers a master problem restricted to a Agra A, Christiansen M, Delgado A (2013) Mixed integer formula-
tions for a short sea fuel oil distribution problem. Transportation
small subset of scenarios and a subproblem that checks Sci. 47(1):108–124.
whether there are violated scenarios is presented. Sev- Agra A, Andersson H, Christiansen M, Wolsey L (2013) A mar-
eral improvement strategies are introduced. In partic- itime inventory routing problem: Discrete time formulations
ular, since the first stage has, in general, multiple alter- and valid inequalities. Networks 64(4):297–314.
Agra A, Christiansen M, Delgado A, Hvattum LM (2015) A mar-
native optimal solutions corresponding to a single set itime inventory routing problem with stochastic sailing and port
of routing decisions, we discuss approaches to choose times. Comput. Oper. Res. 61:18–30.
the most robust alternative solution in relation to a Agra A, Christiansen M, Hvattum LM, Rodrigues F (2016) A MIP
given criteria. An iterated local search heuristic based based local search heuristic for a stochastic maritime inventory
routing problem. Paias A, Ruthmair M, Voss S, eds. Computa-
on local branching ideas is introduced that allows us tional Logistics, Lecture Notes Comput. Sci., Vol. 9855 (Springer
to obtain good-quality robust solutions for all of the International Publishing, Cham, Switzerland), 18–34.
tested instances. Agra A, Santos MC, Nace D, Poss M (2016) A dynamic programming
A computational study based on a set of benchmark approach for a class of robust optimization problems. SIAM J.
Optim. 26(3):1799–1823.
instances shows the effectiveness of the decomposition Agra A, Christiansen M, Figueiredo R, Hvattum LM, Poss M,
procedure with the improvement strategies in solving Requejo C (2013) The robust vehicle routing problem with time
the robust MIRP. This study also shows that, in general, windows. Comput. Oper. Res. 40(3):856–866.
protecting the solutions against possible delays makes Andersson H, Hoff A, Christiansen M, Hasle G, Løkketangen A
(2010) Industrial aspects and literature survey: Combined inven-
the instances harder to solve, since the running times tory management and routing. Comput. Oper. Res. 37:1515–1536.
and the number of iterations tend to increase when the Atamtürk A, Zhang M (2007) Two-stage robust network flow and
number of links that can suffer a delay increases. Some design under demand uncertainty. Oper. Res. 55(4):662–673.
Ben-Tal A, El Ghaoui L, Nemirovski AS (2009) Robust Optimization Halvorsen-Weare EE, Fagerholt K, Rönnqvist M (2013) Vessel routing
(Princeton University Press, Princeton, NJ). and scheduling under uncertainty in the liquefied natural gas
Ben-Tal A, Goryashko A, Guslitzer E, Nemirovski A (2004) Ad- business. Comput. Indust. Engrg. 64:290–301.
justable robust solutions of uncertain linear programs. Math. Li M, Wang Z, Chan F (2016) A robust inventory routing policy under
Programming 99(2):351–376. inventory inaccuracy and replenishment lead-time. Transporta-
Bertsimas D, Sim M (2003) Robust discrete optimization and network tion Res. Part E: Logist. Transportation Rev. 91:290–305.
flows. Math. Programming 98:49–71. Liebchen C, Lübbecke M, Möhring R, Stiller S (2009) The concept
Bertsimas D, Sim M (2004) The price of robustness. Oper. Res. of recoverable robustness, linear programming recovery, and
52(1):35–53. railway applications. Ahuja RK, Möhring R, Zaroliagis C, eds.
Bertsimas D, Thiele A (2006) A robust optimization approach to Robust and Online Large-Scale Optimization, Lecture Notes Com-
inventory theory. Oper. Res. 54(1):150–168. put. Sci., Vol. 5868 (Springer, Berlin Heidelberg), 1–27.
Bertsimas D, Brown DB, Caramanis C (2011) Theory and applications Maggioni F, Potra F, Bertocchi M (2017) A scenario-based framework
of robust optimization. SIAM Rev. 53:464–501. for supply planning under uncertainty: Stochastic programming
versus robust optimization approaches. Comput. Management
Bertsimas D, Gupta S, Tay J (2016) Scalable robust and adaptive
Sci. 14(1):5–44.
inventory routing. Optim.-Online.
Papageorgiou DJ, Nemhauser GL, Sokol J, Cheo M-S, Keha AB (2014)
Bertsimas D, Iancu DA, Parrilo PA (2011) A hierarchy of near-optimal
MIRPLib—A library of maritime inventory routing problem
policies for multistage adaptive optimization. IEEE Trans. Auto-
instances: Survey, core model, and benchmark results. Eur. J.
matic Control 56(12):2809–2824. Oper. Res. 234(2):350–366.
Bienstock D, Özbay N (2008) Computing robust basestock levels. Rakke J, Andersson H, Christiansen M, Desaulniers G (2015) A new
Discrete Optim. 5:389–414. formulation based on customer delivery patterns for a maritime
Billionnet A, Costa M-C, Poirion P-L (2014) 2-stage robust {MILP} inventory routing problem. Transportation Sci. 42(2):384–401.
with continuous recourse variables. Discrete Appl. Math. 170: Roldán RF, Basagoiti R, Coelho LC (2016) Robustness of inventory
21–32. replenishment and customer selection policies for the dynamic
Cheng L, Duran MA (2004) Logistics for world-wide crude oil trans- and stochastic inventory-routing problem. Comput. Oper. Res.
portation using discrete event simulation and optimal control. 74:14–20.
Comput. Chemical Engrg. 28:897–911. Roldán RF, Basagoiti R, Coelho LC (2017) A survey on the inventory-
Christiansen M, Fagerholt K (2002) Robust ship scheduling with mul- routing problem with stochastic lead times and demands.
tiple time windows. Naval Res. Logist. 49(6):611–625. J. Appl. Logic 24:15–24.
Christiansen M, Fagerholt K (2009) Maritime inventory routing prob- Russell RA (2017) Mathematical programming heuristics for the pro-
lems. Floudas CA, Pardalos PM, eds. Encyclopedia of Optimization duction routing problem. Internat. J. Production Econom. 193:
(Springer, Boston), 1947–1955. 40–49.
Christiansen M, Fagerholt K (2014) Ship routing and scheduling Sherali HD, Al-Yakoob SM (2006a) Determining an optimal fleet mix
in industrial and tramp shipping. Toth P, Vigo D, eds. Vehicle and schedules: Part I—Single source and destination. Karlof JK,
Routing: Problems, Methods, and Applications, MOS-SIAM Series ed. Integer Programming: Theory and Practice (CRC Press, Boca
Optim., 2nd ed., Vol. 18 (SIAM, Philadelphia), 381–408. Raton, FL), 137–166.
Christiansen M, Nygreen B (2005) Robust inventory ship routing by Sherali HD, Al-Yakoob SM (2006b) Determining an optimal fleet mix
column generation. Desrosiers J, Desaulniers G, Solomon M, and schedules: Part II—Multiple sources and destinations, and
eds. Column Generation (Springer, New York), 197–224. the option of leasing transshipment depots. Karlof JK, ed. Integer
Christiansen M, Fagerholt K, Nygreen B, Ronen D (2013) Ship rout- Programming: Theory and Practice (CRC Press, Boca-Raton, FL),
ing and scheduling in the new millennium. Eur. J. Oper. Res. 167–194.
228(3):467–483. Solyali O, Süral H (2017) A multi-phase heuristic for the production
Coelho L, Cordeaux JF, Laporte G (2014) Thirty years of inventory routing problem. Comput. Oper. Res. 87:114–124.
Solyali O, Cordeau JF, Laporte G (2012) Robust inventory routing
routing. Transportation Sci. 48(1):1–19.
under demand uncertainty. Transportation Sci. 46(3):327–340.
Delage E, Iancu DA (2015) Robust multistage decision making.
Thiele A, Terry T, Epelman M (2010) Robust linear optimization with
Aleman DM, Thiele AC, eds. Tutorials in Operations Research
recourse. Optim.-Online.
(INFORMS, Catonsville, MD), 20–46. Zeng B, Zhao L (2013) Solving two-stage robust optimization prob-
Desaulniers G, Rakke JG, Coelho LC (2016) A branch-price-and-cut lems using a column-and-constraint generation method. Oper.
algorithm for the inventory-routing problem. Transportation Sci. Res. Lett. 41(5):457–461.
50(3):1060–1076. Zhang C (2015) Robust optimization with applications in maritime
Fischetti M, Lodi A (2003) Local branching. Math. Programming 98: inventory routing. Unpublished doctoral thesis, School of Indus-
23–47. trial and Systems Engineering, Georgia Institute of Technology,
Gabrel V, Murat C, Thiele A (2014) Recent advances in robust opti- Atlanta.
mization: An overview. Eur. J. Oper. Res. 235(3):471–483. Zhang C, Nemhauser G, Sokol J, Cheon MS, Keha A (2018) Flexible
Gorissen BL, Yanikoğlu I, den Hertog D (2015) A practical guide to solutions to maritime inventory routing problems with delivery
robust optimization. Omega 53:124–137. time windows. Comput. Oper. Res. 89:153–162.
Halvorsen-Weare EE, Fagerholt K (2011) Robust supply vessel plan- Zhang C, Nemhauser G, Sokol J, Cheon MS, Papageorgiou D (2015)
ning. Reiners T, Pahl J, Voss S, eds. Network Optimization, Lecture Robust inventory routing with flexible time window allocation.
Notes Comput. Sci., Vol. 6701 (Springer, Berlin), 559–574. Optim.-Online.

Robust Optimization For A Maritime Inventory Routing Problem2018Transportation Science

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Robust Optimization For A Maritime Inventory Routing Problem2018Transportation Science

Uploaded by

Copyright:

Available Formats

This article was downloaded by: [157.193.8.

33] On: 11 January 2021, At: 02:50

Robust Optimization for a Maritime Inventory Routing

To cite this article:

Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-

Copyright © 2018, INFORMS

Please scroll down for article—it is on subsequent pages

Robust Optimization for a Maritime Inventory Routing Problem

1. Introduction research has formed the basis of several surveys: Papa-

variables: ximjnv is 1 if ship v travels from node (i, m) O

and Q̄ i , respectively. ¯ A im ≤ t im ≤ Bim , ∀ (i, m)∈S A . (19)

denotes the number of times vehicle v visits port i dur-

No improvements Improve Opt1 Improve Opt2

Instance ∆ Γ Cost Seconds Iter Seconds Iter Seconds Iter

G1 0.2 0 645.8 975 0 — — — —

14,000 Improve Opt2 Decomposition algorithm

Table 5. Cost of the Robust Solution for Instance E3 in

In Table 8 we present the results obtained for the

Inst. TimeDA CostDA TimeILS CostILS TimeDA CostDA TimeILS CostILS

C4 114 564 62 564 507 595 233 595

You might also like