European Journal of Operational Research 237 (2014) 8290

Contents lists available at ScienceDirect

European Journal of Operational Research

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

Discrete Optimization

Formulations for a problem of petroleum transportation

Luiz Aizemberg , Hugo Harry Kramer, Artur Alves Pessoa, Eduardo Uchoa
Production Engineering Department, Fluminense Federal University, Rua Passo da Ptria 156, 24210-240 Niteri, RJ, Brazil

a r t i c l e

i n f o

Article history:
Received 11 March 2013
Accepted 19 January 2014
Available online 27 January 2014
Keywords:
Mixed Integer Programming
Transportation problem
Column generation-based heuristic

a b s t r a c t
Oil tankers play a fundamental role in every offshore petroleum supply chain and due to its high price, it
is essential to optimize its use. Since this optimization requires handling detailed operational aspects,
complete optimization models are typically intractable. Thus, a usual approach is to solve a tactical level
model prior to optimize the operational details. In this case, it is desirable that tactical models are as precise as possible to avoid too severe adjustments in the next optimization level. In this paper, we study
tactical models for a crude oil transportation problem by tankers. We did our work on the top of a previous paper found in the literature. The previous model considers inventory capacities and discrete lot
sizes to be transported, aiming to meet given demands over a nite time horizon. We compare several
formulations for this model using 50 instances from the literature and proposing 25 new harder ones.
A column generation-based heuristic is also proposed to nd good feasible solutions with less computational burden than the heuristics of the commercial solver used.
2014 Elsevier B.V. All rights reserved.

1. Introduction
Oil tankers play a fundamental role in every offshore petroleum
supply chain. Its optimization is usually divided in three levels:
strategic, tactical and operational. Strategic decisions deal with
eet sizing, facility location and capacity sizing. Tactical decisions
deal with production and distribution planning, transportation
mode selection, storage allocation and order picking strategies.
Finally, operational decisions deal with shipment and vehicle
dispatching (Ghiani, Laporte, & Musmanno, 2004). When dealing
with a planning level, all decisions made in the previous levels
are assumed to be known. Since each level considers more details
than the previous ones, past decisions usually have to be adjusted.
Hence, it is important to make both the strategic and tactical models as detailed as possible to guarantee low adjustments in the
operational level.
In this paper, we study a tactical optimization model for crude
oil distribution by tankers. The problem consists of scheduling the
shipments through routes linking platforms (offshore production
sites) and terminals (onshore consumer sites). The objective is to
ship the products from the platforms to supply the terminals with
minimum transportation cost for a nite planning horizon. For
each site, the inventory levels must lie between a lower and an
upper bound to avoid the lack or excess of product. Also, for each
site, the demand or the production is given for the whole planning
Corresponding author. Tel.: +55 2126295708.
E-mail addresses: luizaizemberg@gmail.com (L. Aizemberg), hugoharry
@gmail.com (H.H. Kramer), artur@producao.uff.br (A.A. Pessoa), uchoa@prod
ucao.uff.br (E. Uchoa).
0377-2217/$ - see front matter 2014 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ejor.2014.01.036

horizon. The products are shipped only by oil tankers and we

assume that the global capacity of the eet is unlimited. For each
transportation, one has to decide the route, the lot size and the
delivery day. This problem has already been addressed in Rocha
(2010) and Rocha, Grossmann, and Poggi de Arago (2011).
This paper has two aims. The rst aim is to compare mathematical formulations for the transportation problem proposed in Rocha
et al. (2011). Firstly, a basic formulation is given, where the inventory level for each day in each site is represented by a continuous
variable and the shipments are represented by binary variables.
This formulation is similar to many others found in the literature.
We refer to Sawik (2011) for an overview of such papers. We also
test several variations of the formulation proposed in Rocha et al.
(2011), referred to as Knapsack Cascading. Lastly, we devise a new
formulation that outperforms all previous ones when given to a
Mixed Integer Programming (MIP) solver.
The second aim of this paper is to propose a column generationbased heuristic approach. We note that, for difcult instances, the
commercial solver used is not capable of nding good feasible
solutions at the beginning of the branch-and-bound algorithm tree.
Finding a good feasible solution at the very beginning of the tree is
crucial to reduce the search by cutting off nodes of such tree,
increasing the probability of proving the optimality in a reasonable
computational time.
We did our work on the top of a previous paper found in the literature (Rocha et al., 2011). The instances used to test the formulations are those proposed by Rocha et al. Our paper shows a clear
evolution concerning the results as the formulations are improved,
leading to the generation of new and harder instances (based on
the existent ones) to evaluate the potential of such formulations.

83

L. Aizemberg et al. / European Journal of Operational Research 237 (2014) 8290

1.1. Literature review

Several models for optimizing the transportation of petroleum
products are found in literature, considering different characteristics. In Colvin and Maravelias (2011), the authors consider task
selection, scheduling, and resource planning decisions, exploring
risk management methods, and develop a multi-stage stochastic
programming model. To optimize such models, Mixed Integer Programming is the most used technique. In Magato, Arruda, and
Neves (2004) and Boschetto et al., 2010, MIP is applied to schedule
the activities in a real oil pipeline network. The former uses illustrative instances with four products, while the latter uses large instances, where more than 14 oil derivatives and ethanol are
transported. In Moro and Pinto (2004), MIP is used to optimize the
short-term crude oil scheduling. The authors deal with the separation of brine from the oil, interface separation between different
types of oil, and energy consumption in the crude distillation units.
MIP is also used in a real-time approach to dispatch petroleum tank
trucks (Brown & Graves, 1981) and to optimize the long-term oil
derivative scheduling under uncertainty (MirHassani, 2008). Finally, in Banaszewski, Tacla, Pereira, Arruda, and Enembreck (2010)
and Aizemberg et al. (2011), the authors deal with the mediumterm planning of oil derivatives transportation in the Brazilian oil
industry multimodal supply chain network. The former applied an
auction based multiagent algorithm, while the latter applied MIP.
Column generation based heuristics (CGH) are found in literature to deal with many different problems: multi-item scheduling
(Bahl, 1983), vehicle routing (Mourgaya & Vanderbeck, 2007), cutting stock (Furini, Malaguti, Durain, Persiani, & Toth, 2012), sensor
placement (Yavuz, Aras, Kuban, & Ersoy, 2010), generalized assignment (Moccia, Cordeau, Monaco, & Sammarra, 2009) and many
others. In general, such approaches share similar main characteristics. First, the linear programming relaxation is solved by column
generation. Then, a feasible solution is constructed for the original
problem using the obtained columns. For example, in Choi and
Tcha (2007) the authors apply a column generation heuristic for
the heterogeneous eet vehicle routing problem which consists
of three steps: (i) solve the continuous relaxation; (ii) set the variables of the nal restricted master as binaries; and (iii) solve the
resulting MIP problem. In Joncour, Michel, Sadykov, Sverdlov, and
Vanderbeck (2010), the authors review generic classes of column
generation-based heuristics.
1.2. Paper organization
The remainder of the paper is organized as follows. Section 2
presents the problem denitions and the mathematical formulations, including the DantzigWolfe decomposition formulation
used by the proposed heuristic. Section 3 shows the column generation-based heuristic procedure, including dynamic programming
algorithm for the pricing subproblem. Section 4 discusses the obtained results. At last, Section 5 contains the conclusions of this
work.

in classes according to their capacities. Let C be the set of all tanker

classes. It is assumed that only oil tankers with fullled capacities
are used in the transportation. The objective is to minimize the total transportation costs.
The formulations described in the next subsections consider the
following notation:
Indices:





p: Platform, p 2 P;
t: Terminal, t 2 T;
c: Class of tanker, c 2 C;
d: day, d 2 f1; . . . ; Dg.

Subsets:
 Tp # T: Terminals allowed to send tankers to a platform
p 2 P;
 Pt # P: Platforms allowed to receive tankers from a terminal t 2 T;
 Cp # C: Classes of tankers allowed to pick up oil from platform p 2 P.
Input data:









ep;0 : Initial inventory at platform p 2 P;

et;0 : Initial inventory at terminal t 2 T;
Pp;d P 0: Production at platform p 2 P, in day d;
C t;d P 0: Consumption at terminal t 2 T, in day d;
CAPp;d : Maximum inventory capacity at platform p 2 P, in
day d;
CAPt;d : Maximum inventory capacity at terminal t 2 T, in day
d;
V c : Capacity of a tanker of class c 2 C;
F c : Transportation cost per day for a tanker of class c 2 C
(tankers must travel back from the platform to the terminal
after each transportation);
Dp;t : Transportation time between a platform p 2 P and a terminal t 2 T in days.

Other denitions will be added as needed by each formulation.

2.1. Natural formulation
For this model the decision variables are dened as follows.
zc;d
p;t 2 f0; 1g is a binary variable indicating if a tanker of class c 2 C
is assigned to ship the product from platform p 2 P to terminal
t 2 T at day d. The variables ep;d 2 R and et;d 2 R represent the
inventory levels at platform p 2 P and terminal t 2 T at day d, respectively. Thus, the natural formulation (NF ) is given by the following:
D
XX XX
c;d
2F c Dp;t zp;t
;

NF min
subject to

ep;d ep;d1 P p;d 

2. Mathematical formulations

X X

c;d
V c zp;t

8p 2 P; d 2 f1;...; Dg;

t2Tpc2Cp

et;d et;d1  C t;d

This section presents the mathematical formulations considered for the problem in study.
The optimization problem dened in Rocha et al. (2011) is the
following. Let P be the set of platforms and T the set of terminals.
The planning horizon consists of D days. A unique product should
be shipped from platforms to terminals to satisfy given demands.
The inventory levels of platforms and terminals are limited by a
lower and an upper bound, i.e., the lack or surplus of the product
is not allowed at any point of the network. Oil tankers are grouped

p2P t2Tpc2Cp d1

X X

c;dDp;t

V c zp;t

8t 2 T; d 2 f1;...;Dg; 3

p2Ptc2Cp

0 6 ep;d 6 CAPp;d
0 6 et;d 6 CAP t;d
zc;d
p;t 2 f0; 1g

8p 2 P; d 2 f1;...;Dg;
8t 2 T; d 2 f1;.. .;Dg;

8p 2 P; t 2 Tp; c 2 Cp; d 2 f1;. ..;Dg:

5
6

The objective function (1) aims to minimize the total transportation costs. Constraints (2) calculate the inventory balance at each
platform, while constraints (3) calculate the inventory balance at
each terminal. Constraints (4) and (5) assure that inventory levels

84

L. Aizemberg et al. / European Journal of Operational Research 237 (2014) 8290

at each platform and terminal at each day will not be less than
their minimum storage capacity nor greater than their maximum
storage capacity.
Some of the problem assumptions deserve discussion:
 no more than one tanker of each class can be used in each route
and day: this assumption is not realistic, we just followed the
previous paper. But it is easy to drop this assumption, just setting the binary variables as integer;
 unlimited number of tankers in each class: we believe this
assumption is realistic because it is based on a real problem
where the oil company rents the tankers;
 the instances use no realistic transportation and (un)loading
times: we could use realistic values, but this will not change
the formulation performance.
The NF is the starting point for a second formulation, which
needs some additional denitions listed on the following. Dene
Lp as the greatest common divisor (GCD) of all the shipping capacities of all classes of tankers that are allowed to pick up oil from
platform p. Let kmin p; d be minimum number of lots of size Lp that
should be shipped before day d from platform p to avoid its maximum inventory capacity to be exceeded. This number is given
by the formula:

2.2. Knapsack cascading formulation

For this third formulation, which is adapted from Rocha et al.
(2011), some of the previous denitions are considered plus others
that follow.

DAp;d AC p;d  CAPp;d ;

DAt;d CAPt;d  AC t;d :
Thus, the Knapsack Cascading (KC) formulation is given by:

KC min

D
XX XX
2F c Dp;t zc;d
p;t ;

p2P t2Tpc2Cp d1

subject to
d X
XX

d X
XX

V c zc;p;ts 6 AC p;d

8p 2 P; d 2 f1;...;Dg;

10

V c zc;p;ts P DAp;d

8p 2 P; d 2 f1;...;Dg;

11

t2Tp s1 c2Cp
d X
XX

d
X
ep;0
Pp;s :

c;sDp;t

V c zp;t

P AC t;d

8t 2 T; d 2 f1;...;Dg; 12

p2Pt s1 c2Cp

s1

Also, let kmax p; d be the maximum number of lots of size Lp

allowed to be shipped until day d from platform p avoiding its
inventory level to be less than zero. Such number is calculated by:


AC p;d
;
kmax p; d
Lp
With the denitions given above, constraints (4) are replaced
with:

AC p;d  kmax p; dLp 6 ep;d 6 AC p;d  kmin p; dLp

d 2 f1; . . . ; Dg:

8p 2 P;
7

Now, dene Lt as the GCD of all the shipping capacities of all

classes of tankers that are allowed to supply terminal t. Let
kmin t; d be the minimum number of lots of size Lt that must be received by terminal t until day d such that its inventory level is not
negative. This number is given by:



AC t;d
;
Lt

where AC t;d is the accumulated amount of product at terminal t

from day 0 to day d and is given by:

AC t;d et;0 

t2Tp s1 c2Cp

where AC p;d is the accumulated amount of product at platform p

from day 0 to day d and is given by:

kmin t; d

8t 2 T; d 2 f1;. ..;Dg:

With the substitutions suggested above, the second formulation

for the problem, which will be called Basic Formulation (BF ), is
given by the objective function (1) subject to constraints (2), (3),
(7), (8) and (6).

AC p;d  CAPp;d
;
kmin p; d
Lp

AC p;d

AC t;d kmin t;dLt 6 et;d 6 AC t;d kmax t;dLt

d
X
C t;s :

s1

Let kmax t; d be the maximum number of lots of size Lt allowed

to be received by terminal t until day d without exceeding its
inventory capacity. Such quantity is calculated by the following
formula:



CAPt;d  AC t;d
;
kmax t; d
Lt
Given this, constraints (5) are replaced with:

d X
XX

c;sDp;t

V c zp;t

6 DAt;d

8t 2 T; d 2 f1;...;Dg;

13

p2Pt s1 c2Cp

zc;d
p;t 2 f0;1g

8p 2 P; t 2 Tp;c 2 Cp; d 2 f1;... ;Dg: 14

The objective function (9) minimizes the total transportation

cost. Constraints (10) avoid the inventory levels at each platform
to be less than the minimum capacity. Constraints (11) ensure that
the inventory levels at each platform never exceed their maximum
inventory capacity. Constraints (12) avoid the inventory levels at
each terminal to be less than the minimum capacity and constraints (13) ensure that the inventory levels at each terminal
never exceed their maximum inventory capacity. Finally, the z
variables are dened by (14) as binaries.
From the KC formulation, it is possible to derive a fourth formulation replacing constraints (10) and (11) by the new constraints
(15) and (16) that follow:
d X
XX

V c zc;p;ts 6 kmax p;dLp

8p 2 P; d 2 f1;... ;Dg

15

V c zc;p;ts P kmin p;dLp

8p 2 P; d 2 f1;...;Dg

16

t2Tp s1 c2Cp
d X
XX
t2Tp s1 c2Cp

and substituting (12) and (13) by (17) and (18) below:

d X
XX

c;sDp;t

V c zp;t

6 kmax t;dLt

8t 2 T; d 2 f1;...;Dg

17

P kmin t;dLt

8t 2 T; d 2 f1;...;Dg:

18

p2Pt s1 c2Cp
d X
XX

c;sDp;t

V c zp;t

p2Pt s1 c2Cp

The resulting formulation will be called Rounded Knapsack

Cascading (RKC) formulation.

L. Aizemberg et al. / European Journal of Operational Research 237 (2014) 8290

formulation, proving that it has the same linear relaxation solution

set as the RKC over the z variables. h

2.3. Accumulated rounded cascading formulation

The variables of this newly proposed formulation are dened as
follows: xc;d
p;t is an integer variable representing the number of
tankers of class c 2 Cp leaving platform p 2 P to terminal
t 2 Tp from day 1 to day d. The zc;d
p;t variables are also considered
as previously dened. Let xc;0
p;t 0. The Accumulated Rounded
Cascading (ARC) formulation is given by:
ARC min

XX X

2F c Dp;t xc;D
p;t ;
p2P t2Tpc2Cp

85

The same proof can be applied for NF ; KC and AC, where AC is

the accumulated cascading formulation without rounding, leading
to the following proposition:
Proposition 2.2. NF ; KC and AC have the same linear relaxation.

19

2.4. DantzigWolfe decomposition

subject to
c;d
c;d1
xp;t
xp;t
zc;d
8p 2 P; t 2 Tp;c 2 Cp; d 2 f1;...;Dg;
p;t
X X
c;d
V c xp;t
6 kmax p;dLp 8p 2 P; d 2 f1;...;Dg;

20
21

t2Tpc2Cp

X X

c;d
V c xp;t
P kmin p;dLp

8p 2 P; d 2 f1;...;Dg;

22

t2Tpc2Cp

X X

c;dDp;t

6 kmax t;dLt

8t 2 T; d 2 f1;...;Dg;

23

c;dDp;t

P kmin t;dLt

8t 2 T; d 2 f1;...;Dg;

24

V c xp;t

p2Ptc2Cp

X X

V c xp;t

p2Ptc2Cp
c;d
2 Z
xp;t

8p 2 P; t 2 Tp; c 2 Cp; d 2 f1;...;Dg;

c;d
zp;t
2 f0;1g or 0;1

8p 2 P; t 2 Tp; c 2 Cp; d 2 f1;...;Dg:

25

26

The objective function (19) is to minimize the total transportation cost. Constraints (20) dene the x variables in terms of the z
variables. Constraints (21) and (22) ensure that the inventory level
in each platform p at day d is not less than the minimum inventory
capacity and does not exceed the maximum inventory capacity.
Constraints (23) and (24) avoid the inventory level in each terminal
t at day d to exceed the maximum inventory capacity and to be less
than zero. At last, constraints (25) dene the x variables as nonnegative integers and constraints (26) dene the z variables as binaries
or continuous variables in the range zero and one. Note that constraint (20) ensures the integrality of variables z regardless of their
denitions.
Proposition 2.1.
relaxation.

BF ; RKC and ARC have the same linear

Proof. We will prove the proposition in two steps. First, we will

show that RKC is obtained from BF without changing its linear
relaxation solution set over the z variables. Second, we will show
that ARC is obtained from RKC without changing its linear relaxation solution set over the z variables. Now, consider the BF . Given
the constraints (2) and expanding ep;d1 , we will have:

ep;d AC p;d 

d X
XX

V c zc;p;ts

8p 2 P;

d 2 f1; . . . ; Dg

t2Tp s1 c2Cp

Here, we present a DantzigWolfe decomposition (DWD)

(Dantzig & Wolfe, 1960) for the problem using the (RKC) formulation as starting point. The resulting reformulation is primarily used
to derive improved lower bounds for the original problem via Column Generation (CG). The results obtained with this formulation
are used as a benchmark to analyze the linear relaxation of the
implemented formulations of this paper. Furthermore, we observe
that the stronger linear relaxation provided by DWD helps to select
z variables that are suitable for building good feasible solutions.
Motivated by this observation, we propose a CG based heuristic
in Section 3 which demonstrates the importance of the DWD
formulation.
The following notation is considered for the reformulation: let
Sp be the set of all feasible solutions for a platform p 2 P, let V t;d
p;s
be the amount of product shipped from a platform p 2 P to terminal t 2 Tp at a day d in a solution s 2 Sp , and let C p;s be the cost of a
solution s 2 Sp for a platform p 2 P. Binary variables kp;s indicate if a
solution s 2 Sp for a platform p 2 P is used. The reformulation obtained after applying the DantzigWolfe decomposition, the so
called DantzigWolfe Master DWM, is the following:
DWM min

XX

C p;s kp;s ;

27

p2P s2Sp

subject to

X
kp;s 1 8p 2 P;

28

s2Sp



AC t;d
Lt 8t 2 T; d 2 f1;...;Dg;
L
t
p2P s2Sp s1


d
XXX
CAP t  AC t;d
Lt 8t 2 T; d 2 f1;...;Dg;
V t;p;ssDp;t kp;s 6
L
t
p2P s2S s1
d
XXX

V t;p;ssDp;t kp;s P

29

30

kp;s 2 f0;1g

8p 2 P; s 2 Sp :

31

The objective function (27) aims to minimize the sum of transportation costs. Constraints (28) ensure that exactly one solution
s 2 Sp variable is chosen for each platform p 2 P. Constraints (29)
avoid the inventory level at each terminal to be less than zero. Constraints (30) ensure that the maximum inventory capacity of each
terminal will not be exceeded. Finally, constraints (31) dene the k
variables as binaries.

From (7),

AC p;d  kmax p;dLp 6 AC p;d 

d X
XX

V c zc;p;ts

6 AC p;d  kmin p;dLp

t2Tp s1 c2Cp

8p 2 P; d 2 f1; .. .; Dg
which is equivalent to (15) and (16). The same procedure can be
used to prove that (3) and (8) have the same linear relaxation solution set of (17) and (18) over the z variables. Thus, RKC has the
same linear relaxation solution set over the z variables as BF .
Now, consider the RKC. Dene the x variables as follows:

xc;d
p;t

d
X
zc;p;ts

s1

and replace the right-hand side of the previous equality by xc;d

p;t in
the RKC. Including the constraints (20), we obtain the ARC

2.4.1. Pricing subproblem

Given the large number of variables in DWM formulation, its direct use is impractical for large instances. To overcome this issue, a
Column Generation (GC) procedure is often used in order to solve
the linear programming relaxation of DWM. This procedure is
started by considering only a subset of its variables. The problem
with such subset of variables is called the Restricted Master Problem (RMP). At each iteration of the CG procedure, the RMP is solved
and the dual variables of its optimal solution are used to generate
new k variables by solving the Pricing Subproblem. A new k variable
is added to RMP if its reduced cost, given by the objective function
value of the Pricing Subproblem, is negative. If, at a given iteration,
it is no longer possible to generate a k variable with negative reduced cost, the procedure is nished and the current solution for
RMP is the optimal solution for the linear relaxation of DWM.

86

L. Aizemberg et al. / European Journal of Operational Research 237 (2014) 8290

Let pp be the vector containing the dual variables associated to

constraints (28), ht;d be the vector containing the dual variables
associated to constraints (29), and rt;d be the vector containing
the dual variables associated to constraints (30). The binary variable zc;d
p;t indicates if the platform p sends a tanker of class c to terminal t at day d. Thus, for each platform p, a Pricing Subproblem is
dened as follows.

SP p min

D
X XX
c;d
2F c Dp;t  V c ht;dDp;t rt;dDp;t zp;t
 pp ; 32
t2Tpc2Cp d1

subject to
d X
XX

Fig. 1. Dynamic programming with two lot sizes.

zc;p;ts V c 6 AC p;d 8d 2 f1; .. . ; Dg;

33

t2Tp s1 c2Cp
d X
XX

zc;p;ts V c P DAp;d

8d 2 f1; . . . ;Dg;

34

t2Tp s1 c2Cp

zc;d
p;t 2 f0; 1g

8t 2 Tp; c 2 Cp; d 2 f1; . . . ;Dg:

35

The purpose of the Pricing Subproblem is to generate a k variable to be added to the RMP with the lowest reduced cost, which
is represented by the objective function (32). Constraints (33)
avoid the inventory level at each platform to be less than zero. Constraints (34) ensure that the maximum inventory capacity at each
platform will not be exceeded. At last, z variables are dened as
binary by (35). Even though the Pricing Subproblem is presented
as an Integer Linear Program, it can be solved more efciently by
dynamic programming.
3. Column generation-based heuristic
This section presents the column generation-based heuristic
(CGH). First, we show the dynamic programming algorithm to
solve the subproblems faster than the MIP of subSection 2.4.1.
Next, we show the heuristic procedure.
3.1. Pricing algorithm
In the following denitions we assume that the elements of the
set C are arbitrarily numbered from 1 to jCj. We dene our dynamic
programming cost table for platform p with Costp d; c; e being the
smallest partial reduced cost from day d to the end of the time
horizon, considering only the tanker classes c; . . . ; jCj at day d
(and all classes for the next days), and inventory level e at the
beginning of day d. Let Costp d; c; e 0 for d D 1. The dynamic
programming recursion is the following:
(
!
)
8
X
X c;d
>
>
>

min
Cost
d;c

1;e

P

V
a
C
a
p
c t
p;d
>
p;t t ; c 1
>
a2P 1;p;c;d;e
>
>
t2T
t2T
>
>
(
!
)
>
<
X
X c;d
min Cost p d;c 1;e 
Cost p d;c; e
V c at
C p;t at ; 1 < c < jCj
a2P 2;p;c;d;e
>
>
t2T
t2T
>
>
(
!
)
>
>
>
X
X c;d
>
>
>
V c at
C p;t at ; c jCj;
: min Cost p d 1; 1; e 
a2P 3;p;c;d;e

t2T

t2T

where

P 1;p;c;d;e fa 2 f0; 1gjTj j0 6 e Pp;d 

X
V c at g;

capacity of the platforms. For the harder class of instances, its performance is quite superior to when a MIP is used. Fig. 1 shows an
example with two lot sizes C f1; 2g. The daily production is always integer. There are two terminals with route to all platforms.
Each lot size can be sent only once per day, for each terminal. In
this example, the inventory capacity of the platform is CAPp 4
units, the daily production is P p;d 2 units, d 2 f1; 2; 3g, which is
added between the rst and the second lot size, and the initial
inventory is 2 units. The rst lot has size V 1 1 unit and costs
1.5. The second lot has size V 2 2 units and costs 2. Both can be
sent to all terminals, with the same cost. If there are two or more
terminals to send a lot size with different costs, the algorithm
chooses always the cheapest one according to the reduced costs.
In Fig. 1, the costs for each inventory level, day and lot size are
represented by the circles. At d 1 and c 1, there is an initial
inventory of 2 units and the lowest cost is Costp 1; 1; 2 4, which
is also the optimal cost. From d 1 and c 1 to d 1 and c 2,
there are three possibilities: (i) to send one lot (at 1, for some t),
resulting in inventory of 3 units and cost 3 1:5; (ii) to send two lots
(at 1; 8t 2 T), resulting in inventory of 2 units and cost 2 3; or
(ii) do not send a lot at 0; 8t 2 T, resulting in inventory of 4
units and cost 4. The third option is chosen. From d 1 and
c 2 to d 2 and c 2, no lot is sent. Then, one lot is sent from
d 2 and c 2 to d 3 and c 1, and from d 3 and c 2,
resulting in an inventory of 4 units at the end of the time horizon.
The main advantage of the dynamic programming procedure is
that the economy of scale on the lot sizes does not impact on the
runtime, differently of the MIP formulation where we observe a
major impact. This happens because the magnitude of the economy of scale only affects the reduced cost values, not the number
of calculations made by the algorithm.
For the case where the daily production is fractional in at least
one day, we need to do a preprocessing step. We accumulate the
daily productions, and if the accumulated value is fractional, we
decrease the inventory capacity by one unit to have space for the
fractional part. The adjusted daily production is the integer part
of the accumulated production, not added yet to the inventory.
Table 1 shows an example with fractional production where the
inventory capacity has 6 units. The rst column shows the daily
original production, which is fractional and equal 2.75 for all days.
The second column shows the accumulated production. Note that
there is only one integer value. Third column shows the original
capacity, 6 units. Fourth column shows the adjusted capacity,

t2T

P 2;p;c;d;e

X
fa 2 f0; 1g j0 6 e 
V c at g;
jTj

Table 1
Dealing with fractional production.

t2T

P 3;p;c;d;e fa 2 f0; 1gjTj j0 6 e 

X
V c at 6 CAPp g;
t2T

and C c;d
p;t 2F c Dp;t  V c ht;dDp;t rt;dDp;t . The optimal reduced cost
for platform p is given by Costp 1; 1; ep;0 .
The Dynamic Programming algorithm is pseudopolynomial. Its
computational complexity depends on the maximum inventory

Original
production

Accumulated
production

Original
capacity

Adjusted
capacity

Adjusted
production

2.75
2.75
2.75
2.75
2.75

2.75
5.5
8.25
11
13.75

6
6
6
6
6

5
5
5
6
5

2
3
3
3
2

L. Aizemberg et al. / European Journal of Operational Research 237 (2014) 8290

which is equal the original capacity if the accumulated production

is integer, or equal the original capacity minus one, if the accumulated capacity is fractional. The fth column shows the adjusted
production, which is equal the integer part of the accumulated production not yet added to the inventory.
3.2. Column generation-based heuristic procedure
The proposed heuristic uses the column generation procedure
described in SubSection 2.4 with the pricing algorithm presented
in the previous subsection to choose only some z variables of
ARC with good chance to appear in the optimal solution. This
reduction in the number of variables allows the branch-and-bound
algorithm to nd fast a feasible solution, or prove that no solution
can be found with these variables.
Such Column Generation-based heuristic is shown in Algorithm
1. First, the linear relaxation of the DWM formulation is solved by
column generation (line 1). Next, the reduced costs of all columns
present in the nal RMP are calculated using the values of the dual
variables in the last iteration of the column generation algorithm
(line 2). As these columns must have non-negative reduced costs,
we then sort such columns in ascending order according to their
reduced costs (line 3). The parameters min and max are initialized
in line 4. In line 5, the parameter k is initialized. This parameter
indicates that the kth rst columns of the nal RMP must be converted into variables of ARC. At each iteration of the main loop
(lines 616), the ARC formulation is populated with only the z
variables corresponding to the k columns with least positive reduced costs in the nal RMP (line 7). The ARC formulation run
for up to ve seconds (line 8). If a new feasible solution is obtained
(line 9), we keep the solution if it improves the best found so far. If
the branch-and-bound enumeration was not completed within the
given ve seconds (line 12), we nish the loop. Otherwise, we increase k by ten (line 15).
It may happen that some platform has no chosen z variable in
the reduced ARC formulation. If this is the case, the resulting
MIP will be trivially infeasible and a larger value for k will be tried.
When k is larger than the number of RMP columns with reduced
cost zero, every platform will have at least one converted column.
The time limit used in line 8 has been chosen based on preliminary experiments where we observed that most best feasible
solutions found by the algorithm were attained when solving a
MIP exactly in less than ve seconds. Thus, this time limit seems
to give a good compromise between solution quality and runtime.
Algorithm 1. Column Generation based heuristic
1: Run column generation to solve linear relaxation of DWM
2: Calculate the reduced cost of the columns using the dual
variables of the last iteration
3: Sort the columns according to their reduced costs
4: min
30; max
150
5: k
min
6: while k 6 max do
7: Convert k columns of DWM with least positive reduced
costs into variables of ARC
8: Run ARC for up to ve seconds
9: if feasible solution found then
10:
Keep the solution if the best found
11: end if
12: if enumeration incomplete within ve seconds then
13:
Exit while
14: end if
15: k
k 10
16: end while

87

4. Computational results
In this section we show the results obtained by the described
formulations using the instances proposed in Rocha et al. (2011)
and a new set of harder instances. All the tests, including those
for the formulation with the best performance proposed by Rocha
et al. (2011) (HullRel), were run in a PC with an Intel Core 2 Quad
Q8300 2.50 gigahertz CPU, 2 gigabytes RAM under Windows 7 OS
using ILOG CPLEX 12.1. All instances have 9 platforms and 2 terminals, and are divided into 4 classes of 25 instances each with the
following characteristics:





only one lot size (easy);

one lot size per route (medium);
two lot sizes per route (hard);
ve lot sizes per route and all platforms are allowed to send
product to all terminals (harder new instances).

In this work, the results obtained for the rst class of instances
(easy) will not be reported, due to the fact that such instances are
easily solved by BF combined with a commercial solver. The results for each instance class are shown separately. The tables contain, for each formulation, the average results obtained. Among the
results, two performance measures are shown, given by:

gapLB 100 

UBb  LBf
;
UBb

36

gapUB 100 

UBf  UBb
;
UBb

37

where UBb is the best integer feasible solution found among all formulations, LBf is the lower bound for a given formulation, and UBf is
the best integer feasible solution found by a given formulation.
The results obtained for the medium and hard classes of instances are shown in Tables 2 and 3, respectively. In the tables,
the column Formulation shows the formulations tested. The formulation ARCbin is the ARC formulation with z variables as binaries, ARCcont is the ARC formulation with z variables as
continuous between 0 and 1 and AC is the ARC formulation without rounding the inventory. Column Time (seconds) shows the
average execution time in seconds for each class of instances. For
all tests performed, the maximum execution time allowed for each
instance is 720 seconds. Column gapLB rootlp (%) shows the average gapLB at the root node before CPLEX cuts, calculated using the
gapLB of each instance given by (36) considering LBf as the lower
bound at the root node before CPLEX cuts. Column gapLB root
(%) shows the average gapLB at the root node after CPLEX cuts, also
calculated using the gapLB of each instance given by (36) considering LBf as the lower bound at the root node after CPLEX cuts.
Column gapLB nal (%) contains the average gapLB at the end of
the execution time of each instance and it is also calculated by
(36), where LBf is the lower bound at the end of the execution time
of each instance. Column gapUB (%) shows the average gapUB obtained at the end of executions and calculated by (37) for each instance. Column No. of nodes contains the average number of
nodes solved at the branch-and-bound algorithm. Finally, column
Inst. solved shows the number of instances of each class solved
to optimality.
Concerning the gapLB root lp , the results include the CPLEX
preprocessing routines. If they were not used, NF ; KC and AC
(see Proposition 2.2) would provide the same relaxation at root
node of the branch-and-bound tree, and the same would happen
with BF ; RKC and ARC (see Proposition 2.1). Indeed, for the
medium and hard instance classes, the average values of the linear
programming relaxation of the rounded formulations are 0.42 and
3.04, respectively, and for the formulations without rounding are

88

L. Aizemberg et al. / European Journal of Operational Research 237 (2014) 8290

Table 2
Average results and number of instances solved for the mathematical formulations and results taken from Rocha et al. (2011) (HullRel) for the medium class of instances.
Formulation

Time (seconds)

gapLB root lp

gapLB root (%)

gapLB nal (%)

gapUB (%)

No. of nodes

Inst. solved

NF
BF
KC
RKC
AC
ARCbin
ARCcont
DWM
HullRel

720
685.47
405.50
405.50
30.65
31.76
30.61

351.53

1.41
0.36
0.35
0.35
0.34
0.33
0.33
0.33

1.19
0.30
0.23
0.23
0.19
0.20
0.18

0.83
0.24
0.12
0.12
0.00
0.00
0.00

0.00
0.00
0.00
0.00
0.00

0.04

2,595,996
3,944,016
569,080
785,543
17,288
12,319
14,663

0
2
11
11
24
24
24

13

Table 3
Average results and number of instances solved for the mathematical formulations and results taken from Rocha et al. (2011) (HullRel) for the hard class of instances.
Formulation

Time (seconds)

gapLB root lp

gapLB root (%)

gapLB nal (%)

gapUB (%)

No. of nodes

Inst. solved

NF
BF
KC
RKC
AC
ARCbin
ARCcont
DWM
HullRel

720.00
720.00
687.51
664.28
103.93
111.13
102.94

691.40

3.14
2.98
2.84
2.84
2.38
2.38
2.38
1.24

2.74
2.62
1.76
1.76
1.32
1.30
1.30

1.98
1.95
1.11
1.11
0.77
0.77
0.77

1.07
1.07
0.04
0.04
0.04

1.24

1,294,071
1,326,605
157,950
156,079
55,632
52,975
57,962

0
0
2
2
22
22
22

1.47 and 3.19. Without CPLEX preprocessing, DWM would have

the best gapLB rootlp for all instance classes. Yet, for the hard class
of instances, DWM found the best average gapLB at the root node
even when the CPLEX preprocessing routines and cuts are used
for the other formulations.
In Tables 2 and 3 one can observe that the NF and BF have the
worst performance in comparison with the others. The gapUB of
them is not shown because these formulations found few feasible
solutions, so average value is not appropriate. The BF was capable
of solving only two instances to optimality, both of the medium
class. This fact impacts in the average execution time for this formulation, which is close to the maximum allowed for each instance. For the hard class of instances, as this formulation was
not able to solve any instance to optimality, the average time is exactly equal to the maximum allowed for each instance. The NF did
not solve any instance, so the average time is exactly 720 for both
classes of instances. On the other hand, since these are more compact formulations with respect to the remaining ones in the sense
that it has less nonzero coefcients in the matrix, a signicantly
larger number of nodes were solved at the branch-and-bound
algorithm.
Comparing KC with RKC and AC with ARCcont , the performance
are similar. We can conclude that formulation NF is dominated by
BF but the effect of the inventory rounding in KC and AC is not
noticeable, even for GCD of lot sizes greater than one.
In comparison with BF , the RKC formulation achieved a better
performance, being able to solve 13 out of 50 instances, where 11
instances belong to the medium class and 2 are hard instances.
This occurs because in the RKC formulation the constraints
(10)(13) are recognized by CPLEX as knapsack constraints
(for (11) and (12) all binary variables are replaced with their
complements to derive equivalent knapsack constraints). As a
result, CPLEX is able to add specialized cuts, e.g. cover cuts. This
effect can be observed by comparing the root lower bounds of
BF and RKC in Tables 2 and 3.
The best performance was obtained by the ARC formulation,
outperforming the literature results reported in Rocha et al.
(2011). This formulation was able to solve 24 instances of the medium class and 22 hard instances in a signicantly small average

Table 4
Size of all formulations for an instance of the harder class.

Number of variables
Number of constraints
Number of nonzero coefs
a
b

NF

BF

KC

RKC

ARC

3030
330
6060

3030
330
6060

2700
660
167,400

2700
660
167,400

5400
3360a
18,900b

80% of these constraints are in constraints (20).

43% of these variables are in constraints (20).

execution time. For the 4 instances not solved by this formulation,

it found 3 of the best known feasible solutions. For the 49 remaining instances, the gapUB is equal to zero. Thus, the average gapUB is
equal to zero for the medium class of instances and close to zero
for the hard instances.
The success of this formulation can be explained by three
main reasons: (i) the number of nonzero coefcients in ARC is
much smaller than in RKC (and comparable to that of BF ) because each x variable in the constraints (21)(24) replaces a
sum of many z variables; (ii) as in the RKC, CPLEX recognizes
the constraints (21)(24) as knapsack constraints; and (iii) the
x variables are more suitable for branching. To explain the item
(i), we observe that the number of nonzero coefcients of
NF ; BF and ARC is OjPkTkCjD while for KC and RKC is
OjPkTkCjD2 . Table 4 compares the sizes of all formulations for
an instance of the harder class. To explain the item (iii), we refer
to Fig. 2. The graph of Fig. 2(a) represents the z variables associated to one route for the whole time horizon: the arcs represent
the variables, the rst column of nodes represents the source
platform and the second column of nodes represents the
destination terminal. Now, assume that z15 has a fractional value
in a relaxed solution for the RKC formulation. If z15 is selected
as a branching variable and forced to assume the value one,
the relaxed solution will probably change signicantly. However,
when z15 is forced to be zero, it is likely that the relaxation gives
the same lower bound if there is sufcient slack in the inventories of both p1 and t 1 to send the same amount of product one
day latter, i.e. to set z16 to the same value as z15 had before

89

L. Aizemberg et al. / European Journal of Operational Research 237 (2014) 8290

Table 6
Average results and number of feasible solutions found for the column generationbased heuristic with and without violation allowed in terminals for the harder class of
instances generated.
Formulation

Time (seconds)

gapUB (%)

Feasible found

CGH with violation

CGH no violation

11.5
13.2

0.12
0.40

16
14

Min

XX X

2F c Dp;t xc;D
p;t

p2P t2Tpc2Cp

D


XX
min
max
C v iol v iolt;d v iolt;d

38

t2T d1
max

We also include the new variables v iolt;d in the constraints (23)

min
and v iolt;d in the constraints (24), as follows:
X X
c;dD
max
V c xp;t p;t  v iolt;d 6 kmax t;dLt 8t 2 T; d 2 f1;...;Dg 39
p2Ptc2Cp

X X

c;dDp;t

V c xp;t

min

v iolt;d P kmin t;dLt

8t 2 T; d 2 f1;.. .;Dg 40

p2Ptc2Cp

Fig. 2. (a) Representation of a set of z variables corresponding to the same route and
tanker class. (b) Branching on z variables. (c) Branching on x variables.

branching. This observation is depicted in Fig. 2(b). This characteristic, when replicated in several parts of the solution, generates an undesirable symmetry that prevents nishing the
branch-and-bound enumeration even in instances where the root
gap is very small. As shown in Fig. 2(c), the use of x variables
breaks this symmetry. Because of that, the ARC formulation
was tested in two ways: with the z variables as binaries
(ARCbin ), and with these variables as continuous between zero
and one (ARCcont ). Comparing both ways in Tables 2 and 3, one
can see that the ARCcont formulation shows a slight improvement
regarding the average gapLB at the root node in medium and hard
instances, and a little improvement concerning the average nal
gapLB in the hard instances. Thus, although the difference between
the two assumptions for the z variables is not expressive, we consider that the ARCcont formulation obtained the best results for the
instances of the medium and hard classes.
The results obtained for all formulation for the harder class of
instances are shown in Table 5. For this class of instances, we test
the original ARC and a modied version of this formulation. The
modication consists of allowing both inventory shortages and
maximum inventory capacities violations in the consumption sites,
keeping the violation cost high enough to avoid changing the optimal solution. For this modication, we include two new sets of
continuous variables:
max

v iolt;d : maximum capacity inventory violation at terminal

t and day d,
min
 v iolt;d : inventory shortage at terminal t and day d.


Let C v iol be the cost of inventory violations. As we do not want to

change the optimal solution, we use a large enough value. The new
objective function will be as follows:

For the harder class of instances, the GCD is always one. Rounding the inventory just drops the fractional part of the accumulating
production or consumption. So, we show only the results of the
rounding formulations. To nd optimal or near optimal solutions
for all instances of this class (to be used as the UBb value), we
run each instance for many hours, using as initial feasible solution
the best solution found in previous tests. Again the ARC has the
best performance. Comparing this formulation with and without
violation, we can see a good improvement in the quality of the
feasible solutions. This occurs because, when allowing violation,
the number of feasible solutions increases, facilitating the CPLEX
heuristic procedures to construct and improve feasible solutions.
Only the knapsack cascading formulations nd feasible solutions
for all instances, but the average gapUB of the ARC is better because
when this formulation nds a feasible solution for an instance, it
usually improves the solution faster than the knapsack cascading
formulations.
Concerning the DWM, it obtained the best average root relaxation, even when the CPLEX preprocessing routines and cuts are
considered for the other formulations. Without CPLEX preprocessing and cuts, the average value of the linear programming relaxation for all formulations is 8.82.
The results obtained for the column generation-based heuristic
are shown in Table 6 for the harder class of instances. Column
Feasible Found shows the number of instances where the heuristic nds a feasible solution, and column gapUB (%) shows the
average gapUB for these solutions. We also test the heuristic including violation in the ARC. Comparing the CGH with and without
violation, we can see a good improvement when the violation is
allowed. This occurs because the number of feasible solutions increases. The average gapUB for both cases shows that, as expected,
the heuristic nds good feasible solutions. For the instances where
the heuristic nds feasible solutions, the gapUB of the ARCv iol formulation is 0.69. Moreover, this formulation reaches feasibility
for 6 out of the 9 instances where the heuristic fails to nd feasible
solutions. However, the average time to obtain a feasible solution

Table 5
Average results and number of instances solved for the mathematical formulations for the harder class of instances generated.
Formulation

Time (seconds)

gapLB root lp

gapLB root (%)

gapLB nal (%)

gapUB (%)

No. of nodes

Inst. solved

BF
RKC
ARC
ARCv iol
DWM

720.00
720.00
637.64
642.36

8.49
8.26
7.81
7.81
4.74

7.50
6.53
5.88
5.89

6.37
5.34
5.16
5.19

10.9
6.20
1.85

562,356
107,204
137,416
110,018

0
0
3
3

90

L. Aizemberg et al. / European Journal of Operational Research 237 (2014) 8290

in these cases is 245.9 seconds. The average run time of both

variants of the heuristic is low, with a minimum of 7 seconds
and a maximum of 16.78 seconds.
One may think that, as the knapsack cascading formulation
nds feasible solutions for all instances, the heuristic should be
used with this formulation. But our tests show that the
accumulated cascading formulation has a better performance. This
happens because the heuristic decreases the variable set, so it is
easier to nd a feasible solution. Comparing both formulations,
when ARC nds a feasible solution, usually the solution is better
than RKC.
5. Conclusions
This work presented a modeling study for the oil transportation
problem. Three classes of instances taken from the literature were
used to test the formulations. Given the good results obtained, a
new class was generated with harder instances.
Six formulations were devised: the Natural Formulation (NF ),
the Basic Formulation (BF ), the Knapsack Cascading Formulation
(KC), the Rounded Knapsack Cascading Formulation (RKC), the
Accumulated Rounded Cascading Formulation (ARC), and the
DantzigWolfe Master Formulation (DWM). The rst ve formulations were tested with the CPLEX branch-and-cut algorithm, and
the best results found used the ARC formulation and its variants.
Without the CPLEX cuts, the DWM formulation achieved the best
lower bounds, which motivated the development of a column generation-base heuristic.
In comparison with the results reported by Rocha et al. (2011),
the results obtained in this work show a good improvement. This
achievement is due only to a better understanding of the problem,
incorporated in the mathematical formulation, reducing the computational effort.
Rounding the inventories improves the formulation only for
NF , not being noticeable when applied in RKC and ARC.
The harder class of instances has been used to evaluate the ability of the best formulations to nd good feasible solutions when
the optimality could not be proved. In this case, the ARC formulation achieved the best performance when allowing inventory
capacity violation in the demand sites, but keeping the violation
cost high enough to ensure that the cost of the optimal solution
will not change. The column generation-based heuristic used for
the harder class of instances found good feasible solutions in a
small computational time. We showed that it is also possible to improve the quality of the feasible solutions found by this heuristic
allowing inventory capacity violations.
References
Aizemberg, L., Uchoa, E., Pessoa, A., Rocha, R., Coutinho, R., & Paula, U. C., Jr. (2011).
Optimisation Model for the problem of oil transportation with local search by

MIP and simulation. In Proceedings of XLIII SBPO 2011, UbatubaSP, Brazil, in

Portuguese.
Bahl, H. C. (1983). Column generation based heuristic algorithm for multi-item
scheduling. IIE Transactions, 15(2), 136141. http://dx.doi.org/10.1080/
05695558308974624.
Banaszewski, R. F., Tacla, C. A., Pereira, F. R., Arruda, L. V., & Enembreck, F. (2010).
Planning transport of crude oil derivatives with simultaneous auctions. In 2010
IEEE international conference on systems man and cybernetics (SMC)
(pp. 28202827). IEEE.
Boschetto, S. N., Magato, L., Brondani, W. M., Neves, F., Jr., Arruda, L. V. R., BarbosaPvoa, A. P. F. D., et al. (2010). An operational scheduling model to product
distribution through a pipeline network. Industrial & Engineering Chemistry
Research, 49(12), 56615682.
Brown, G., & Graves, G. (1981). Real-time dispatch of petroleum tank trucks.
Management Science, 27(1), 1932.
Choi, E., & Tcha, D.-W. (2007). A column generation approach to the heterogeneous
eet vehicle routing problem. Computers & Operations Research, 34(7),
20802095.
Colvin, M., & Maravelias, C. T. (2011). R&D pipeline management: Task
interdependencies and risk management. European Journal of Operational
Research, 215(3), 616628. http://dx.doi.org/10.1016/j.ejor.2011.06.023. ISSN
0377-2217.
Dantzig, G. B., & Wolfe, P. (1960). Decomposition principle for linear programs.
Operations Research, 8(1), 101111.
Furini, F., Malaguti, E., Durain, R. M., Persiani, A., & Toth, P. (2012). A column
generation heuristic for the two-dimensional two-staged guillotine cutting
stock problem with multiple stock size. European Journal of Operational
Research, 218(1), 251260. http://dx.doi.org/10.1016/j.ejor.2011.10.018. ISSN
0377-2217.
Ghiani, G., Laporte, G., & Musmanno, R. (2004). Introduction to logistics systems
planning and control (Vol. 14). Wiley.
Joncour, C., Michel, S., Sadykov, R., Sverdlov, D., & Vanderbeck, F. (2010). Column
generation based primal heuristics. Electronic Notes in Discrete Mathematics,
36(0), 695702. http://dx.doi.org/10.1016/j.endm.2010.05.088. ISSN 15710653.
Magato, L., Arruda, L. V. R., Neves, F., et al. (2004). A mixed integer programming
approach for scheduling commodities in a pipeline. Computers & Chemical
Engineering, 28(12), 171185.
MirHassani, S. (2008). An operational planning model for petroleum products
logistics under uncertainty. Applied Mathematics and Computation, 196(2),
744751.
Moccia, L., Cordeau, J.-F., Monaco, M. F., & Sammarra, M. (2009). A column
generation heuristic for a dynamic generalized assignment problem. Computers
& Operations Research, 36(9), 26702681. http://dx.doi.org/10.1016/
j.cor.2008.11.022. ISSN 0305-0548.
Moro, L., & Pinto, J. (2004). Mixed-integer programming approach for short-term
crude oil scheduling. Industrial & Engineering Chemistry Research, 43(1), 8594.
Mourgaya, M., & Vanderbeck, F. (2007). Column generation based heuristic for
tactical planning in multi-period vehicle routing. European Journal of
Operational
Research,
183(3),
10281041.
http://dx.doi.org/10.1016/
j.ejor.2006.02.030. ISSN 0377-2217.
Rocha, R. (2010). Petroleum supply planning: models, reformulations and algorithms,
Ph.D. thesis, Departamento de Informtica, PUC-Rio.
Rocha, R., Grossmann, I. E., & Poggi de Arago, M. V. S. (2011). Cascading knapsack
inequalities: Reformulation of a crude oil distribution problem. Annals of
Operations Research, 118.
Sawik, T. (2011). Scheduling in supply chains using mixed integer programming. Wiley.
ISBN 9781118029091. <http://books.google.com.br/books?id=BuOLXKt7cKsC>.
Yavuz, B., Aras, N., Kuban, I., & Ersoy, C. (2010). A column generation based heuristic
for sensor placement, activity scheduling and data routing in wireless sensor
networks. European Journal of Operational Research, 207(2), 10141026. http://
dx.doi.org/10.1016/j.ejor.2010.05.020. ISSN 0377-2217.