You are on page 1of 15

This article was downloaded by: [129.127.145.

240] On: 23 October 2017, At: 12:01


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

The Profitable Arc Tour Problem: Solution with a Branch-


and-Price Algorithm
Dominique Feillet, Pierre Dejax, Michel Gendreau,

To cite this article:


Dominique Feillet, Pierre Dejax, Michel Gendreau, (2005) The Profitable Arc Tour Problem: Solution with a Branch-and-Price
Algorithm. Transportation Science 39(4):539-552. https://doi.org/10.1287/trsc.1040.0106

Full terms and conditions of use: http://pubsonline.informs.org/page/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.

© 2005 INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
TRANSPORTATION SCIENCE informs ®
Vol. 39, No. 4, November 2005, pp. 539–552
issn 0041-1655  eissn 1526-5447  05  3904  0539 doi 10.1287/trsc.1040.0106
© 2005 INFORMS

The Profitable Arc Tour Problem: Solution with


a Branch-and-Price Algorithm
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

Dominique Feillet
Laboratoire d’Informatiquè d’Avignon, 339 Chemin des Meinajariés, Agroparc BP 1228, 84911 Avignon, France,
dominique.feillet@univ-avignon.fr

Pierre Dejax
École des Mines de Nantes and Communications and Cybernetic Research Institute of Nantes,
4, rue Alfred Kastler, 44307 Nantes, France, pierre.dejax@emn.fr

Michel Gendreau
CRT Université de Montréal, C.P. 6128, succursale Centre-ville, Montréal, Canada H3C 3J7, michelg@crt.umontreal.ca

I n this article, we introduce a new arc routing problem that we call the profitable arc tour problem. This
problem is defined on a graph in which profits and travel costs are associated with the arcs. The objective
is to find a set of cycles in the graph that maximizes the collection of profit minus travel costs, subject to
constraints limiting the number of times that profit is available on arcs and the maximal length of cycles. The
problem is related both to constrained flow problems and to vehicle-routing problems. We tackle it from this
standpoint and propose a branch-and-price algorithm for its solution. In the column-generation phase, the issue
of the collection decisions while traveling through the arcs is addressed. In the branching phase, the fact that
viewing solutions in terms of flow variables regularly induces an integer flow matrix leads us to introduce a
branching method called the flow-splitting method. Finally, the relationships of this problem with constrained
flow optimization are taken into account in an initial phase of the algorithm.
Key words: arc-routing problem; routing problem with profits; column generation; branch and price; bounded
cycle cover
History: Received: October 2003; revision received: August 2004; accepted: August 2004.

1. Introduction objective function is the sum of the collected profits


In this paper, we introduce a new arc-routing problem minus the sum of the travel costs of its arcs. A cycle is
that we call the profitable arc tour problem (PATP). feasible when its total length does not exceed a given
This problem is defined on a graph in which prof- value lmax , where the length of a cycle is the sum of
its and travel costs are associated with the arcs. The the lengths of its arcs. To improve legibility of the
objective is to find a set of cycles in the graph that text, the length of an arc vi  vj  ∈ A is supposed to be
maximizes the collection of profit minus travel costs, equal to its cost cij throughout the article; subsequent
subject to constraints limiting the number of times sections would be straightforward to adapt without
that profit is available on arcs and the maximal length this assumption. Note that in a feasible cycle some
of cycles. Let G = V  A be a complete directed vertices and arcs can possibly appear several times.
graph, where V = v1   vn is the vertex set of the The number of cycles belonging to a solution and col-
graph and A = vi  vj  1 ≤ i j ≤ n is its arc set. Let lecting profit on an arc vi  vj  ∈ Q is upper bounded
cij ≥ 0 be the cost associated with arc vi  vj . The cost by qij .
matrix is assumed to satisfy the triangle inequality It is important to note that a given cycle can be
and is subsequently called the travel cost matrix. Let selected several times in the solution and that a cycle
Q ⊂ A be a subset of the arcs of the graph. In addi- may contain arcs belonging to Q without collecting
tion to cij , each arc vi  vj  ∈ Q receives a profit value their profit (which might already be collected by other
pij ≥ 0 and a value qij ∈  indicating the number of cycles of the solution). In this last case, only the travel
times that this profit can be collected. Each time profit costs are counted on these arcs. Figure 1 displays an
is collected on an arc vi  vj , a value pij is gained. By instance of the PATP and its optimal solution.
analogy with flow problems, qij is called capacity of The context of this study is a tactical freight
arc vi  vj . The aim of the PATP is to find a collec- transportation-planning problem arising in the car
tion of feasible cycles in the graph G, maximizing the industry. In this context, a set of trips need to be
objective function, where the impact of a cycle in the planned for transporting freight between plants. Trips
539
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
540 Transportation Science 39(4), pp. 539–552, © 2005 INFORMS

profit 15
capacity 30
1 2
V = extremities of 10 x 10 square
profit 20 Cycle length limit = 35
capacity 10
: Arc with profit
: Arc without profit
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

profit 18
capacity 50 Optimum :
10 times cycle 1 4 3 1 (C1)
40 times cycle 1 4 3 1 (C2)

co
s
Cycle C1 : profit 55, cost 34

t1
4
Cycle C2 : profit 35, cost 34
cost 10
3 4
profit 17
capacity 70

Figure 1 An Instance of the PATP and Its Optimal Solution

can either be round trips or direct trips. A round trip compulsory. These problems are described in detail
has the same origin and destination and is restricted in Feillet, Dejax, and Gendreau (2004a). Classical
to a given length. A direct trip is not constrained but examples of such problems are the selective traveling
is more expensive, even if, for example, a direct trip salesman problem (e.g., Fischetti and Toth 1988) and
i → j is cheaper than a round trip i → j → i leaving the prize-collecting traveling salesman problem (e.g.,
the truck empty on its way back. It is easy to show Balas 1989). Note that routing problems with prof-
that this problem can be expressed as a PATP, with the its are essentially defined in a node-routing context,
above set Q, capacities qij and profits pij representing except for Deitch and Ladany (2000). However, it is
the set of freight transportation demands, the num- well known that transformations are possible in many
bers of times these transportation operations have to cases between node- and arc-routing problems (see,
be planned, and the differences in cost between round for example, Laporte 1997). Such transformations are
trips and direct trips (see Feillet, Dejax, and Gendreau precisely used by Deitch and Ladany (2000). Yet they
2002 for more details). Apart from this freight trans- are costly in term of size of the graph and should
portation application, the PATP also appears to be a priori be avoided. We will see in subsequent sec-
relevant to telecommunication network optimization, tions that it is possible to make use of the similarities
when cycles are used for reliability considerations,
between the PATP and routing problems with profits
as will be shown further. Throughout this article, we
without resorting to such transformations.
focus, however, on the formal specification of the
The second difference leads to an important remark.
PATP and leave practical applications aside.
Two of the main reasons for which vehicle-routing
As presented above, the PATP naturally relates to
problems are not as simple as min-cost network flow
arc-routing problems, especially the capacitated arc-
problems are that in vehicle-routing problems a lim-
routing problem (CARP). In the CARP, a fleet of vehi-
cles constrained by a load capacity and based at a ited number of vehicles is available and that these
depot is used to service demand present on some vehicles have to initiate their tour at some prede-
arcs—see, for example, Eiselt, Gendreau, and Laporte fined places. A simple illustration of this is provided
(1995) for a detailed review. There are two main dif- by the traveling salesman problem, which changes
ferences between PATP and CARP: First, in PATP, not to an assignment problem when these two condi-
all demands need to be satisfied; second, for the PATP, tions are relaxed (subtour elimination constraints are
the size of the fleet is unlimited, and it is not based removed). Thus, although defined as a routing prob-
at a depot. Let us consider more carefully these two lem, the PATP (where these two conditions do not
points. exist) is no more than a constrained min-cost net-
The first difference introduces the idea of profit and work flow problem. Indeed, it is easy to demon-
links the PATP to the so-called class of routing prob- strate that if the tour length constraint is relaxed, the
lems with profits. This class gathers routing prob- PATP reduces to the minimum cost circulation prob-
lems where two conflicting objectives are pursued lem, a well-known polynomial network flow prob-
together—minimization of travel costs and maximiza- lem (see, for example, Ahuja, Magnanti, and Orlin
tion of collected profit—knowing that profit is avail- 1993), for which several efficient algorithms exist. This
able on vertices (or arcs) and that collection is not remark induces three questions: What literature is
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
Transportation Science 39(4), pp. 539–552, © 2005 INFORMS 541

available on flow problems with an additional con- networks at a minimal cost. In spite of its similari-
straint? Does the solution of such a constrained flow ties with the PATP, the BCCP differs from it in many
problem enable one to find an optimal set of cycles? ways: The graph is undirected, it is not complete,
Is the PATP a polynomial problem? This third ques- and all the edges have to be covered. Furthermore,
tion is addressed in §2, where we prove that the PATP the tour length limit is not as important as in the
is NP-hard. For the time being, let us consider the two PATP, since the problem remains NP-hard after its
other questions. relaxation (Thomassen 1997). Nonetheless, this prob-
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

Several papers focus on flow problems with one lem and many similar problems that are related to
or several additional constraints. Let us mention the same context (the minimum-cost cycle-cover prob-
Belling-Seib, Mevert, and Muller (1988), Cohen and lem, Thomassen 1997; the cycle cover problem, Labbé,
Megiddo (1994), Yan (1996), and more recently Laporte, and Soriano 1998; the multiring design prob-
Mathies and Mevert (1998). However, in all cases, lem, Kennington, Nair, and Rahman 1999; the two-
the additional constraints involve any combination of connected network with bounded meshes problem,
flow variables, whereas the tour length constraint of Fortz, Labbé, and Maffioli 2000; and the cardinal-
the PATP cannot be expressed this way. Thus, this par- ity constrained covering traveling salesman problem,
ticular constraint has a deep impact on the structure Patterson and Rolland 2003) share with the PATP
of the problem and really distinguishes it from flow the border area between routing problems and flow
problems. problems.
Nevertheless, relaxing the constraint on tour length Actually, the backbone of this research is the solu-
may eventually prove to be an effective solution tion of the PATP through a procedure based on a
approach. This point brings us back to our second branch-and-price approach. It is organized as follows.
question: How difficult is it to derive a family of In §2, we propose a mathematical formulation of the
cycles from a solution expressed in terms of flow? PATP and address the question of its belonging to
If the cycles are not constrained, the answer is easy. the NP-hard class. Sections 3 and 4 focus on the
A simple algorithm is to select a vertex receiving algorithmic part of the paper: Section 3 describes the
flow, to follow a unit of flow until coming back to column-generation procedure included in the branch-
a vertex that has already been visited (flow con- and-price algorithm, while §4 is dedicated to the
servation ensures that this will occur), and to put branching phase. Computational results are reported
aside the resulting cycle. The complete set of cycles in §5, and the conclusion follows.
is obtained by repeating this simple procedure until The contributions of this research are manifold. It
flow is exhausted (see, for example, Ahuja, Magnanti, presents and analyzes a new problem with impor-
and Orlin 1993). Unfortunately, this algorithm does tant applications and provides insights into similar
not necessarily generate cycles with a constrained problems. It also describes a column-generation pro-
length. Indeed, nothing forces the followed path to cedure explicitly devised for arc routing and quanti-
come back quickly toward one of the previously vis- fies the drawbacks of a node-routing transformation.
ited vertices. Finally, it introduces a new branching policy that can
Not being aware of any efficient algorithm to gener- be applied to problems where nonbinary flows are
ate constrained cycles from a flow or to solve a circu- present on arcs.
lation problem with an additional constraint limiting
the cycle length, we chose not to address the PATP
from a flow perspective. Thus, the solution algorithm 2. Mathematical Formulation and
we have developed is designed as a routing problem Complexity of the PATP
algorithm, even if it takes advantage of the flow sub- The original situation of the PATP, between routing
structure of the PATP in its initial phase. The devel- problems and flow problems, results in some diffi-
opment of a complete flow-based algorithm is left to culties to formulate it with linear programming (LP).
future research. Typically, routing problems can be formulated with
It is worth mentioning that several other prob- binary variables xijk that indicate whether vehicle k
lems also consist in finding a set of cycles in a travels through arc vi  vj , and flow problems can be
graph. In particular, several telecommunication prob- formulated with variables xij , representing the flow
lems deal with covering the edges of an undirected value on the corresponding arc. As mentioned above,
graph. Among them, the bounded cycle-cover prob- the tour length constraint cannot be expressed by
lem (BCCP; Hochbaum and Olinick 2001) is probably using only xij variables. With xijk variables, it becomes
the most closely connected to the PATP. It consists easy to express, and the whole problem can be for-
of finding a minimum cost set of cycles of limited mulated. The major inconvenience of this formulation
length that cover the arcs of a network. Its interest is that the number of vehicles is not bounded a priori.
lies in ensuring the reliability of telecommunication Consequently, even if it is possible to determine an
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
542 Transportation Science 39(4), pp. 539–552, © 2005 INFORMS

upper bound (the sum of arc capacities for example), Feillet (2001) for a complete statement of the proof.
the k index will have to be allowed to take unneces- Briefly, items of the bin-packing problem are intro-
sarily high values and the model will not reflect the duced as arcs defining the set Q and cycles selected
genuine structure of the problem. in the solution of the PATP represent bins. Costs are
In the model that we describe below, we propose to defined properly to establish a relationship between
introduce path variables (binary variables associated the value of the PATP solution and the number of bins
with the selection of a cycle in the solution). Before we (cycles) used.
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

do so, we need to introduce the notion of collection


cycle.
3. A Column-Generation Algorithm
Definition 1. A collection cycle is an elementary
The model that is presented in §2 cannot be used
cycle of the graph, along with an associated charac-
directly for the solution of the PATP, except when
teristic vector indicating on which arcs of this cycle fairly small instances are considered. This stems from
profit is collected. the fact that the size of the set  grows exponentially
By elementary cycle, we mean that no vertex with instance size. Among the methods devised for
appears more than one time in the cycle. The notion of solving large-scale integer programming problems,
collection cycle permits to explicitly settle the vehicle the branch-and-price method has often been used suc-
behavior regarding the collection of profit. In the fol- cessfully, especially in the field of routing problems.
lowing, we will talk about collection cycles or cycles Branch and price is a particular case of branch and
when this is not ambiguous. Let  be the set of col- bound. At each node of the search tree, an upper
lection cycles having a length not greater than lmax in bound is computed by solving the linear relaxation
graph G. For every collection cycle rk ∈ , let ck be of the original problem; a column-generation method
its cost and pk be its profit, i.e., respectively, the sum is used to solve the linear programs that contain too
of the travel costs of the arcs it uses and the sum of many variables to be solved in a classical way.
the profits it collects. Finally, for every arc vi  vj  ∈ Q, Barnhart et al. (1998) and Vanderbeck (2000) focus
let akij = 1 if rk collects profit on arc vi  vj  and akij = 0 on the method per se and detail heavily many of its
otherwise. facets, outside a specific field of application. In the
We propose the following model for the PATP, area of vehicle routing, several branch-and-price algo-
which we call the master problem (MP): rithms have been developed (e.g., Desrosiers et al.
 1995; Savelsbergh and Sol 1998; Desaulniers et al.
maximize pk − ck xk (1)
rk ∈
1998a; and Boland, Clarke, and Nemhauser 2000). For
 vehicle-routing problems with profits, let us mention
subject to akij xk ≤ qij vi  vj  ∈ Q (2) Butt and Ryan (1999) and Gueguen (1999). Finally,
rk ∈
Barnhart, Hane, and Vance (2000) propose a branch-
xk ∈  rk ∈  (3) and-price solution procedure for a multicommodity
flow problem, where each commodity has to follow
where xk is the number of times that the collection a unique path between a given origin and a given
cycle rk ∈  appears in the solution. The objective (1) destination.
is to maximize the collected profit minus travel costs. In branch-and-price algorithms, an important part
Constraints (2) enforce arc capacities. is played by the column-generation procedure,
Note that, in the model, cycles are restricted to devoted to the solution of the linear relaxation of the
being elementary. The PATP definition does not spec- model. In this section, we present this part of the algo-
ify such a requirement, but, because any nonelemen- rithm, which we call CAP (for column generation for
tary cycle can be broken down into a set of elementary an arc-routing problem with profits). The branching
cycles, this constraint does not modify the problem. part of the algorithm is presented in §4.
We will explain in §3.3 why we have decided to
devise the model in this fashion. 3.1. Column-Generation Principle
An important point in the study of a new prob- The backbone of the method is a subproblem allowing
lem is to determine its complexity. Basing on the bin- the generation of interesting columns. Before further
packing problem, which is well known to be NP-hard developing this point, let us briefly describe how the
(see Garey and Johnson 1979), we have demonstrated method can be applied in our context.
that the PATP is NP-hard. We introduce MP1 , the restriction of MP to a
subset of variables 1 ⊂ . MP is then identically
Theorem 1. The PATP is NP-hard.
MP. MP1  is called the restricted master prob-
The transformation used and the proof of NP- lem. Let MPLP 1  be the linear relaxation of MP1 
hardness are rather simple but a bit too heavy to and D1  be its dual program. D is then the dual
be detailed here. The interested reader is referred to program of MPLP , that is, the dual of the linear
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
Transportation Science 39(4), pp. 539–552, © 2005 INFORMS 543

relaxation of MP. The optimal solution of MPLP 1 , dynamic programming algorithm are subject to mod-
with the simplex algorithm for instance, provides an ifications. Yet, on the whole, the same principles
optimal solution (∗ ) for D1 . This solution is equiv- are involved: Labels, representing partial paths, are
alently a solution of D, not necessarily feasible. sequentially extended with a checking of feasibility
When every constraint deriving from the collection based on resource constraints and the use of domi-
cycles of \1 is satisfied, the solution (∗ ) is feasible nance rules discarding dominated labels.
for D, and therefore optimal, because D is more In our situation, the set of paths to be explored is
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

constrained than D1 . When one or several con- the set of elementary cycles with a length not greater
straints deriving from the collection cycles of \1 than lmax . Within this set, we search for cycles with a
are violated, the principle of the column-generation negative cost value, with a cost evaluation described
method is to identify one or several of these con- further.
straints, with the help of a subproblem, to integrate Let us call d the set of feasible cycles (regarding
the corresponding variables in the set 1 . Thus, solv- the above conditions) visiting vertex vd and visiting
ing alternately MPLP 1  and the subproblem allows none of the vertices with a lower index. d 1 ≤ d ≤
convergence toward dual feasibility. The algorithm n − 1 defines a partition of the subproblem solution
terminates when the subproblem solution attests that space. Note that the size of d reduces drastically
there are no more violated constraints and therefore while d increases. For the subproblem solution, sets
that the current dual solution is feasible for D. d will be successively explored. Searching for a
In our situation, dual constraints are of the form cycle with negative cost in d remains to solving
 an elementary shortest path problem with resource
akij ij ≥ pk − ck  constraints (ESPPRC), using an appropriate cost func-
vi  vj ∈Q
tion that we have to specify. Vertex vd then stands for
where ij is the nonnegative dual variable associated a depot (the paths go from vd to vd ), and resources
with arc vi  vj  in the constraint set (2). The purpose will ensure that the paths are elementary and have
of the subproblem is thus to find collection cycles a length not greater than lmax . This problem can be
rk ∈  such that solved with a dynamic programming algorithm pro-
posed in Feillet et al. (2004b). This algorithm extends

0 < pk − ck − akij ij a standard solution algorithm devoted to the solution
vi  vj ∈Q of the nonelementary path version of the problem (see
Desrochers, Desrosiers, and Solomon 1992), by adding
It consists equivalently of columns with a positive resources enforcing the elementary path condition.
reduced cost in MPLP , when the basic solution is Note that this partitioning of the subproblem on the
the optimal solution of MPLP 1 . In the following, basis of the departure point is the usual way of tack-
we call these columns collection cycles with a positive ling multidepot vehicle routing problems with col-
reduced cost. Using the notation kij = 1 when vi  vj  umn generation (see, for example, Desaulniers et al.
is included in rk and kij = 0 otherwise, this condition 1998b).
can be expressed as: As specified above, the cost matrix has to be de-
  fined carefully to take into account the collection deci-
kij cij + akij ij − pij  < 0 sions on arcs and to ensure that a negative cost cycle
vi  vj ∈A vi  vj ∈Q
is indeed a collection cycle with a positive reduced
One of the key points determining the efficiency of cost. To this aim, when an arc vi  vj  is traveled
a column-generation algorithm is the quickness of the through, we expect it to cost cij when no collection is
subproblem in finding interesting columns. performed on it and cij + ij − pij when the available
profit pij is collected. As the collection is compulsory,
3.2. Subproblem Solution Procedure Description an important issue is how to decide whether profit
As previously indicated, column-generation meth- has to be collected.
ods have been widely used in the field of vehi- Decisions regarding the collection of profit on arcs
cle routing. The most successful subproblem solution are taken with respect to the cost matrix of the sub-
approach appears to be its modeling as a shortest problem. Let us call (c 1 ) this matrix. We define:
path problem with resource constraints, paired up • cij1 = cij for all arcs vi  vj  ∈ A\Q
with a dynamic programming resolution (see, e.g., • cij1 = cij for all arcs vi  vj  ∈ Q with ij − pij ≥ 0
Desrochers, Desrosiers, and Solomon 1992). Depend- • cij1 = cij + ij − pij for all arcs vi  vj  ∈ Q with
ing on which specific routing problem is to be solved, ij − pij < 0.
the resources that are considered in the constrained The profit will be collected when ij − pij < 0, which
shortest path problem and the functioning of the means that it is more interesting to collect the profit
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
544 Transportation Science 39(4), pp. 539–552, © 2005 INFORMS

while traveling through the arc than to pass with- 3.4. Generation of an Initial Set of Columns
out doing so. In other words, among the two possible The previous column-generation procedure is pre-
traversal modes of the arc (going from vi to vj with ceded by an initial phase that generates a first set
or without collection), only the most interesting is of columns. This phase exploits the flow substruc-
retained. It is important to note that establishing such ture of the PATP. It consists in solving a sequence
a dominance relation does not hinder the functioning of assignment problems, each one detecting poten-
of the algorithm, because the resource consumption tially interesting collection cycles in the graph. These
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

involved by the two traversal modes allows a priority assignment problems are successively solved within
order between them to be established. As an exam- two greedy heuristics H1 and H2 and a heuristic
ple, such a behavior would not be possible for other column-generation procedure.
arc-routing problems like the CARP, for which per- Heuristics H1 and H2 are both variants of a single
forming service on an arc involves the consumption solution scheme. They work iteratively in a greedy
of a resource (connected to the filling of the vehicle) fashion. At each iteration, collection cycles enabling
that passing without service does not induce. a profit are sought. For this purpose the following
assignment problem is defined.
3.3. Discussion About the Elementary Let us consider G1 , a duplicate of the graph G to
Path Condition which self-loops are added: G1 = V  A1  with A1 =
As explained above, when column generation is used A ∪ vi  vi  vi ∈ V . A cost matrix c 1 is associated
in the context of vehicle-routing problems, the sub- with A1 . For every arc vi  vj  ∈ A, cij1 is either cij1 − pij1
problem is often modeled as a shortest path problem or cij1 , depending on whether arc vi  vj  has profit or
with resource constraints (SPPRC). Hence, although not. For every arc vi  vi  ∈ A1 , cii1 = 0. The assignment
routes typically need to be elementary, the solu- problem is defined as follows: To every vertex of the
tion space of the subproblem includes nonelemen- graph a successor must be assigned, such that each
tary routes. This allows the subproblem solution to vertex receives a single successor, each vertex receives
speed up, since constraints enforcing the elementary a single predecessor, and the assignment cost is min-
path condition strongly penalize dominance rules. imized. A solution then consists of an arc set repre-
As a consequence, the restricted MP must be able senting precedence relations between vertices. Due to
to integrate the nonelementary cycles that the sub- the specific constraints, this arc set composes a set
problem is likely to generate; that is, the variable of elementary cycles and self-loops. Seeing that self-
set of the MP has to include nonelementary routes. loops have a zero cost value, elementary cycles are
nonpositive in the optimal solution. Thus, the opti-
The cost definition ensures that this modification of
mal resolution of this assignment problem provides
the MP is possible; indeed, visiting vertices twice (or
one or several disjoint elementary cycles of negative
more) is expensive and the optimal solution will natu-
cost, when these exist. Such an assignment problem
rally satisfy the elementary path condition. However,
is well known to have a polynomial complexity and
this property is not true for the linear relaxation of
can, for example, quickly be solved using the simplex
the MP, which deteriorates the quality of the bound-
algorithm (see Ahuja, Magnanti, and Orlin 1993).
ing scheme. The backbone of greedy heuristics H1 and H2 is
In §2, we decided to include only elementary cycles the solution of this assignment problem. It is used as
in the MP. Yet, in our case, it is possible to use the follows. Solving the assignment problem, a set of neg-
SPPRC as a subproblem instead of the ESPPRC, with- ative cost cycles can be obtained. Actually, as the cost
out changing the variable set of the MP. Indeed, we definition takes into account the profit collected, it is
have already mentioned that a nonelementary cycle is simple to derive a collection cycle set from the com-
equivalently a set of elementary cycles. Moreover, it is puted cycle set. Once these collection cycles are found,
easy to demonstrate that a feasible elementary cycle data are updated for the next iterations: For every
of negative cost exists if and only if a feasible nonele- arc vi  vj  ∈ Q belonging to a collection cycle, qij is
mentary cycle with negative cost also exists. Thus, decreased by h units, where h is the higher value such
the solution of the SPPRC provides one or several that no qij becomes negative; for each arc vi  vj  for
feasible nonelementary cycles that themselves pro- which qij has been changed to 0, cij1 is set to cij instead
vide one or several feasible elementary cycles of neg- of cij − pij (no more profit is available). This update
ative cost. The correct behavior of the procedure is corresponds to putting aside the collection cycles in a
therefore ensured. However, the value of the length greedy way. The assignment problem resolution and
limit lmax then becomes the only limit to the combina- the updates are performed repeatedly, until only self-
torial explosion of the state space of the subproblem. loops are found.
Hence, we have preferred to consider the subproblem Unfortunately, the constraint limiting the length of
as an ESPPRC in the CAP algorithm. In §5, the two the cycles is not considered in the assignment prob-
approaches are numerically compared. lem, which might thus generate infeasible cycles. The
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
Transportation Science 39(4), pp. 539–552, © 2005 INFORMS 545

way we proceed with these too-long cycles leads to then often consists of selecting an arc with a fractional
the two variants H1 and H2. In H1, infeasible cycles flow value and deriving one branch where this arc is
are put aside in a greedy fashion, as in other cycles, enforced and another branch where it is forbidden.
but are not considered in the resulting set of cycles. To do so, inadequate columns are forced to zero in
In H2, infeasible cycles are not put aside as long the MP, and inadequate arcs are removed from the
as there exists a feasible cycle in the current itera- subproblem graph. More generally, when this scheme
tion; when all cycles are nonfeasible, a single one is cannot be applied, the branching policy can consist
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

put aside. It is interesting to note that these heuris- of bounding the flow value for one or several arcs.
tics, which do not take into account the tour length A new constraint, expressed in terms of path vari-
limit, can be considered as greedy versions of a flow ables, then has to be added to the MP. This new con-
algorithm like the Ford and Fulkerson’s one (1956): straint defines a limit on the sum of the path variables
In both cases, new paths are found repeatedly for that traverse the arc(s) considered. Also, the dual vari-
increasing flow value. The assignment-based method able of this constraint has to be taken into account
that is included in H1 and H2 is particularly well during the subproblem phase by adequately adding
suited to our situation for two reasons. On the one its value on corresponding arcs.
hand, each assignment problem solution can provide In our situation, vertices are generally visited more
several negative cost elementary cycles. On the other than once and this last branching policy should
hand, for the same reason, the method tends to favor appear as a good approach. Unfortunately, the pre-
short cycles. liminary tests computed with procedure CAP showed
Actually, the initial set of columns is completed that expressing fractional solutions in terms of arc
with another procedure, which we call ENUM(3) and flow variables often results in an integer-flow matrix.
which consists of enumerating all the feasible elemen- It is thus not easy to determine a branching con-
tary cycles containing no more than three arcs, pro- straint based on arc flow variables discarding the cur-
vided that the collected profit is larger than the travel rent fractional solution. This situation is quite original
costs. because an integer arc-flow matrix usually denotes a
In the same way that we use an assignment prob- nonfractional solution. It thus immediately raises a
lem inside greedy heuristic procedures H1 and H2, question: Can a fractional solution exhibiting an inte-
we are able to insert it in a heuristic column-gener- ger flow variable matrix trivially be transformed into
ation procedure. We just need to copy the column- a feasible (integer) solution?
generation procedure of §3.1 but change the subprob- Unfortunately, the answer to this question is no,
lem into an assignment problem. The cost matrix def- at least by using only the current path variables; some
inition is the same as in §3.2, except for the zero cost situations might occur where a set of fractional path
self-loops that are added. The procedure ends when variables associated to an optimal solution leads to
no more interesting cycle is returned. an integer flow matrix and it is not possible to find
We call HCG the previous heuristic column-gener- an integer solution of the same cost using these vari-
ation procedure, initiated with the union of the collec- ables (a detailed counterexample is provided in Feillet
tion cycle sets resulting from H1, H2, and ENUM(3). 2001). Transforming a fractional solution into an inte-
HCG enables the generation of a sizeable set of ger by introducing new feasible cycles might be possi-
columns. The efficiency of using this set as an initial ble, but this seems rather difficult. We propose instead
set for the exact column-generation procedure will be to develop a branching scheme adapted to the case
demonstrated in the experiments described in §5. where the arc flow matrix is an integer.
To this end, we propose a method that we call the
4. Branching Scheme flow-splitting method. This method essentially pro-
In branch-and-price procedures, column generation is vides a theoretical tool ensuring the completeness of
applied at each node of the search tree. It necessitates the search without resorting to branching on path
being able to comply with the change of problem variables. Basically, when branching on arcs is not
structure due to the branching decisions. Typically, possible, we propose branching on short sequences of
branching on the problem variables does not suit the arcs (instead of long sequences when path variable
resolution of shortest path type subproblems (see, for branching is used). In §4.1, we describe this method;
example, Barnhart et al. 1998). The rule that is prefer- we then describe the remainder of the algorithm,
ably used is to express the fractional solution in terms in §4.2. The complete branch-and-price algorithm is
of arc flow variables and to branch on these. The val- called solution of an arc routing problem with profits
ues of the arc flow variables are obtained by super- (SARPP).
imposing the flow generated by the columns of the
solution. Usually, in routing contexts, the vertices are 4.1. Flow-Splitting Method
visited one time and consequently admit a unique The purpose of this subsection is to describe how we
successor (in integer solutions). The branching rule propose to tackle branching when the arc flow matrix
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
546 Transportation Science 39(4), pp. 539–552, © 2005 INFORMS

computed from a fractional solution may be an inte- Set of feasible cycles Level 0
ger. The key point of the method is this simple obser-
vation: The solution computed at a node of the tree
is an integer if and only if the flow associated to any
subset of path variables is an integer. A direct conse- P(1) P(2) P(i1) P(n-1) Level 1
quence of this observation is that, when the solution is
fractional, it is possible to select a subset of path vari-
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

ables such that its associated flow will be fractional. P(i1, i2) ¬ P(i1, i2) Level 2
Such a subset can then be used to generate a fractional
flow matrix and derive a branching. As mentioned
previously, the subset usually selected in the literature
is exactly the whole set of path variables, which in
this case may result in an integer flow matrix. On the
Level t
contrary, selecting subsets of Size 1 corresponds to P(i1, i2,…, it) ¬ P(i1, i2,…, it)
branching on the MP variables, which is generally
awkward, as we have already mentioned. Thus, our Figure 2 Flow-Splitting Principle
approach is a compromise approach between these
two extremes.
i1  i1 +1  i1  n−1, such that i1  i contains
The counterpart of using a subset of columns to
cycles beginning with the sequence vi1 → vi . These
derive a branching, that is, introducing a new con-
subsets define a partition of i1  (note that sequences
straint on a subset of columns, is that the new dual
vi1 → vi with i < i1 are not considered due to the d
variable has only to be handled by columns of that
definition; see §3.2). Two situations can then arise:
subset. As we will see with our policy, the subprob-
• There exists i2 such that the flow associated with
lem will have to be decomposed to treat the different
i1  i2  is fractional. We then use this set to derive
sets of columns. In a first step, let us describe how we a branching, and we call ¬i1  i2  the complemen-
propose to select the subset of variables. The way the tary set.
subproblems are adapted so that they only address • For every vertex vi i1 + 1 ≤ i ≤ n − 1, the flow
the chosen variable sets and the form of the branching associated with i1  i is an integer. We then select a
constraints will be discussed in the next subsection. set i1  i2  holding variables with a fractional value,
At first, we make use of the fact that in the subprob- and we split it in the same manner.
lem phase, we successively search for new columns We decompose the set of feasible cycles until we find,
in sets d, 1 ≤ d ≤ n − 1 (see §3.2). Even if this at level t, a subset of cycles beginning with a same
decomposition is not explicit in the MP formulation, it sequence s = vi1 → vi2 → · · · → vit for which the flow
highlights a natural splitting of the variable set, each is fractional. Note that t is expected to keep a low
variable being connected to a single set d. Thus, value in practice. We call i1   it , or more briefly
at a given node of the search tree and for a fractional s, this set and we use it for performing the branch-
solution, we start by searching for a set d such that ing in the way described below.
the flow associated to its variables is fractional. Com- As far as we know, our approach is original, even if
putationally, determining whether some flow matri- it is not new: When the subproblem is made of several
ces are fractional is very fast and simple. If we find components, branchings might be operated on a sin-
a set d with a fractional flow matrix, we use this gle component (see for example Barnhart et al. 1998
set to derive the branching, as explained in the sub- or Desaulniers et al. 1998a).
sequent section.
When the flow is an integer for every set d, 4.2. Complementary Information
let us describe how we proceed (see Figure 2). Note Using the flow-splitting method, we have to deal with
that practically this situation is rather rare and that new kinds of subproblems and new constraints.
this part is mainly for a theoretical matter. Splitting If, at a given node of the search tree, the level
Level 0 corresponds to the complete set of feasi- reached by the flow-splitting method is t, the fol-
ble cycles. This set is automatically partitioned into lowing subsets of columns will have to be explored
n − 1 subsets 1  n − 1, which corresponds for the search of new columns: 1  i1 − 1,
to Splitting Level 1. Let us assume that the flow is an i1 + 1  n − 1, i1   it , and ¬i1   ik 
integer for all these sets at Level 1. We select a set for 2 ≤ k ≤ t. Note that we have to explore n + t − 2
i1  holding variables with a fractional value, and sets, instead of the n − 1 initial ones, which is not
we split it. For this purpose, we partition its cycles so detrimental as long as t is low. To comply with
into subsets, depending on the second visited vertex. the structure of these different subsets, we have to
We thus obtain new subsets that we respectively call adapt the shortest path solution algorithm. We have
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
Transportation Science 39(4), pp. 539–552, © 2005 INFORMS 547

to deal with two new kinds of subproblems. For the Finally, an important point is that to accelerate
search of paths of the type i1   it , one just needs the solution, the column-generation procedure is not
to initiate the dynamic programming procedure with automatically executed at each node of the search tree.
the label associated to sequence vi1 → · · · → vit . Then Indeed, let us consider the case when the solution of
every label will be an extension of this original label. the linear relaxation of the restricted MP by the sim-
For the search of paths of the type ¬i1   ik , plex algorithm at a node N provides the same value
one just needs to initiate the dynamic program- that the optimal solution of the linear relaxation of the
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

ming procedure with the label associated to sequence MP for the father of N does. We know that the opti-
vi1 → vi2 → · · · → vik−1 and to remove arc vik−1  vik . mal solution value of the linear relaxation of the MP
The branching constraints will be defined as fol- at node N is between these two values. In this case,
lows. The flow-splitting method provides a subset of we are sure to have the optimal solution in hand, and
path variables s such that the flow matrix M s com- we do not need to generate new columns.
puted by superimposing the flow generated by these In practice, experiments show that flow-splitting at
variables is fractional. New constraints are deter- a level greater than 1 occurs only exceptionally. For
mined from matrix M s . A first possibility would have the experiments, we therefore use a simplified ver-
been to select an arc vi  vj  such that Mijs is fractional sion of the algorithm, designed in the following man-
and to add constraints bounding the flow issued from ner: When a node of the search tree needs a Level 2
s on this arc. However, Mijs might be almost inte- splitting, the node is postponed. Using this simple
gral, which would limit the impact of the new con- rule, the algorithm always happens to find the opti-
straints (computational experiments will show that mal solution before having to explore the postponed
the impact of the branching constraints is an impor- nodes (by always finding a feasible solution with a
tant issue here). Instead of an arc, we prefer to select value as good as the upper bound associated to these
an arc set. This set is selected such that the fractional nodes).
part of the flow value traversing this set is not null,
but as close as possible to 0.5. For this purpose the
line of M s having the largest number of fractional 5. Computational Results
values is selected. Calling vi the vertex correspond- The following computational study is developed with
ing to this line and vi1 the first vertex of s, candidate three objectives. The first is to validate our solution
sets are constructed by successively adding the out- approach; the second is to compare it with several
going arcs of vi (from the left to the right). Among the variants; the third is to evaluate the effects of the
resulting sets, the instance parameters, to highlight the PATP structure.
 set vi  vi1  vi  vi1 +1   vi  vj  Experiments were performed on randomly gen-
such that = i1 ≤p≤j Mips is the more fractional, is
retained. Again, note that arcs vi  v1   vi  vi1 −1  erated Euclidean instances. Vertices are randomly
are not considered because they cannot be traversed located on a square map with a width of 500 dis-
by cycles of the set s. tance units. The cost of an arc is the distance between
This permits two branches to be defined with the its two endpoints, with a precision of 10−1 . Unless
respective constraints: the opposite is explicitly stated, the other parame-
  k ters are set as follows. Each arc vi  vj  ∈ A belongs
ip xk ≤   to Q pij > 0 with a probability 50%; pij is then chosen
vp ∈vi1  vj rk ∈s
uniformly in the range $cij  1 5 × cij %, and the arc
and capacity qij is selected uniformly in $1 200%. The max-
  imal length lmax of a collection cycle is set to 1,500.
kip xk ≥   + 1 For each experiment, results indicate the mean over
vp ∈vi1  vj rk ∈s
10 instances. In most cases, results are only given
These constraints are introduced explicitly in the when the solution time is no more than one hour.
MP. During the subproblem phase, when the set s Another stopping criterion is met when the depth of
or one of its subset is evaluated, these constraints have the search tree exceeds 3,000.
to be taken account. Noting !, the dual variable associ- Experiments are performed on a PC Pentium 333
ated to one of these constraints, we simply have to add MHz with 128 MB of RAM. Even if specialized algo-
! to the cost of arcs vi  vi1  vi  vi1 +1   vi  vj  . rithms exist for solving assignment problems, they are
To complete the description of the algorithm, let us solved as linear programs using CPLEX 6.0 (ILOG
now detail how the tree is explored. At each step, 1998). This is also used for the solution of linear relax-
we consider the values of the linear relaxation for the ations of restricted MPs. The interface between CPLEX
fathers of the tree leaves and we select the leaf having and the programming language (C++) is realized with
the maximal value. In case of equality between sev- the help of LPToolKit 3.6 (EURODECISION 1998).
eral nodes, the deeper is selected. When equality still We begin the presentation of the results by exam-
occurs, the last node to enter the tree is chosen. ining in §5.1 the efficiency of the procedure SARPP,
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
548 Transportation Science 39(4), pp. 539–552, © 2005 INFORMS

that is, its ability to find the optimal solution quickly. to conjecture that the duality gap would remain very
In §5.2, we assess several variants of SARPP. The com- small whatever instances are used, which is usually
putational study ends with the study of the efficiency in favor of a quick finding of the integer solution.
of SARPP with respect to instance parameters, in §5.3. However, we note that the bottleneck of SARPP is
Computing times in the tables always correspond to not the search of the linear relaxation value, but the
the time spent since the beginning of the algorithm, branching phase. Indeed, CAP permits one to solve
even when it has several phases. instances with up to 65 vertices in about an hour,
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

whereas SARPP does not go over 25 vertices. Note


5.1. Efficiency of SARPP that very few iterations are needed in CAP and that
Table 1 evaluates the efficiency of the solution pro- most of the time is spent solving the subproblem.
cedure SARPP. The algorithm is applied to instances During the branching phase, the constraints added
with up to 25 vertices. Table 2 focuses on the column- to the model do not seem to be very influen-
generation solution procedure CAP, at the root of the tial. Indeed, for all computed instances, they never
search tree. resulted in a decrease of the optimal relaxed solu-
We first observe that the gap between the linear tion value. The lack of influence of these constraints
relaxation value and the optimal solution value is and the huge combinatorics of the problem (arc flow
always null. More generally, this observation leads us matrices that can be computed by superimposing the
path variables are not binary, as it often happens,
Table 1 Efficiency of SARPP but can reach high values) explain the difficulty in
reaching the integer optimal solution. In particular,
CAP SARPP
computing times are very variable, depending on the
n Q c 1
t 2
d 3
nd 4
iter 5
col 6 c 7 t 2 instances. However, the weakness of the added con-
5 10 4 0.00 0 22 0 0 0 0 47048 0 0 24 straints has an important advantage. While going
10 45 1 0.00 0 84 0 3 0 3 0 0 236952 9 0 87 down the tree, we never had to backtrack; indeed,
15 105 3 0.00 2 17 43 1 43 1 0 0 573179 3 3 79
the search tree was reduced to a chain going from
20 199 1 0.00 5 96 144 2 144 2 0 0 1123103 4 17 15
25 305 1 0.00 14 91 1003 5 1003 5 0 0 1848166 4 1256 43 the root node to an integer solution whose value was
1
exactly the value of the lower bound at the root node
Gap (percentage) between the optimal solution and the linear relaxation
value.
and, as such, was certified to be optimal. Equivalently,
2
Total CPU time in seconds. the number of spread-out nodes was always exactly
3
Depth of the optimal solution in the search tree. equal to the depth reached to find the optimal solu-
4
Number of nodes spread out in the search tree. tion in the search tree. Nevertheless, the branching
5
Number of calls of the subproblem (root excepted).
6
phase takes a long time, because of the high value of
Number of columns generated while exploring the search tree (root
excepted). this depth. Basically, our opinion is that there exists
7
Value of the optimal solution. a very large number of optimal integer solutions, but
there are so many more fractional solutions with the
Table 2 Efficiency of CAP
same cost that SARPP has difficulties finding them.
With this standpoint, trying to obtain an integer solu-
HCG CAP tion from the fractional solution provided by CAP by
n Q c 1 t 2 iter 3 col 4 c 5 t 2 fixing some path variables is certainly a promising
perspective.
5 10 4 100 00 0 20 1 8 7 47048 0 0 22
10 45 1 97 97 0 78 2 6 108 1 236952 9 0 84
Another important remark concerns the quality of
15 105 3 98 79 1 98 2 4 491 3 573179 3 2 17 the column set obtained at the root of the search tree.
20 199 1 99 13 5 23 2 2 1143 2 1123103 4 5 96 For all the computed instances, it was never necessary
25 305 1 99 81 11 65 1 9 2746 6 1848166 4 14 91 to generate new columns during the search. Thus,
30 440 2 99 95 24 39 1 8 3946 8 2635932 5 34 73
SARPP acted exactly as a branch-and-bound algo-
35 605 9 99 95 46 27 2 6049 0 3727594 2 71 89
40 786 9 99 99 71 20 2 5945 8 4972050 5 122 30 rithm. Furthermore, the subproblem was never called,
45 992 0 99 97 120 21 1 9 9258 2 5989998 9 231 94 since the relaxed solution value at each node of the
50 1228 9 99 99 203 93 1 9 11478 9 7638764 7 446 54 search tree was always identical to its father node.
55 1504 3 99 97 340 59 1 9 19197 2 9438438 0 876 80 Incidentally, note that using the standard branch-
60 1764 0 99 97 526 24 1 9 33596 7 10970388 7 1941 16
65 2074 3 99 98 831 55 1 8 35859 4 12823662 9 3609 24
and-bound procedure of CPLEX as a complement
of CAP instead of SARPP (which would provide a
1
Value obtained with HCG in percent of the linear relaxation value (≤100). heuristic solution in the general case but an opti-
2
Total CPU time in seconds.
3
Number of calls of the subproblem.
mal one for the above instances solved by SARPP)
4
Number of columns generated in 1 . permitted us to solve instances with up to 45 ver-
5
Linear relaxation value. tices. Note also that SARPP was used with success
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
Transportation Science 39(4), pp. 539–552, © 2005 INFORMS 549

in the context of the freight transportation prob- are less convincing. CG(ENUM(3)) does not suc-
lem described in §1: Real size instances with up ceed in solving instances when the number of ver-
to 20 plants (set V ) and 200 freight transportation tices reaches 25. Intuitively, the set computed with
demands (set Q) were solved in a few seconds. ENUM(3) suffers of the lack of lengthy cycles and,
consequently, the branching constraints appear to be
5.2. Comparison Between Several less effective. Furthermore, as the number of arcs
Solution Approaches in solution cycles is certainly related to instances
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

While describing CAP, we insisted on defining an ini- characteristics, CG(ENUM(3)) might be less robust
tial set of columns exploiting the constrained flow than CAP. Thus, when the size of the instances
structure of the PATP and of developing the column- increases, generating an initial set of columns with
generation procedure on arcs. It is now time to check HCG appears to be more effective.
these assertions. Subsections 5.2.1 and 5.2.2 respec- 5.2.2. Comparison Between Column Generation
tively focus on these subjects. The subsequent §5.2.3 on Arcs and on Nodes. Table 4 shows the effect of
considers another question, that of the respective transforming the arc-routing problem into a node-
efficiency of ESPPRC and SPPRC modelings of the routing problem. Before describing the results, let us
subproblem during the column generation process. recall briefly how to operate such a transformation.
In a first step, a transformed graph G = V   A  is
5.2.1. Impact of the Initial Set of Columns. Table 3
constructed. The vertex set V  is such that each ver-
compares five variants of the procedure CAP. In each
tex in V  corresponds to an arc in Q. The arc set A
variant, the column-generation procedure (called CG)
is such that the graph G is complete. A new distance
is initialized with a different set of columns 1 .
matrix is computed for G , using the distance matrix
• CG()—1 = . associated to G. The notion of collection cycle is no
• CG(ENUM(2))—1 is the set of all the collection longer necessary, since collection is on vertices and
cycles with two arcs and a positive cost. it is never necessary to visit a vertex without collect-
• CG(ENUM(3))—1 is the set obtained using ing profit, as a direct arc exists between every pair of
ENUM(3). vertices. That aside, the algorithm remains identical.
• CG(H)—1 is the set obtained using H1 and H2. In particular, in the dynamic programming procedure
• CAP—Procedure included in SARPP; 1 is the used for the subproblem solution, the fact that the
set obtained using HCG. graph G is derived from a geographical graph G is
Even if the initial set of columns seems to concern not exploited. This exploitation would have permit-
essentially the MP linear relaxation computation (at ted more efficiency, but our goal is to compare the
the tree root), it also seems interesting to evaluate its arc routing approach with a node-routing approach
impact on the branching phase. following an automatic transformation. In Table 4,
Table 3 demonstrates the necessity of having CAP-node is the solution procedure devoted to the
columns before starting the column-generation proce- node-routing problem.
dure. Procedures CG() and CG(ENUM(2)) are not Results clearly demonstrate the interest of using the
competitive, except for small instances. CG(H) is not geographic graph. When the problem is transformed
far better. On the other hand, the use of ENUM(3) into a node-routing problem, the number of vertices
proves very efficient. Even so, when we focus on com- increases significantly and the subproblem solution
puting times for finding an integer solution, results procedure becomes inefficient. Thus, it confirms the

Table 3 Impact of Initial Set of Column for the Solution of the PATP

CG() CG(ENUM(2)) CG(ENUM(3)) CG(H) CAP


1 2 1 2 1 2 1 2 1
n t t t t t t t t t t 2

5 0 04 0 05 0 03 0 05 0 02 0 03 0 21 0 23 0 22 0 24
10 0 08 0 11 0 07 0 09 0 06 0 08 0 66 0 69 0 84 0 87
15 2 16 53 32 1 30 37 93 0 31 3 14 2 47 17 38 2 17 3 79
20 7 54 4 37 884 26 1 01 34 67 7 64 287 74 5 96 17 15
25 134 15 51 31 9 85 48 22 14 91 1256 43
30 736 38 22 92 256 37 34 73
35 45 25 1940 79 71 89
40 131 90 122 30
45 209 85 231 94
1
CPU time in seconds to find the linear relaxation value.
2
Total CPU time in seconds to obtain the integer optimal solution value.
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
550 Transportation Science 39(4), pp. 539–552, © 2005 INFORMS

Table 4 Comparison of Solutions in the Geographic Graph and in the because it enables a large decrease in the number
Transformed Graph of negative cost cycles during the subproblem phase.
CAP CAP-node Indeed, in many cases, no such cycle remains in the
graph and a unique iteration of the subproblem is
n it 1 col 2 t 3 Q it 1 col 2 t 3 performed. Our opinion is that applying the SPPRC-
5 1 8 7 0.22 10 4 1 8 7 0 24 based approach in cases where a large number of neg-
10 2 6 108 1 0.84 45 1 2 6 171 4 3 64 ative cost cycles remain in the graph would be far less
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

15 2 4 491 3 2.17 105 3 2 4 1005 8 221 93 efficient.


20 2 2 1143 2 5.96 191 1 2 3 1735 4 1276 60
Finally, it is important to recall that, even for small
1
Number of calls of the subproblem. instances, when lmax tends toward infinity, the size
2
Number of columns generated in 1 . of the state space also tends toward infinity when
3
CPU time in seconds for the solution of the linear relaxation.
nonelementary paths are authorized, while it is not
true in the opposite case. However, this is not a
necessity of developing specific procedures designed sufficient reason to dismiss CAP-SPPRC because §1
for arc-routing problems, even if, obviously, establish- showed that the PATP can be solved in polynomial
ing bridges between node- and arc-routing problems time when the length constraint is redundant.
is also necessary.
5.3. Scenarios on the Instance-Building Parameters
5.2.3. Relaxation of the Elementary Constraint of
Besides the previous tests, this subsection is devoted
the Subproblem. Table 5 compares using an ESPPRC
to developing some understanding of the evolution of
procedure with using a SPPRC procedure to solve the
the solution value and of the efficiency of SARPP, with
subproblem. In the last case, we call CAP-SPPRC the
respect to instance parameters. We successively eval-
procedure dedicated to the solution of the linear relax-
uate the impact of the length limit lmax , of the demand
ation of the PATP.
density (percentage of arcs belonging to Q), of the
To summarize results, it seems more efficient to use
maximal arc capacity, and of the profit value. We do
the SPPRC instead of the ESPPRC for the subproblem
not present heavily detailed results here, just some
inside the column-generation procedure CAP. Even
conclusions; detailed results can be found in Feillet
so, computing times are similar and the use of the
(2001).
SPPRC does not enable us to solve larger instances.
lmax value—A low value of lmax strongly reduces the
It is also important to notice that the use of HCG is
number of feasible cycles in the graph and permits
all the more worthwhile in the case of CAP-SPPRC
a more efficient solution. Nevertheless, an increase
of lmax is not completely negative. Indeed, the larger
Table 5 Comparison Between the Use of the ESPPRC and the SPPRC lmax is, the smaller the gap between HCG result and
as a Subproblem in CAP
the linear relaxation value is and the less numerous
CAP CAP-SPPRC the subproblem calls are. Thus, in many cases, a sin-
1 2 3 1
gle iteration proves that the HCG solution is optimal.
lmax n it col t it col 2 t 3
It is interesting to mention also that when using HCG
1,500 10 2 6 108 1 0 84 3 4 118 9 0 74 instead of ENUM(3) it is all the more important that
15 2 4 491 3 2 17 3 5 340 1 2 07 lmax value is high.
20 2 2 1143 2 5 96 3 804 4 4 73 Demand density—Instances can be solved faster
25 1 9 2746 6 14 91 2 5 1446 3 13 28
30 1 8 3946 8 34 73 2 2148 6 27 27
when demand density is either low or medium.
35 2 6049 0 71 89 2 4 3360 4 58 67 Capacity values—These values do not seem to have
40 2 5945 8 122 30 2 7 4697 1 100 26 a significant impact on the solution efficiency.
45 1 9 9258 2 231 94 2 4 7020 6 177 65 Profit values—Overall, the higher profit values are,
3,000 10 1 2 87 9 0 95 1 4 102 4 0 87 the harder the problem is. Yet, even with very low
15 1 4 937 0 2 55 1 7 596 7 1 97 profit values, the problem remains difficult to solve.
20 1 6 1283 5 6 94 1 8 943 0 5 42
Actually, profit values impact essentially the interest
25 1 6 5264 8 102 96 1 7 1618 7 14 76
30 1 5 4335 4 157 65 2 3 2391 6 38 84 of visiting arcs without collection. The lower profit
6,000 10 1 2 87 5 0 83 1 2 180 3 0 82
values are, the less arcs are visited without collection.
15 1 267 4 1 97 1 267 4 1 95
20 1 1 6759 2 275 03 1 1 1412 3 5 62
25 1 2 1598 5 24 58 1 2 1870 0 18 49 6. Conclusion
30 1 2 3812 4 129 25 1 2 2367 5 59 67 In this article, we have introduced a new problem,
1
Number of calls of the subproblem.
PATP, which is an arc-routing problem with profits,
2
Number of columns generated in 1 . and we have addressed the solution of this problem
3
CPU time in seconds to find the linear relaxation value. with a branch-and-price based procedure.
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
Transportation Science 39(4), pp. 539–552, © 2005 INFORMS 551

Several of the results presented with respect to our Boland, N. L., L. W. Clarke, G. L. Nemhauser. 2000. Asymmetric
column-generation approach (symmetry management traveling salesman problem with replenishment arcs. Eur. J.
Oper. Res. 123(2) 408–427.
while partitioning the variable set according to the
Butt, S. E., D. M. Ryan. 1999. An optimal solution procedure for
origin node, flow substructure exploitation) can be the multiple tour maximum collection problem using column
generalized to similar problems, where the aim is to generation. Comput. Oper. Res. 26(4) 427–441.
find a collection of constrained cycles in a graph. In Cohen, E., N. Megiddo. 1994. Algorithms and complexity analysis
the same fashion, the way collection decisions are for some flow problems. Algorithmica 11 320–340.
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

fixed for the subproblem, using dual variable values, Deitch, R., S. P. Ladany. 2000. The one-period bus routing prob-
is instructive for the use of column generation for lem: Solved by an effective heuristic for the orienteering tour
problem and improvement algorithm. Eur. J. Oper. Res. 127(1)
solving related problems. 69–77.
Branch-and-price methods have seen a rapid devel- Desaulniers, G., J. Desrosiers, I. Ioachim, M. M. Solomon, F. Soumis,
opment in recent years, but their use is still con- D. Villeneuve. 1998a. A unified framework for deterministic
fined to some particular problems. With this work, time constrained vehicle routing and crew scheduling prob-
we have contributed to the exploration of branch- lems. C. T. Laporte, G. Laporte, eds. Fleet Management and
Logistics. Kluwer Academic Publishers, Norwell, MA, 57–93.
and-price possibilities. In particular, the original struc-
Desaulniers, G., J. Lavigne, F. Soumis. 1998b. Multi-depot vehicle
ture of the PATP implied the introduction of a new scheduling problems with time windows and waiting costs.
branching method, which we called the flow-splitting Eur. J. Oper. Res. 111(3) 479–494.
method. Desrochers, M., J. Desrosiers, M. M. Solomon. 1992. A new opti-
Our computational experiments allowed us to as- mization algorithm for the vehicle routing problem with time
sess the efficiency of the SARPP procedure for solv- windows. Oper. Res. 40(2) 342–354.
ing the PATP and to better understand the inherent Desrosiers, J., Y. Dumas, M. M. Solomon, F. Soumis. 1995. Time
constrained routing and scheduling. M. Ball, T. Magnanti,
complexity of the problem. Results show that SARPP C. Monma, G. Nemhauser, eds. Handbooks in OR and MS, Vol. 8.
is able to solve, within a reasonable computing time, Elsevier Science Publishers B.V., Amsterdam, North-Holland.
instances with up to 30 vertices, that is, instances with Eiselt, H. A., M. Gendreau, G. Laporte. 1995. Arc routing problems.
hundreds of arcs providing profit that could not be Part II: The rural postman problem. Oper. Res. 43(3) 399–414.
tackled by merely transforming the problem into a EURODECISION. 1998. LPToolKit, User’s Guide and Technical Refer-
ence, 3rd ed. Eurodécision, Versailles, France.
node-routing one.
Feillet, D. 2001. Problèmes de tournées avec gains: Ètude et appli-
As a final remark, let us recall that the defini-
cation au transport inter-usines. Doctoral dissertation, Labora-
tion of PATP has been motivated by an application toire Productique Logistique, École Centrale Paris.
in freight transportation. However, many variants Feillet, D., P. Dejax, M. Gendreau. 2002. Planification tactique de
could be derived while facing other practical situa- transport de marchandises interusine. J. Européen des Systèmes
tions. Similar problems especially exist in the field of Automatisés 36(1) 149–168.
telecommunication network reliability, as noted in §1. Feillet, D., P. Dejax, M. Gendreau. 2004a. Traveling salesman prob-
lems with profits. Transportation Sci. 39(4) 540–553.
In our opinion, the results of this paper and the
Feillet, D., P. Dejax, M. Gendreau, C. Gueguen. 2004b. An exact
SARPP algorithm can possibly be adapted to these algorithm for the elementary shortest path problem with
other contexts. resource constraints: Application to some vehicle routing prob-
lems. Networks. 44(3) 216–229.
Fischetti, M., P. Toth. 1988. An additive approach for the optimal
Acknowledgments solution of the prize-collecting travelling salesman problem.
The authors gratefully acknowledge helpful comments from B. Golden, A. Assad, eds. Vehicle Routing: Methods and Studies.
three anonymous referees who helped improve the quality Elsevier Science Publishers B.V., Amsterdam, North-Holland,
of this paper. 319–343.
Ford, L. R., D. R. Fulkerson. 1956. Maximum flow through a net-
work. Canadian J. Math. 8 399–404.
References Fortz, B., M. Labbé, F. Maffioli. 2000. Solving the two-connected net-
work with bounded meshes problem. Oper. Res. 48(6) 866–877.
Ahuja, R. K., T. L. Magnanti, J. B. Orlin. 1993. Network Flows: Theory,
Algorithms, and Applications. Prentice Hall, Englewood Cliffs, Garey, M. R., D. S. Johnson. 1979. Computers and Intractability:
NJ. A Guide to the Theory of NP-Completeness, Vol. 50. Freeman,
Balas, E. 1989. The prize collecting traveling salesman problem. San Francisco, CA.
Networks 19(6) 621–636. Gueguen, C. 1999. Méthodes de résolution exacte pour les prob-
Barnhart, C., C. A. Hane, P. H. Vance. 2000. Using branch-and-price- lèmes de tournées de véhicules. Doctoral dissertation, Labora-
and-cut to solve origin-destination integer multicommodity toire Productique Logistique, École Centrale Paris.
flow problems. Oper. Res. 48(2) 318–326. Hochbaum, D. S., E. V. Olinick. 2001. The bounded cycle cover
Barnhart, C., E. L. Johnson, G. L. Nemhauser, M. W. P. Savelsbergh, problem. INFORMS J. Comput. 13(2) 104–119.
P. H. Vance. 1998. Branch-and-price: Column generation for ILOG. 1998. ILOG, Using the CPLEX Callable Library, 6th ed. ILOG
solving huge integer programs. Oper. Res. 46(3) 316–329. Inc., CPLEX Division, Incline Village.
Belling-Seib, K., P. Mevert, C. Muller. 1988. Network flow prob- Kennington, J. L., V. S. S. Nair, M. H. Rahman. 1999. Optimization-
lems with one side constraint: A comparison of three solution based algorithms for finding minimal cost ring covers in sur-
methods. Comput. Oper. Res. 15(4) 381–394. vivable networks. Comput. Optim. Appl. 14 219–230.
Feillet, Dejax, and Gendreau: The Profitable Arc Tour Problem
552 Transportation Science 39(4), pp. 539–552, © 2005 INFORMS

Labbé, M., G. Laporte, P. Soriano. 1998. Covering a graph with Savelsbergh, M., M. Sol. 1998. Drive: Dynamic routing of indepen-
cycles. Comput. Oper. Res. 25(6) 499–504. dent vehicles. Oper. Res. 46(4) 474–490.
Laporte, G. 1997. Modeling and solving several classes of arc rout- Thomassen, C. 1997. On the complexity of finding a minimum cycle
ing problems as traveling salesman problems. Comput. Oper. cover of a graph. SIAM J. Comput. 26(3) 675–677.
Res. 24(11) 1057–1061. Vanderbeck, F. 2000. On Dantzig-Wolfe decomposition in integer
Mathies, S., P. Mevert. 1998. A hybrid algorithm for solving net- programming and ways to perform branching in a branch-and-
work flow problems with side constraints. Comput. Oper. Res. price algorithm. Oper. Res. 48(1) 111–128.
25(9) 745–756. Yan, S. 1996. Approximating reduced costs under degeneracy in
Downloaded from informs.org by [129.127.145.240] on 23 October 2017, at 12:01 . For personal use only, all rights reserved.

Patterson, R., E. Rolland. 2003. The cardinality constrained covering a network flow problem with side constraints. Networks 27(4)
traveling salesman problem. Comput. Oper. Res. 30 97–116. 267–278.

You might also like