Downloaded from orbit.dtu.

dk on: Nov 29, 2023

Recent exact algorithms for solving the vehicle routing problem under capacity and
time window constraints

Baldacci, Roberto; Mingozzi, Aristide; Roberti, Roberto

Published in:
European Journal of Operational Research

Link to article, DOI:

10.1016/j.ejor.2011.07.037

Publication date:
2012

Document Version
Publisher's PDF, also known as Version of record

Link back to DTU Orbit

Citation (APA):
Baldacci, R., Mingozzi, A., & Roberti, R. (2012). Recent exact algorithms for solving the vehicle routing problem
under capacity and time window constraints. European Journal of Operational Research, 218(1), 1-6.
https://doi.org/10.1016/j.ejor.2011.07.037

General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright
owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

 Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
 You may not further distribute the material or use it for any profit-making activity or commercial gain
 You may freely distribute the URL identifying the publication in the public portal

If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately
and investigate your claim.
 European Journal of Operational Research 218 (2012) 1–6

Contents lists available at SciVerse ScienceDirect

European Journal of Operational Research

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

Invited Review

Recent exact algorithms for solving the vehicle routing problem under capacity and
time window constraints
Roberto Baldacci a, Aristide Mingozzi b,⇑, Roberto Roberti c
a
DEIS, University of Bologna, Via Venezia 52, 47521 Cesena, Italy
b
Department of Mathematics, University of Bologna, Via Sacchi 3, 47521 Cesena, Italy
c
DEIS, University of Bologna, Viale Risorgimento 2, 40136 Bologna, Italy

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

Article history: This paper provides a review of the recent developments that had a major impact on the current state-of-
Received 19 January 2011 the-art exact algorithms for the vehicle routing problem (VRP). The paper reviews mathematical formu-
Accepted 22 July 2011 lations, relaxations and recent exact methods for two of the most important variants of the VRP: the
Available online 6 August 2011
capacitated VRP (CVRP) and the VRP with time windows (VRPTW). The paper also reports a comparison
of the computational performances of the different exact algorithms for the CVRP and VRPTW.
Keywords:  2011 Elsevier B.V. All rights reserved.
Vehicle routing
Exact algorithms
Survey

1. Introduction In this paper, we survey the most effective exact methods pro-
posed for the CVRP and the VRPTW in the literature in the last ten
Vehicle routing problem (VRP) is a generic name given to a whole years. Section 2 reviews exact methods for the CVRP while Sec-
class of problems involving the design of optimal routes for a fleet of tion 3 deals with exact methods for the VRPTW. Conclusions are re-
vehicles to service a set of customers subject to side constraints. The ported in Section 4.
VRP is a central problem in the physical delivery of goods and ser-
vices. In practice, several variants of the VRP exist, depending on
the nature of the transported goods, the quality of service required, 2. The capacitated vehicle routing problem (CVRP)
and the characteristics of customers and vehicles. Some typical com-
plications are heterogeneous vehicles located at different depots, The CVRP can be described as follows. An undirected graph
customers incompatible with certain vehicle types, customers G = (V0 , E) is given where V0 = {0, 1, . . . , n} is the set of n + 1 vertices
accepting delivery within specified time windows, multiple-day and E is the set of edges. Vertex 0 represents the depot, and the ver-
planning horizons and vehicles performing multiple routes. In all tex set V = V0 n{0} corresponds to n customers. A nonnegative cost
cases, the objective is to supply the customers at minimum cost. dij is associated with each edge {i, j} 2 E. Each customer i 2 V re-
The simplest and most studied member of the VRP family is the quires a supply of qi units of goods from depot 0 (we assume
capacitated VRP (CVRP). In the CVRP, a fleet of identical vehicles lo- q0 = 0). A set of m identical vehicles of capacity Q is stationed at
cated at a central depot has to be optimally routed to supply a set the depot and must be used to supply the customers. A route is de-
of customers with known demands. Each vehicle can perform at fined as a least-cost elementary cycle R = (0, i1, . . . , ih, 0) of graph G
most one route and the total demand of the customers visited by passing through the depot and such that the total demand of the
a route cannot exceed the vehicle capacity. Another important var- customers visited does not exceed the vehicle capacity. The cost
iant of the VRP is the VRP with time windows (VRPTW) that gener- c(R) of a route is equal to the cost of the solution to the Traveling
alizes the CVRP by imposing that each customer is visited within a Salesman Problem (TSP) defined by the set R of vertices.
specified time interval, called time window. The CVRP is to design m routes of minimum cost, one for each
Several heuristics have been proposed for the CVRP and its variants vehicle, so that all customers are visited exactly once; the CVRP
in the literature. For surveys of both exact and heuristic methods for is NP-hard as it is a natural generalization of the TSP.
these VRPs up to 2002, the reader is referred to Toth and Vigo (2002). Many formulations have been proposed for the CVRP (see Achu-
than and Caccetta, 1991; Toth and Vigo, 2002; Kara et al., 2004), but
⇑ Corresponding author. not all of them have been used to derive exact algorithms. Currently,
E-mail addresses: [email protected] (R. Baldacci), [email protected] (A. the most successful exact methods for the CVRP are based on the
Mingozzi), [email protected] (R. Roberti). two-index flow formulation, the two-commodity flow formulation

0377-2217/$ - see front matter  2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2011.07.037
2 R. Baldacci et al. / European Journal of Operational Research 218 (2012) 1–6

proposed by Baldacci et al. (2004), and the set partitioning formula- and two variables yij and yji representing the vehicle load and the
tion proposed by Balinski and Quandt (1964). These algorithms are empty vehicle space (i.e., yji = Q  yij) on edge {i, j}, whenever
described in Sections 2.1 and 2.2. In Section 2.3, a comparison of xij = 1. The resulting model involves (2n + jEj + 3) constraints. It
the computational performances of these algorithms is reported. can be shown that (i) any solution (x, y) of the LP-relaxation of this
model is also a feasible solution of the LP-relaxation of F when
k(S) = q(S)/Q is used in constraints (3), and (ii) (x, y) satisfies the
2.1. Branch-and-cut algorithms
generalized multistar inequalities not implied by F.
Baldacci et al. (2004) described a BC algorithm based on this
The branch-and-cut (BC) algorithms for the CVRP are based either
model using only rounded capacity inequalities and reported com-
on the two-index flow or the two-commodity flow formulation.
P putational results showing that their BC algorithm is competitive
Let S ¼ fS : S # V; jSj P 2g, and let qðSÞ ¼ i2S qi be the total de-
with the algorithm of Naddef and Rinaldi (2002).
mand of customers in S 2 S and k(S) the minimum number of vehi-
cles of capacity Q needed to service all customers in S. Moreover,
2.2. Algorithms based on the set partitioning formulation
let d(S) be the cutset defined by S (i.e., d(S) = {{i, j} 2 E : i 2 S, j R S
or i R S, j 2 S}.
The Set Partitioning (SP) formulation of the CVRP was originally
Let xij be an integer variable that takes value in {0, 1}, for all
proposed by Balinski and Quandt (1964) and associates a binary
{i, j} 2 En{{0, j} : j 2 V}, and value in {0, 1, 2}, for all {0, j}, j 2 V, with
variable with each feasible route. The Balinski and Quandt formu-
xij = 1 when edge {i, j} is traversed and x0j = 2 when the single cus-
lation is the following.
tomer route (0, j, 0) is in the solution.
Let R be the index set of all routes. Let ai‘ be a binary coefficient
The two-index vehicle flow formulation of the CVRP is as
equal to 1 if vertex i 2 V belongs to route ‘ 2 R and 0 otherwise
follows.
(note that a0‘ = 1, 8‘ 2 R). Each route ‘ 2 R has an associated cost
X c‘ which is the sum of the costs of the edges traversed. Let n‘ be a
ðFÞ zðFÞ ¼ min dij xij ð1Þ
binary variable that is equal to 1 if route ‘ 2 R is in the optimal
fi;jg2E
X solution and 0 otherwise.
s:t: xij ¼ 2; 8h 2 V; ð2Þ Model SP is as following:
fi;jg2dðfhgÞ
X X
xij P 2kðSÞ; 8S 2 S; ð3Þ ðSPÞ zðSPÞ ¼ min c ‘ n‘ ð7Þ
fi;jg2dðSÞ ‘2R
X X
x0j ¼ 2m; ð4Þ s:t: n‘ 6 m; ð8Þ
j2V ‘2R
X
xij 2 f0; 1g; 8fi; jg 2 E n ff0; jg : j 2 Vg; ð5Þ ai‘ n‘ ¼ 1; 8i 2 V; ð9Þ
‘2R
x0j 2 f0; 1; 2g; 8f0; jg; j 2 V: ð6Þ
n‘ 2 f0; 1g; 8‘ 2 R: ð10Þ
Constraints (2) are the degree constraints for each customer, and
constraints (3) are the capacity constraints. Whenever k(S) = q(S)/Q, Constraint (8) requires that at most m routes are selected, and con-
constraints (3) are called fractional capacity inequalities. Constraint straints (9) specify that each customer i 2 V must be covered by one
(4) states that m vehicles must leave and return to the depot. route.
The LP-relaxation of F can be strengthened if k(S) is computed as In the following, we use R‘ and E(R‘) to indicate the subset of
k(S) = dq(S)/Qe, where dxe denotes the smallest integer not less than vertices of graph G visited and the edges traversed by route
x. In this case, constraints (3) are called rounded capacity con- ‘ 2 R, respectively.
straints. Several authors have proposed valid inequalities to rein- Formulation SP cannot be used directly to solve nontrivial CVRP
force the lower bound achieved by the LP-relaxation of F. Some instances because of the large number of potential routes. Model
of these inequalities are derived by extending, to formulation F, SP is very general and can take into account several route con-
the successful results of polyhedral combinatorics developed for straints (e.g., time windows) because the route feasibility is implic-
the TSP (see Grötschel and Padberg, 1985). itly considered in the definition of the route set R.
Augerat et al. (1995) were the first to describe an exact BC algo- The optimal solution cost of the LP-relaxation of SP, called LSP,
rithm for the CVRP based on formulation F strengthened by valid dominates the lower bounds provided by the LP-relaxation of for-
inequalities such as the generalized capacity constraints, hypotour mulation F when k(S) = q(S)/Q (see Baldacci et al., 2004) because
inequalities, comb inequalities, path-bin inequalities. The BC algo- any LSP solution implicitly satisfies a variety of valid inequalities
rithm of Augerat et al. (1995) was able to solve, for the first time, that are not implied by the LP-relaxation of F. In particular, Baldac-
a CVRP instance involving 135 customers. An improved version ci et al. (2004) have shown that LSP implies fractional capacity, gen-
of the BC algorithm of Augerat et al. (1995) is described in Naddef eralized large multistar and knapsack large multistar inequalities.
and Rinaldi (2002). Letchford and Salazar González (2006) have shown that LSP im-
Ralphs et al. (2003) described a BC algorithm based on the two- plies some hypotour-like inequalities by projection.
index formulation and on the addition of rounded capacity con- Moreover, Baldacci et al. (2004) have shown that any SP solu-
straints in a cutting plane fashion. tion n can be transformed into an F solution x by setting:
Lysgaard et al. (2004) described a BC algorithm based on formu-
X
lation F, strengthened by valid inequalities, including the rounded xij ¼ g‘ij n‘ ; 8fi; jg 2 E; ð11Þ
capacity, generalized capacity, framed capacity, strengthened comb, ‘2R
multistar, partial multistar, extended hypotour inequalities, and Gom-
ory mixed integer cuts. Their BC algorithm solved several instances where the coefficients g‘ij are defined as follows: (i) if ‘ is a single
not solved by Augerat et al. (1995). customer route (0, h, 0) then g‘0h ¼ 2 and g‘ij ¼ 0; 8fi; jg 2 E n f0; hg;
Baldacci et al. (2004) proposed a two-commodity flow formula- (ii) if ‘ is not a single customer route, then g‘ij ¼ 1; 8fi; jg 2 EðR‘ Þ
tion of the CVRP which extends the TSP model introduced by Finke and g‘ij ¼ 0; 8fi; jg 2 E n EðR‘ Þ.
et al. (1984). This formulation requires, for each edge {i, j} 2 E, a Let F be the family of valid inequalities for model F that can be
binary variable xij that is equal to 1 if the edge is in the solution, expressed in a general form as:
 R. Baldacci et al. / European Journal of Operational Research 218 (2012) 1–6 3

X
atij xij P bt ; t 2 F : ð12Þ respect to (u0 , v0 , g0 ) and zUB is a valid upper bound on the CVRP;
fi;jg2E (ii) add all constraints (14) and (15) whose associated slacks are
not in the optimal basis of LSP0 .
Thus, using Eq. (11), inequalities (12) become the following inequal- c using a general purpose integer programming
3. Solve problem SP
ities for the SP model:
X solver.
atij g‘ij n‘ P bt ; t 2 F : ð13Þ
fi;jg2E The effectiveness of this method relies on the quality of the dual
b depends on the
solution (u0 , v0 , g0 ) achieved as the size of subset R
2.2.1. The exact algorithm of Fukasawa et al. (2006) gap zUB  z0 .
Fukasawa et al. (2006) described an exact algorithm based on The core of the algorithm of Baldacci et al. (2008) is the method
the SP model where the variables correspond to the set of q-routes, for solving SP0 . They propose to solve LSP0 using three CCG proce-
introduced by Christofides et al. (1981), while the constraints cor- dures, called H1, H2 and H3, that produce three lower bounds LB1,
respond to the set partitioning constraints (8), (9) and valid LB2 and LB3 (such that LB1 6 LB2 6 LB3) corresponding to the costs
inequalities (13), such as rounded capacity inequalities, framed of three different dual solutions of LSP0 . The three procedures are
capacity, strengthened comb, multistar, partial multistar, generalized executed in sequence, and the dual solution produced by proce-
large multistar and hypotour inequalities, all presented in Lysgaard dure Hk is used to generate the master problem of procedure
et al. (2004) for formulation F. Hk+1, k = 1, 2.
A (q, i)-path is a nonnecessarily elementary path, starting from The first two procedures H1 and H2 ignore clique inequalities
the depot, visiting a set of vertices (without 2-vertex loops) of total (15) and use rounded capacity constraints instead of strengthened
demand equal to q, and ending at vertex i. Let f(q, i) be the cost of a capacity inequalities (14). H1 replaces the route set R with the set
least-cost (q, i)-path; functions f(q, i) can be computed by using dy- of all q-routes whereas H2 uses elementary routes instead of q-
namic programming (DP) (see Christofides et al., 1981). A q-route routes. The CCG method used by H1 and H2 differs from standard
is a (q, 0)-path of cost f(q, 0). CCG algorithms based on the simplex as it uses a dual ascent heu-
Because the resulting formulation has an exponential number of ristic to find a near-optimal dual solution of the master problem.
both columns and rows, Fukasawa et al. used a column-and-cut gen- H3 is a CCG method based on the simplex to solve LSP0 including
eration (CCG) method to compute the lower bound and a branch- both inequalities (14) and (15), that inherits the master problem
and-cut-and-price (BCP) algorithm to solve the CVRP. Their exact from H2. In practice, H3 is fast as it requires few iterations to con-
method decides at the root node, according to the best balance be- verge to an optimal dual solution of LSP0 .
tween running time and bound quality, to solve the CVRP either A key component of the algorithm of Baldacci et al. (2008) is the
using the BC method of Lysgaard et al. (2004) or their BCP algorithm. procedure, called GENROUTE, that solves the pricing problem in H2
and H3 and generates the final reduced problem SP. c

2.2.2. The exact algorithm of Baldacci et al. (2008) Let P be the set of all elementary paths of minimum cost from
Baldacci et al. (2008) proposed an exact algorithm based on mod- the depot such that q(P) 6 dQ/2e + qr(P), 8P 2 P, where qðPÞ ¼
P
el SP, strengthened by the following two types of valid inequalities. i2VðPÞ qi , r(P) and V(P) represent the load, the terminal customer
and the subset of visited customers of path P, respectively.
2.2.2.1. Strengthened capacity inequalities. These inequalities are GENROUTE is a two-phase procedure based on the observation
obtained by lifting the rounded capacity constraints of F expressed that every feasible route can be obtained by combining a pair of
in the form of inequalities (13). Let RðSÞ be the index subset of paths P; P 2 P such that: (a) rðPÞ ¼ rðPÞ; (b) VðPÞ \ VðPÞ ¼
routes visiting at least one customer of set S 2 S. The strengthened f0; rðPÞg; (c) qðPÞ þ qðPÞ 6 Q þ qrðpÞ . GENROUTE uses bounding
capacity inequalities are: functions, based on the q-path relaxation, to limit the size of the
X path set P, by avoiding the generation of paths that cannot be in
n‘ P dqðSÞ=Q e; 8S 2 S: ð14Þ any optimal CVRP solution and an algorithm for generating the
‘2RðSÞ routes which avoids the enumeration of all path pairs.

2.2.2.2. Clique inequalities. Let H ¼ ðR; EÞ be the conflict graph 2.2.3. The exact algorithm of Baldacci et al. (in press)
where each vertex corresponds to a route and the edge set E con- Baldacci et al. (in press) extended the exact method of Baldacci
tains an edge f‘; ‘0 g; 8‘; ‘0 2 R; ‘ < ‘0 , such that R‘ \ R‘0 –f0g. Let C be et al. (2008) to solve the CVRP and VRPTW by means of the follow-
the set of all cliques of H. ing improvements.
Clique inequalities are:
X 1. ng-route relaxation. The ng-route relaxation is a new state-
n‘ 6 1; 8C 2 C: ð15Þ space relaxation that improves the q-path and t-path relax-
‘2C
ations proposed for the CVRP and VRPTW, respectively. This
Let SP0 denote the problem obtained by adding strengthened relaxation provides ng-routes, that are used as an alternative
capacity and clique inequalities to problem SP, and let LSP0 denote to q-routes in procedure H1 and in computing better bounding
the LP-relaxation of SP. Let u = (u0, u1, . . . , un) be the vector of dual functions used in procedure GENROUTE. The ng-path relaxation
variables associated with constraints (8) and (9), where u0 6 0 consists of partitioning the set of all possible q-paths ending at
and ui 2 R, i 2 V. Moreover, let vS P 0, S 2 S, and gC 6 0, C 2 C, be vertex i according to a mapping function that associates with
the dual variables of constraints (14) and (15), respectively. each q-path a subset NG of the vertices visited that depends
The exact method of Baldacci et al. (2008) can be described as on the order in which they are visited.
follows. Let Ni # V be a set of selected customers for vertex i (according
to some criterion) such that Ni3i and jNij 6 D(Ni), where D(Ni) is
1. Solve relaxation LSP0 , and let (u0 , v0 , g0 ) be an optimal dual solu- a parameter. For a given path P = (0, i1, . . . , ik), let P(P) be the
tion of LSP0 of cost z0 . subset of V(P) containing customer ik and every customer ir,
2. Define the reduced problem SP c resulting from SP0 as follows: (i) r = 1, . . . , k  1, of P that belongs to all sets N irþ1 ; . . . ; N ik . An
replace the route set R with the largest subset R b # R such that ng-path (NG, q, i) is a nonnecessarily elementary path
b where c0 is the reduced cost of route ‘ with
c0‘ < zUB  z0 ; ‘ 2 R, P = (0, i1, . . . , ik1, ik) starting from the depot, visiting a subset of
‘
4 R. Baldacci et al. / European Journal of Operational Research 218 (2012) 1–6

customers of total demand equal to q such that NG = P(P), end- BCM was not able to solve to optimality 3 instances solved by
ing at customer ik, and such that ik R P(P0 ), where FLL and by BMR. Table 1 indicates that the lower bounds of BCM
P0 = (0, i1, . . . , ik1). Let f(NG, q, i) the cost of a least-cost ng-path and BMR are on average superior to the lower bounds of FLL in
(NG, q, i). Any (NG, q, 0)-path is called ng-route. Functions all classes of instances considered. Notice that FLL solved 78 in-
f(NG, q, i) can be computed using DP recursions. stances to optimality, but 29 of them were solved using the BC of
2. Subset-Row (SR3) inequalities. Jepsen et al. (2008) introduced Lysgaard et al. (2004). Table 1 indicates that BMR is, on average
Subset-Row (SR3) inequalities for the VRPTW. SR3s represent a faster than the other methods.
subset of clique inequalities. Let C # fC # V : jCj ¼ 3g be the sub-
set of all customer triplets and RðCÞ be the subset of routes 3. The vehicle routing problem with time windows (VRPTW)
serving at least two customers in C. Subset-Row (SR3) inequal-
ities are: The VRPTW is defined on a complete digraph G = (V0 , A), where
X V0 = {0, 1, . . . , n} is the vertex set and A is the arc set. Associated with
n‘ 6 1; 8C 2 C: ð16Þ
‘2RðCÞ
each arc (i, j) included in A is a travel cost dij and a travel time tij > 0,
where tij includes the service time at vertex i. It is assumed that
3. Weak Subset-Row (WSR3) inequalities. WSR3s are a weaker matrices dij and tij satisfy the triangle inequality. Associated with
version of SR3s. Given C 2 C, the route set RðCÞ contains only each vertex i included in V is a demand qi and a time window [ei, li],
those routes traversing at least an edge {i, j} with i, j 2 C. where ei and li represent the earliest and latest time to visit vertex
WSR3 dual variables can be easily considered in solving the i. The time windows are assumed hard.
pricing problem and in computing the bounding functions used A fleet of m identical vehicles of capacity Q stationed at the de-
by GENROUTE. The SR3s and, thus, the WSR3s are separated by pot has to fulfill customer demands. A vehicle route
complete enumeration. Both SR3s and WSR3s are used instead R = (0, i1, . . . , ir, 0), with r P 1, is a simple circuit in G, passing
of clique inequalities. through the depot, visiting vertices V(R) = {0, i1, . . . , ir}, V(R) # V0 ,
4. A new pricing strategy involving multiple dual solutions. and such that the following holds:
Baldacci et al. (in press) propose a new algorithm, more efficient
than GENROUTE, to generate routes of negative reduced costs in (i) The total demand of visited customers does not exceed the
H3 that also uses the dual solution achieved by H2 to eliminate vehicle capacity Q.
routes that cannot be in any optimal solution. Such improve- (ii) The vehicle leaves the depot 0 at time e0, visits each cus-
ment stabilizes the CCG procedure based on the simplex and tomer in V(R) within its time window, and returns to the
improves the final lower bound achieved by H3. depot before l0.
(iii) If the vehicle arrives at i 2 V(R) before ei, the service is
2.3. Computational comparison of exact algorithms for the CVRP delayed to time ei.

In this section, we report a computational comparison of the re- The cost of route R is equal to the sum of the travel costs of the
sults obtained by Lysgaard et al. (2004) (hereafter called LLE), arc set, A(R), traversed by route R.
Fukasawa et al. (2006) (FLL) and the two algorithms of Baldacci The VRPTW consists of designing at most m routes of minimum
et al. (2008) (BCM) and Baldacci et al. (in press) (BMR), on six total cost such that each customer is visited exactly once. As the
well-known classes of CVRP instances from the literature, called VRPTW reduces to the CVRP if ei = 0 and li = 1, "i 2 V0 , the VRPTW
A, B, E, F, M and P, and available at http://branchandcut.org/VRP/ is NP-hard. Indeed, it is strongly NP-complete to find a feasible
data. solution for the VRPTW if m < n.
The algorithm of BMR was run on an IBM Intel Xeon X7350 Ser- Several exact algorithms have been presented for the VRPTW. A
ver (2.93 GHz–16 GB of RAM). According to SPEC (http://www.spe- review of the exact methods up to 2002 is reported in Cordeau
c.org/benchmarks.html), BMR machine is three times faster than et al. (2002) and Kallehauge (2008). The best exact methods re-
the Pentium 4 2.6 GHz PC of BCM and the Pentium 4 2.4 GHz PC cently published on the VRPTW are based on model SP, described
of FLL and 10 times faster than the Intel Celeron 700 MHz PC of LLE. in Section 2.2, where the route set R contains any least-cost route
Table 1 reports, for each class, the name (Class), the number of in- satisfying time windows constraints.
stances (NP) and, for each exact method, the number of instances Model SP can be strengthened by any valid inequality studied
solved to optimality (Opt), the average percentage deviation of the for the CVRP, as discussed in Section 2, and by the k-path inequal-
lower bound (%LB) and the average computing time in seconds ities introduced by Kohl et al. (1999). The Kohl et al. strategy can be
(Time). For FLL, column OptBC reports the number of instances solved defined as follows.
to optimality by using the BC of Lysgaard et al., and column OptBCP Let k(S) be the minimum number of routes needed to service
reports the number of instances solved by their BCP. The last two the customer subset S 2 S. Kohl et al. (1999) called k-path inequal-
lines of Table 1 report the total number of instances solved by each ities the following generalization of the rounded capacity
method and the averages of lower bounds and computing times. constraints:

Table 1
Computational comparison of the exact methods for the CVRP.

Class BMR BCM FLL LLE

NP Opt %LB Time Opt %LB Time Opt OptBCP OptBC %LB Time Opt %LB Time
A 22 22 99.9 30 22 99.8 118 22 20 2 99.2 1961 15 97.9 6638
B 20 20 99.9 67 20 99.8 417 20 6 14 99.5 4763 19 99.4 8178
E-M 12 9 99.8 303 8 99.4 1025 9 7 2 98.9 126,987 3 97.7 39,592
F 3 2 100.0 164 3 0 3 99.9 2398 3 99.9 1046
P 24 24 99.8 85 22 99.7 187 24 16 8 99.2 2892 16 97.7 11,219
Avg 99.9 92 99.7 323 99.3 17,409 98.4 9935
Tot 81 77 72 78 49 29 56
 R. Baldacci et al. / European Journal of Operational Research 218 (2012) 1–6 5

Table 2 mon instances with 25, 50, and 100 customers. Nonetheless, this
Computational comparison of the exact methods for the VRPTW. method fails to solve several instances with 25 customers.
Class n NP Solved Time Algorithms based on the computation of elementary routes
BMR JPSP DHL BMR JPSP DHL were proposed by Feillet et al. (2004), Danna and Le Pape (2005)
and Chabrier (2006). Righini et al. (2006) proposed a DP method,
C2 50 8 8 7 n.a. 8 79 n.a.
RC2 50 8 8 7 n.a. 27 268 n.a.
called decremental state-space algorithm. This algorithm solves
R2 50 11 11 9 n.a. 124 7086 n.a. the pricing problem by forcing the routes to visit the customers
C1 100 9 9 9 9 25 468 18 of a selected subset V # V at most once. The least-cost route is
RC1 100 8 8 8 8 276 11,004 2150 computed using a label-setting method that expands both forward
R1 100 12 12 12 12 251 27,412 2327
and backward paths from the depot and connects routes in the
C2 100 8 8 7 8 40 2795 2093
RC2 100 8 8 5 6 3767 3204 15,394 middle.
R2 100 11 10 4 8 28,680 35,292 63,068 A significant contribution has been given by Jepsen et al. (2008).
They extended the BCP framework by adding the SR3 inequalities
(described in Section 2.2.3) to the SP master problem. The SR3
X inequalities provide better lower bounds but increase the complex-
b‘ ðSÞn‘ P kðSÞ; 8S 2 S; ð17Þ
‘2R ity of the pricing problem. The pricing problem is solved to optimal-
ity by using standard label setting techniques where each SR3
where b‘(S) = j{(i, j) 2 A(R‘) : i 2 V0 nS, j 2 S}j. These inequalities can be inequality is represented by an additional state variable. To improve
lifted by defining b‘(S) = 1 if V(R‘) \ S – ;, and b‘(S) = 0 otherwise. the performance of the label-setting algorithm, they introduced a
The separation algorithm consists of solving a VRPTW where modified dominance criterion that handles the reduced cost calcula-
only the customers in S need to be serviced and the objective is tion in a reasonable way. Moreover, to reduce the computing time,
to minimize the number of vehicles used. For this reason, most they attempted to solve the pricing problem heuristically. If no route
researchers considered the 2-path inequalities that consist of find- with negative reduced cost is found, then the pricing is solved to
ing any subset S 2 S, such that dq(S)/Qe = 1 and in solving a TSP optimality. The computational results indicate that the algorithm
with time windows (TSPTW) for the set S [ {0}. When the TSPTW of Jepsen et al. (2008) outperforms those of Irnich and Villeneuve
is infeasible, then k(S) = 2 provides a valid 2-path inequality for S. (2006) and Chabrier (2006) and solves eight previously unsolved in-
Other valid inequalities, related to k-path inequalities and called stances. However, the algorithm did not succeed in solving four in-
Reachability cuts, have been investigated by Lysgaard (2006). stances previously solved by different authors.
The exact algorithms for solving the resulting SP model of the The BCP of Jepsen et al. (2008) was improved by Desaulniers
VRPTW use CCG methods for computing the lower bound and et al. (2008) by adding both SR3 and generalized k-path inequali-
either branch-and-price (BP) or BCP algorithms to find an optimal ties and using a tabu search heuristic, before using DP, to rapidly
integer solution. The key component of these algorithms is the generate negative reduced cost routes. Their method outperforms
method for solving the pricing problem. This algorithm consists all other algorithms, remarkably decreasing the computational
of finding a number of VRPTW routes of negative reduced cost with time on Solomon instances with 100 customers, and solving 5 of
respect to the duals of the SP constraints (8), (9) and of the differ- the 10 open Solomon instances. The authors do not report any evi-
ent inequalities separated during the previous iterations. This dence of the performance of their method in solving the Solomon
problem is solved using different DP strategies to find either non- instances with 25 and 50 customers.
elementary or elementary routes. The exact method of Baldacci et al. (in press) described in Sec-
The main difficulty faced by both BP and BCP algorithms is how tion 2.2.3 for the CVRP also solves the VRPTW by simply using dif-
to take into account the duals of the different cuts in solving the ferent route relaxations to consider the time window constraints.
pricing problem. The duals of inequalities (8), (9) (say u0 and These route relaxations are based on the (t, i)-path and the
ui 2 R, i 2 V) and the duals of inequalities defined according to (NG, t, i)-path relaxations.
expressions (13) (say, wt, t 2 F ) can be associated with the arcs A (t, i)-path is a nonnecessarily elementary path, without 2-ver-
of a route. Thus, it is easy to observe that the reduced cost c0‘ of a tex loops, starting from the depot at time e0, visiting a set of cus-
P 0
route ‘ 2 R is equal to Þ dij by defining the modified arc
ði;jÞ2AðR‘P tomers (even more than once) within their time windows, and
0 0
costs dij as dij ¼ dij  2 ðui þ uj Þ  t2F atij wt .
1
ending at vertex i at time ei 6 t 6 li.
It is much more complex to solve the pricing problem when Let f(t, i) be the cost of a least-cost (t, i)-path. A t-route, for a gi-
considering set partitioning inequalities such as clique, subset ven e0 6 t 6 l0, is defined as the (t, 0)-path of cost f(t, 0). Functions
row, lifted versions of rounded capacity and k-path inequalities. f(t, i) can be computed using DP (see Christofides et al., 1981).
A possible way is to associate a binary state variable for each of A (NG, t, i)-path is a nonnecessarily elementary path
these inequalities in the DP algorithm where each variable repre- P = (0, i1, . . . , ik1, ik) starting from the depot at time e0, visiting a
sents the dual of the associated inequality in the reduced cost of subset of customers (even more than once) within their time win-
the emerging path. Unfortunately, the size of the resulting state- dows such that NG = P(P), ending at customer ik at time eik 6 t 6 lik ,
space graph grows and the resulting DP algorithm becomes com- and such that ik R P(P0 ), where P0 = (0, i1, . . . , ik1).
putationally slow. Let f(NG, t, i) be the cost of a least-cost (NG, t, i)-path. The ng-
Kohl et al. (1999) described a BP algorithm which improves the routes for the VRPTW corresponds to the (NG, t, 0)-path e0 6 t 6 l0,
BP of Desrochers et al. (1992) by adding 2-path inequalities to LP- NG # V0 , of cost f(NG, t, 0). Capacity constraints are ignored in both
relaxation of the SP formulation and by using nonelementary (t, i)-path and (NG, t, 0)-path relaxations.
routes without 2-vertex loops in solving the pricing problem. Ir-
nich and Villeneuve (2006) improved the BP of Kohl et al. (1999) 3.1. Computational comparison of exact algorithms for the VRPTW
by using a sophisticated labelling algorithm to forbid routes having
k-vertex loops with k P 3. The numerical experiments on some In this section, we report the computational comparison of the
hard VRPTW instances generated by Solomon (1987) show that exact methods of Jepsen et al. (2008) (hereafter called JPSP),
k-cycle elimination, with k P 3, can substantially improve the low- Desaulniers et al. (2008) (DHL) and Baldacci et al. (in press)
er bounds. The new labelling algorithm has proven to be a key (BMR) on six classes of instances (C1, RC1, R1, C2, RC2 and R2)
ingredient for solving to optimality more than 15 unsolved Solo- (see Solomon, 1987).
6 R. Baldacci et al. / European Journal of Operational Research 218 (2012) 1–6

BMR was run on an IBM Intel Xeon X7350 Server (2.93 GHz– Christofides, N., Mingozzi, A., Toth, P., 1981. Exact algorithms for the vehicle routing
problem based on spanning tree and shortest path relaxations. Mathematical
16 GB of RAM). According to SPEC (http://www.spec.org/bench-
Programming 20 (1), 255–282.
marks.html), the machine used by BMR is three times faster than Cordeau, J.-F., Desaulniers, G., Desrosiers, J., Solomon, M., Soumis, F., 2002. The vrp
the Intel Pentium 4 3.0-GHz PC of JPSP and twice as fast as the Li- with time windows. In: Toth, P., Vigo, D. (Eds.), The Vehicle Routing Problem,
nux PC Dual Core AMD Opteron at 2.6 GHz of DHL. For each class, vol. 9. SIAM Monographs on Discrete Mathematics and Applications,
Philadelphia, pp. 157–193.
Table 2 reports the class name (Class), the number of customers (n), Danna, E., Le Pape, C., 2005. Column generation. In: Desaulniers, G., Desrosiers, J.,
the number of instances (NP), the number of instances solved by Solomon, M.M. (Eds.), Accelerating Branch-and-price with Local Search: A Case
each of the three methods (Solved) and the average computing Study on the Vehicle Routing Problem with Time Windows. Springer, New York,
pp. 90–130.
time in seconds (Time) (n.a. means data is not available). Desaulniers, G., Lessard, F., Hadjar, A., 2008. Tabu search, partial elementarity, and
The table shows that BMR solved all but one Solomon instance generalized k-path inequalities for the vehicle routing problem with time
and closed four open instances. Moreover, BMR outperforms JPSP windows. Transportation Science 42 (3), 387–404.
Desrochers, M., Desrosiers, J., Solomon, M., 1992. A new optimization algorithm for
and DHL: all instances solved by the other methods were solved the vehicle-routing problem with time windows. Operations Research 40 (2),
by BMR and the average time is significantly lower. Instances of 342–354.
classes C2, RC2 and R2 involving 100 customers are more difficult Feillet, D., Dejax, P., Gendreau, M., Gueguen, C., 2004. An exact algorithm for the
elementary shortest path problem with resource constraints: Application to
than instances of classes C1, RC1 and R1 of the same dimension as some vehicle routing problems. Networks 44, 216–229.
they feature wide time windows. Finke, G., Claus, A., Gunn, E., 1984. A two-commodity network flow approach to the
traveling salesman problem. Congress. Numerantium 41, 167–178.
Fukasawa, R., Longo, H., Lysgaard, J., Poggi de Aragão, M., Reis, M., Uchoa, E.,
4. Conclusions Werneck, R., 2006. Robust branch-and-cut-and-price for the capacitated vehicle
routing problem. Mathematical Programming Series A 106, 491–511.
Grötschel, M., Padberg, M., 1985. Polyhedral theory. In: Lawler, E., Lenstra, J.,
In the last decade, some innovative exact approaches for vehicle Rinnooy Kan, A., Shmoys, D. (Eds.), The Traveling Salesman Problem: A Guided
routing problems have been proposed, producing a significant Tour of Combinatorial Optimization. John Wiley & Sons Ltd., Chichester, pp.
231–305.
improvement on the size of the instances that can be solved to
Irnich, S., Villeneuve, D., 2006. The Shortest-Path problem with resource constraints
optimality. Indeed, these algorithms have brought more than one and k-Cycle elimination for k>=3. INFORMS Journal on Computing 18 (3), 391–
hundred the number of customers that may be handled. The key 406.
factor of the success of these approaches is the effective combina- Jepsen, M., Petersen, B., Spoorendonk, S., Pisinger, D., 2008. Subset-row inequalities
applied to the vehicle-routing problem with time windows. Operations
tion of the set partitioning formulation with families of cuts into Research 56 (2), 497–511.
column generation based algorithms. This approach has signifi- Kallehauge, B., 2008. Formulations and exact algorithms for the vehicle routing
cantly improved the quality of the lower bounds that are now very problem with time windows. Computers & Operations Research 35 (7), 2307–
2330.
close to the optimal solution values. Furthermore, set partitioning Kara, I., Laporte, G., Bektas, T., 2004. A note on the lifted Miller–Tucker–Zemlin
based approaches proved quite general as they are easily able to subtour elimination constraints for the capacitated vehicle routing problem.
incorporate additional constraints arising in practical applications. European Journal of Operational Research 158 (3), 793–795.
Kohl, N., Desrosiers, J., Madsen, O.B.G., Solomon, M.M., Soumis, F., 1999. 2-path cuts
for the vehicle routing problem with time windows. Transportation Science 33
References (1), 101–116.
Letchford, A., Salazar González, J., 2006. Projection results for vehicle routing.
Mathematical Programming Series B 105 (2–3), 251–274.
Achuthan, N.R., Caccetta, L., 1991. Integer linear programming formulation for a Lysgaard, J., 2006. Reachability cuts for the vehicle routing problem with time
vehicle routing problem. European Journal of Operational Research 52 (1), 86– windows. European Journal of Operational Research 175 (1), 210–223.
89. Lysgaard, J., Letchford, A.N., Eglese, R.W., 2004. A new branch-and-cut algorithm for
Augerat, P., Belenguer, J.M., Benavent, E., Corberán, A., Naddef, D., Rinaldi, G., 1995. the capacitated vehicle routing problem. Mathematical Programming Series A
Computational results with a branch and cut code for the capacitated vehicle 100, 423–445.
routing problem. Tech. Rep. 1 RR949-M, ARTEMIS-IMAG, Grenoble France. Naddef, D., Rinaldi, G., 2002. Branch-and-cut algorithms for the capacitated VRP. In:
Baldacci, R., Hadjiconstantinou, E., Mingozzi, A., 2004. An exact algorithm for the Toth, P., Vigo, D. (Eds.), The Vehicle Routing Problem, vol. 9. SIAM Monographs
capacitated vehicle routing problem based on a two-commodity network flow on Discrete Mathematics and Applications, pp. 53–81.
formulation. Operations Research 52, 723–738. Ralphs, T., Kopman, L., Pulleyblank, W., Trotter, L., 2003. On the capacitated vehicle
Baldacci, R., Christofides, N., Mingozzi, A., 2008. An exact algorithm for the vehicle routing problem. Mathematical Programming Series B 94, 343–359.
routing problem based on the set partitioning formulation with additional cuts. Righini, G., Salani, M., 2006. Symmetry helps: Bounded bi-directional dynamic
Mathematical Programming Series A 115 (2), 351–385. programming for the elementary shortest path problem with resource
Baldacci, R., Mingozzi, A., Roberti, R., in press. New route relaxation and pricing constraints. Discrete Optimization 3 (3), 255–273.
strategies for the vehicle routing problem. Operations Research. Solomon, M.M., 1987. Algorithms for the vehicle routing and scheduling problem
Balinski, M., Quandt, R., 1964. On an integer program for a delivery problem. with time window constraints. Operations Research 35, 234–265.
Operations Research 12, 300–304. Toth, P., Vigo, D. (Eds.), 2002. The Vehicle Routing Problem. Monographs on Discrete
Chabrier, A., 2006. Vehicle routing problem with elementary shortest path based Mathematics and Applications. SIAM, Philadelphia, PA.
column generation. Computers & Operations Research 33, 2972–2990.