Transportation Research Part E 41 (2005) 131–144

www.elsevier.com/locate/tre

A vehicle routing problem with backhauls and

time windows: a guided local search solution
a,1 b,*,1
Yingjie Zhong , Michael H. Cole
a
Motorola, Inc., Schaumburg, IL 60196, USA
b
Department of Mechanical and Industrial Engineering, Montana State University, Bozeman, MT 59717, USA
Received 20 February 2003; received in revised form 16 October 2003; accepted 9 December 2003

Abstract

This paper presents a guided local search heuristic to solve a vehicle routing problem with backhauls and
time windows. The VRPBTW with and without customer precedence are both considered. Customer
precedence requires that all linehaul customers be visited before any backhaul customer. The basic ap-
proach is to construct an initial infeasible solution and then use a guided local search to improve the
solution feasibility and quality. A new technique––section planning––is used to enhance the feasibility.
Computational results show that the new heuristic can solve problems in which customers are distributed
normally or in clusters. Some of the results are better than the best solutions that have appeared in the
literature.

2004 Elsevier Ltd. All rights reserved.

Keywords: Vehicle routing; Time windows; Metaheuristic

1. Introduction

The Vehicle Routing Problem with Backhauls and Time Windows (VRPBTW) can be stated as
follows: A set of customers with deterministic demand, deterministic time windows, and certain
types of service requirements (pick up and/or delivery) must be served by a homogeneous fleet of
vehicles with fixed capacity starting from and ending at a central depot, which also has a certain
time horizon. For each service requirement (pickup or delivery), each customer must be assigned
exactly once to a vehicle. The customers are classified into two groups: linehaul customers whose

*
Corresponding author. Tel.: +1-406-994-1399; fax: +1-406-994-6292.
E-mail address: mcole@ie.montana.edu (M.H. Cole).
1
Formerly at the University of Arkansas.

1366-5545/$ - see front matter
2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.tre.2003.12.003
132 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144

demand needs to be delivered, and backhaul customers whose demand needs to be picked up (if a
customer requires both pickup and delivery, it is modeled as two separate customers). In this
paper, a vehicle is allowed to arrive at a customer before the relevant time window, but cannot
service the customer until the time window opens. The VRPBTW has two objectives: minimize the
number of routes and minimize the sum total distance of all the routes. The VRPBTW can be
classified into two cases. In the VRPBTW with customer precedence, all of the linehaul customers
must be served before any backhaul customers in each route. In the VRPBTW without customer
precedence, linehaul and backhaul customers can be interspersed on a given route.
This research develops a heuristic approach to the VRPBTW. Optimal methods are not
appropriate since the traditional VRP is an NP-hard problem and, in general, adding time win-
dows and backhauls does not simplify the problem. The new heuristic is a cluster-first, route-
second algorithm, with most effort spent on routing. The two phases are

• Phase 1––Use an adapted sweep algorithm (Fisher and Jaikumar, 1981) to generate an initial
infeasible solution. Use GLSA, a guided local search heuristic (based partly on Voudouris and
Tsang, 1999) to manage 2-opt, 1-move, and 1-exchange (a new method) to improve the initial
solution.
• Phase 2––Enhance feasibility using a new technique called section planning. This technique
inserts new routes until feasibility is attained, and arranges customers within routes to reduce
travel distance. This phase is iterative with feasibility constraints being ‘‘soft’’ in early iterations
and ‘‘hard’’ in later iterations.

The new heuristic was originally developed for the VRPBTW without customer precedence.
After promising results for problems without customer precedence, we extended the heuristic to
consider the VRPBTW with customer precedence.

2. Literature review

A comprehensive review of the VRP can be found in Bodin et al. (1983) and Ball et al. (1995).
Useful techniques for the general VRP are outlined in Golden and Assad (1988) and Aarts and
Lenstra (1997). Reeves (1993) covers modern techniques such as simulated annealing, tabu search,
and genetic algorithms. Various heuristic methods may be found in the literature for both the
VRPTW (Potvin et al., 1996b; Potvin and Bengio, 1996; Russell, 1995; Chiang and Russell, 1997)
and the VRPB (Casco et al., 1988; Deif and Bodin, 1984; Golden et al., 1985; Goetschalckx and
Jacobs-Blecha, 1989; Jacobs-Blecha and Goetschalckx, 1993; Toth and Vigo, 1996).
Several papers deal specifically with the VRPBTW. Gelinas et al. (1995) propose a new
branching strategy for branch-and-bound approaches based on column generation. This algo-
rithm finds optimal solutions to different test problems with up to 100 customers. Potvin et al.
(1996a) design a genetic algorithm to identify an ordering of customers that produces good routes.
Duhamel et al. (1997) design a tabu search heuristic for the VRPBTW with customer precedence.
First a feasible solution is constructed by an adapted version of Solomon’s (1987) insertion
heuristic. Then tabu search and local search and improvement algorithms are used to improve the
solution. All three of these papers detail experiments on problems with customers distributed
 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144 133

normally within the service area. Kontoravdis and Bard (1995) describe a greedy randomized
adaptive search procedure (GRASP) for VRPTW. This algorithm is also capable of solving
VRPBTW without customer precedence. The authors report experimental results for problems
with customers distributed in clusters. Cheung and Hang (2003) develop label matching algo-
rithms for solving the VRPBTW. Their heuristic approach can handle the addition of complex
real-world constraints, such as vehicles of different capacities and penalties for vehicles that arrive
early.
Voudouris and Tsang (1999) develop the GLS (guided local search) metaheuristic and apply it
to the traveling salesman problem. The GLS penalizes solution features (e.g., active arcs) during
each iteration based on the value of a utility function. The penalty acts as a disturbance to an
augmented objective function which is adjusted during each iteration. This reduces the chance
that the solution procedure will get stuck in a local optimum.

3. Algorithm description

3.1. Solution feasibility consideration

In recent years, VRP researchers investigating additional constraints, such as time windows or
backhauls, have concentrated on postponing solution feasibility (Chiang and Russell, 1997).
Logically, this concern is a perfect match for simulated annealing. However, our consideration of
solution infeasibility is different from that of some researchers. Often, infeasibility is taken to
mean that there exist unrouted customers, but that all customers in routes are in feasible routes. In
this paper, the routes themselves may be infeasible. The local search heuristics in Phase 1 enhance
the feasibility of the solution. The section planning technique in Phase 2 enforces feasibility by
inserting a new route into the solution during each iteration. In the most extreme case, the
heuristic would generate a separate route for each customer. Thus, a feasible solution is guar-
anteed.

3.2. Guided local search approach (GLSA)

This subsection develops an augmented objective function and penalty function for the
VRPBTW.

3.2.1. Augmented objective function

The augmented VRPBTW objective function comprises the usual VRP objective function,
Lagrangean multipliers that penalize constraint violations, and a disturbance item derived from
the GLS. Eq. (1) is the augmented objective function for the VRPBTW without customer pre-
cedence. The augmented objective function for the VRPBTW with customer precedence is
identical except for an added term related to precedence violations, kp vp .
X
hðSÞ ¼ gðSÞ þ kt vt þ kc vc þ k pij Iij ðSÞ ð1Þ
ij2S
134 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144

gðSÞ typical VRP cost function

vt violation of due-time constraints
vc violation of capacity constraints
kt and kc Lagrangean multipliers
k multiplier indicating the magnitude of the total disturbance due to arc penalties
pij penalty associated with arc ij
Iij ðSÞ 1 if arc ij is in solution S, 0 otherwise

Parameters kt and kc should increase after each iteration in order to more strongly encourage
feasibility. The value of pij  Iij ðSÞ should eventually decrease to zero. Parameter k should decrease
after each iteration, as well. Note that the augmented objective function does not penalize vio-
lations of ready-time constraints. Although customers can only be served within their time win-
dows, vehicles are allowed to arrive early and wait without penalty.

3.2.2. Penalty function

The penalty pij depends on three factors: the arc length, the arc time precedence, and a dynamic
utility function. An arc’s length dij is the travel distance between nodes i and j. An arc’s time
precedence tij is calculated as described in Fig. 1 (the hard-coded numbers are included to give a
sense of scale).
The initial value of each arc’s penalty, pij , is set equal to the arc’s time-precedence. It is updated
in each iteration as described in Fig. 2 (the hard-coded numbers are included to give a sense of
scale). If an arc develops a very large penalty, it is deleted from future consideration. This eases
the computational burden associated with evaluating the inclusion of the arc into future solutions.

3.3. Solution approach

This section introduces the solution approach for both the VRPBTW with precedence and the
VRPBTW without precedence. The algorithm has two phases.
Set t = dij / vehicle speed

If ( t + customer i ready time > customer j due time)

tij = 200

Else if (t + customer i ready time > customer j ready time)

tij = 0

Else
tij = 20
End if

If (t + customer i due time < customer j due time)

tij -= 20

Else if ( t + customer i due time > customer j due time)

tz += 20

End if

Fig. 1. Pseudocode for calculating time precedence of arc ij.

 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144 135

Set u = Iij(S) · (dij · tij) / (1 + p ij )

If u > 9
pij += 5
End if

if pij > 60
Set Iij = 0 for remainder of algorithm
End if

Fig. 2. Pseudocode for calculating penalty value of arc ij during each iteration.

3.3.1. Phase 1
Phase 1 has four stages:

1. Clustering and initial route construction.

2. Intra-route improvement: 2-opt.
3. Inter-route improvement: 1-move.
4. Inter-route improvement: 1-exchange.

3.3.1.1. Clustering and initial route construction. Clustering and initial route construction are
detailed in Fig. 3 (based, in part, on Fisher and Jaikumar, 1981).

3.3.1.2. Route improvement. This section describes how GLSA guides local search heuristics. The
composite heuristic improves the initial infeasible solution into a much more ordered structure. A
local search heuristic is designed for searching for a high quality solution in a shrunken solution

For i = 1 to NumberOfCustomers
Determine polar angle in relation to depot
Next i

For i = 1 to 50
Select random start angle from U[0,pi/2]
Select random cluster capacity from U[0.6,1.0]*vehicle capacity

Sweep customers into clusters

For j = 1 to NumberOfClusters
Divide cluster into two zones of equal polar angle
Assign each customer to the route, alternating between zones
Next j
Next i

Select initial solution from the 50 candidates such that the number
of routes is low, and violations of customer due times and
precedence are low

Fig. 3. Pseudocode for initial route construction.

136 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144

space––a neighborhood. In solving a complicated problem, a set of local search heuristics based
on different neighborhoods is necessary to increase the search efficiency and effectiveness. The
general idea is to use simple local search heuristics followed by more sophisticated local search
heuristics to yield a better solution quality within a reasonable run time.
This section discusses three local search heuristics that can be categorized into two classes:
intra-route exchange and inter-route exchange. 2-opt and 1-move are common arc exchange
heuristics that are frequently used in solving the VRP, while 1-exchange is a new heuristic de-
signed particularly for the VRPBTW.

3.3.1.3. Intra-route exchange: 2-opt. 2-opt (Croes, 1958) is a local search heuristic for the traveling
salesman problem (TSP) and other combinatorial problems. 2-opt is commonly used in the VRP
without time windows. Potvin et al. (1996b) develop 2-opt in order to maintain time window
feasibility when solving the VRPTW. In Phase 1, 2-opt is sufficient because maintaining solution
feasibility is not necessary. 2-opt is used to decrease the length, time precedence violation, and
customer precedence violations (for the VRPBTW with customer precedence) of each route. 2-opt
does not use the augmented VRPBTW objective function. The neighborhood of 2-opt is all pairs
of arcs in a particular route.

3.3.1.4. Inter-route exchange: 1-move, 1-exchange. Inter-route exchange involves the augmented
objective function. The GLS first sets the initial value of the Lagrangean multipliers, and then
performs 1-move and 1-exchange for a predetermined number of iterations.

3.3.1.5. 1-move. 1-move is the same heuristic as operators (1, 0) and (0, 1) proposed by Chiang and
Russell (1997). The 1-move procedure effectively improves the quality and feasibility of poor
solutions. 1-move deletes one node from a route and inserts it into another route. The evaluation
of 1-move is decided by the length and time precedence of the arc, and the time and capacity
violations change caused by 1-move. The neighborhood of 1-move is all the customers and any
potential customer positions for all pairs of routes in the current solution.

3.3.1.6. 1-exchange. The 1-exchange heuristic deletes two customers from two different routes and
inserts each customer into the best position in the other route. In Fig. 4, customer x in route 1
connects with arcs ðx1 ; xÞ and ðx; x2 Þ, and customer y in route 2 connects with arcs ðy1 ; yÞ and
ðy; y2 Þ.
1-exchange first deletes the two customers, x and y, from route 1 and route 2 respectively.
Then, customer x searches for a good position in route 2 and customer y searches for a good
position in route 1. The best combination of these two insertions (if the result has a better
objective function value than the original configuration) is the result of 1-exchange. The best
position for customer x in route 2 is after customer y2 and the best position for customer y in
route 1 is before customer x1 .
1-exchange intensifies the local search. The neighborhood of 1-exchange is larger than that of
1-move since each of the customers needs to find its best position in the other route. To reduce
the computational burden, if the length of the new arcs is more than twice that of the deleted arcs,
the candidate exchange is cancelled before extensive evaluation of time windows and capacity
constraints.
 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144 137

BEFORE AFTER

x
x
x2
x2 x1
x1 y2 y2
y
y
route 1 route 2 route 1 route 2

y1/depot y1/depot

Fig. 4. Before and after 1-exchange of nodes x and y.

3.3.2. Phase 2
Phase 2 is an iterative procedure that adds a user-specified number of new routes in an effort to
achieve feasibility (see Fig. 5). Each iteration has two main steps: route feasibility enhancement
and route improvement. The feasibility enhancement step deletes customers that create infeasi-
bilities and adds a new route comprising the deleted customers. Section Planning (described
below) is a key part of feasibility enhancement. The improvement step applies the 2-opt, 1-move,
and 1-exchange heuristics.

3.3.2.1. Section Planning for route feasibility enhancement. Section Planning divides each route
into several sections and tries to delete the customer in each section that contributes most to the
violation of due-time and capacity constraints (see Fig. 6). Considering a route section by section
is computationally easier than considering the route as a whole.

For k = 1 to NumberOfNewRoutes(a parameter)

For i = 1 to NumberOfRoutes
Delete precedence violating customers
Delete “big” violators of capacity and time windows
Do Section Planning
Next i
Add one new route comprising deleted customers
Update info

For j = 1 to NumberOfIterations (a parameter)

Do route improvement:
2-opt
1-move
1-exchange (every other iteration)
Next j
Next k

Fig. 5. Pseudocode for Phase 2.

138 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144

If a capacity-violating customer exists then

Divide the route into sections divided by the violating customer(s)

For each section

If (no customers violate due time in this section or next section)

Delete the linehaul or backhaul customer in the section that best

reduces the capacity violation

Else If (a customer violates due time in this section)

Let n = customer with largest due time violation in this section

If (find backhaul customer before n that best reduces capacity

violation)
Delete the backhaul customer
End If

Else If (a customer violates due time in the next section)

Let n = customer with largest due time violation in next section

If (find linehaul customer before n that best reduces capacity

violation)
Delete the linehaul customer
End If

End If

Next

Else If a ready-time-violating customer exists then

Divide the route into sections divided by the violating customer(s)

For each section

If (a customer violates due time in this section)

Let n = customer with largest due time violation in this section
Delete the customer before n that best reduces the due time violation
End If

Next

Else If a due-time-violating customer exists then

[treat entire route as a single section]

Let n = customer with largest due time violation in the entire route
Delete the customer before n that best reduces the due time violation

Else
This route is feasible

End If

Fig. 6. Pseudocode for section planning a route.

The algorithm first defines sections such that each section includes exactly one customer that
violates capacity restrictions. If any customer in the current section or following section also
 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144 139

violates due-time, the algorithm deletes a customer that alleviates both the capacity violation and
the due-time violation. Otherwise, the algorithm deletes a customer that alleviates only the
capacity violation. If the route has no capacity violations, the algorithm defines sections according
to ready-time or due-time violations, and deletes the single customer that best alleviates the worst
violation of due-time.

4. Computational experiments

The two versions of the GLSA heuristic (without customer precedence and with customer
precedence) were programmed in Microsoft Visual C++ 6.0 and executed on a 450 MHz PC.
GLSA was tested on problems sets derived from Solomon’s data sets R1, R2, C2, and RC2. In
Solomon’s original data sets, all the problems are Euclidean with a single service type. As de-
scribed by Kontoravdis and Bard (1995):

• Each data set contains between eight and twelve 100-node problems over a service area defined
on a 100 by 100 grid.
• For R1 and R2, the customers are located uniformly over the service area. Set C2 has clustered
customers and set RC2 has a combination of clustered and randomly placed customers.
• R1 has a short scheduling horizon and a vehicle capacity of 200 units; R2, C2 and RC2 have a
long scheduling horizon and vehicle capacity of 1000, 700, and 1000 units respectively.
• The time window constraints in problem set R1 allow only a small number of customers to be
served by each vehicle. The opposite is true for R2, C2, and RC2. In both cases, the vehicle
capacity constraints are fairly loose.

4.1. VRPBTW without customer precedence

For the VRPBTW without customer precedence, GLSA was tested on data sets developed by
Kontoravdis and Bard (1995). Kontoravdis constructed data sets MR2, MC2, and MRC2 by
designating each customer in Solomon’s R2, C2, and RC2 data sets as either linehaul or backhaul
with equal probability. Their experiments are restricted to derivatives of R2, C2, and RC2 because
‘‘time window constraints are dominant for R1, C1 and RC1, and thus the solution state space
would not be greatly affected after changing half of the customers to backhaul.’’ They reduced the
vehicle capacity to 250 units to ‘‘assure that the capacity constraint had the primary influence on
feasibility.’’
Tables 1–3 compare the performance of GLSA algorithm to the performance of the GRASP
algorithm developed by Kontoravdis and Bard (1995). Two different GLSA runs were per-
formed on each problem and the best solution was selected at the end. The tables show three
measures of performance: number of routes, total route distance, and CPU time (in seconds).
Note that the individual GRASP solution measures are not available, and are thus represented
by averages. The CPU times are not directly comparable, since computing power has increased
considerably since Kontoravdis and Bard implemented their algorithm on a Sun Sparc-10
workstation in 1995.
140 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144

Table 1
Results for MC2 data set
Problem GRASP GLSA
Routes Distance CPU time Routes Distance CPU time
MC201 – – – 5 763.88 57
MC202 – – – 4 1186.24 347
MC203 – – – 4 1096.31 323
MC204 – – – 4 885.73 431
MC205 – – – 5 781.70 169
MC206 – – – 5 860.74 352
MC207 – – – 5 792.96 476
MC208 – – – 5 859.92 393
Average 4 1094.94 130.1 4.625 903.56 319

Table 2
Results for MR2 data set
Problem GRASP GLSA
Routes Distance CPU time Routes Distance CPU time
MR201 – – – 4 1388.73 108
MR202 – – – 4 1198.99 422
MR203 – – – 4 988.92 544
MR204 – – – 4 858.32 450
MR205 – – – 4 1172.53 253
MR206 – – – 4 979.50 406
MR207 – – – 4 912.69 451
MR208 – – – 4 764.52 408
MR209 – – – 4 978.82 516
MR210 – – – 4 1061.36 557
MR211 – – – 4 878.81 692
Average 4 1168.53 122.3 4 1016.66 437

Table 3
Results for MRC2 data set
Problem GRASP GLSA
Routes Distance CPU time Routes Distance CPU time
MRC201 – – – 5 1498.90 73
MRC202 – – – 4 1539.41 493
MRC203 – – – 4 1303.48 713
MRC204 – – – 4 932.48 472
MRC205 – – – 4 1632.04 362
MRC206 – – – 4 1433.43 208
MRC207 – – – 4 1217.20 599
MRC208 – – – 4 1085.57 1134
Average 4.5 1496.91 135.7 4.125 1330.31 507
 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144 141

For data set MC2 (Table 1), GLSA generated more routes than GRASP, but did get a total
distance that is around 19% shorter. For set MR2 (Table 2), GLSA got the same number of routes
as GRASP, and saved about 15% in total route distance. For set MRC2 (Table 3), GLSA gen-
erated fewer routes and also reduced the total distance about 10%. So, for these three data sets,
GLSA outperforms GRASP with respect to route distance, and seems to perform at approxi-
mately the same level as GRASP with respect to number of routes.
GRASP was primarily designed to minimize the number of routes, so it is not overly surprising
that the new GLSA algorithm outperforms GRASP with respect to total route distance for all
three sample data sets. Whether GLSA’s additional computation time is worth the reduction in
total distance depends on the particular case. More interestingly, neither GLSA nor GRASP was
consistently better at minimizing the number of routes. This indicates that both algorithms should
be run if the primary goal is to minimize the total number of routes while secondarily minimizing
the total route distance.

4.2. VRPBTW with customer precedence

For the VRPBTW with customer precedence, GLSA was tested on data sets developed by
Gelinas et al. (1995). As discussed in Duhamel et al. (1997), Gelinas et al. modified the first five
problems of Solomon’s R1 data set by randomly choosing 10%, 30%, and 50% of the 100 clients
to be backhaul customers, leaving other attributes unchanged. They generated additional test
problems by considering only the first 25 and first 50 nodes. The total number of test problems for
the VRPBTW with customer precedence is 45.
Representative results are reported in Tables 4 and 5. Five different runs were performed on
each problem and the best solution was selected at the end. Potvin et al. (1996a) results with a
genetic algorithm GA (run on a Sun Sparc-10 workstation) are also reported. The tables show
three measures of performance: number of routes, total route distance, and CPU time (in sec-
onds). Note that the computation times are not directly comparable since computing power has
increased significantly since 1996.

Table 4
Results for problems derived from R101
Number of Backhaul% Optimal Potvin’s GA GLSA
customers solution Routes Distance CPU time Routes Distance CPU time
25 10 643.4 9 643.4 5.9 9 665.24 2
30 711.1 10 721.8 5.6 10 723.19 4
50 674.5 10 682.3 5.6 10 684.52 3
50 10 1122.3 14 1138.1 16.2 14 1178.73 6
30 1191.5 16 1192.7 16.9 16 1332.44 8
50 1168.6 16 1183.9 17.5 16 1237.65 6
100 10 1767.9 23 1815.0 169.6 24 1848.04 25
30 1877.6 23 1896.6 163.3 24 2034.61 27
50 1895.1 24 1905.9 211.6 25 2057.05 27
142 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144

Table 5
Results for problems derived from R105
Number of Backhaul% Optimal Potvin’s GA GLSA
customers solution
Routes Distance CPU time Routes Distance CPU time
25 10 565.1 7 565.1 5.6 7 577.85 2
30 623.5 8 630.2 6.8 8 635.62 6
50 591.1 7 592.1 5.4 7 618.20 3
50 10 970.6 10 1002.5 14.9 10 1041.89 2
30 1007.5 11 1047.8 15.2 10 1089.00 8
50 993.4 11 1018.0 15.4 11 1048.87 3
100 10 – 17 1621.0 191.4 17 1590.54 34
30 – 16 1652.8 210.6 17 1667.92 55
50 – 18 1706.7 160.4 19 1699.88 64

Table 4 shows results for problems derived from the R101 data set. For this problem set, the
GLSA solutions were typically a few percent worse than the GA solutions, which were a few
percent worse than optimal. The results were similar for problems derived from data sets R102,
R103, and R104.
Table 5 shows results for problems derived from R105 data set. GLSA underperformed Pot-
vin’s GA except for two cases. For the 100 customer/10% backhaul case, GLSA yielded a solution
with the same number of routes but 1.9% lower total distance. For the 100 customer/50%
backhaul case, GLSA yielded a solution with one more route but 0.4% lower total distance.
In contrast to the case of the VRPBTW problem without customer precedence, GLSA did not
perform very well compared to existing algorithms for the VRPBTW with customer precedence.
However, GLSA did outperform the GA algorithm in one instance. Additional research will be
needed before adding GLSA to the toolkit of problem-solving techniques for the VRPBTW with
customer precedence.

5. Conclusions

GLSA is a simple heuristic for solving the vehicle routing problem with backhauls and time
windows (VRPBTW). The first phase of the GLSA heuristic uses guided local search to improve
routes, but still allows time and capacity violations. The second phase of GLSA uses a new
technique, Section Planning, to eliminate time and capacity violations by selectively moving
problematic customers to newly created routes. We compared GLSA to GRASP, developed by
Kontoravdis and Bard (1995) and to GA, developed by Potvin et al. (1996a).
GLSA was originally developed for the VRPBTW without customer precedence. On a series of
test data sets, GLSA provided solutions with 10–19% lower total distance, but occasionally more
routes, than GRASP. GLSA requires more computational effort than GRASP, but the additional
effort might be a reasonable tradeoff for better solution quality. Since the computational burden
of GLSA and other modern heuristics is relatively small and since no heuristic is guaranteed to be
 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144 143

best for a particular data set, researchers and practitioners are advised to solve problems using
several different heuristics.
We extended GLSA to consider the VRPBTW with customer precedence. GLSA did not
perform as well as GA on this class of VRPBTW, although its solution values were generally
within 5% of published best results. This indicates that customer precedence is a major distin-
guishing factor of VRPBTW problems, and that problems with customer precedence might re-
quire very different solution approaches than problems without customer precedence.
In summary, GLSA makes the following contributions to the VRPBTW literature:

• GLSA introduces the Section Planning technique for eliminating time and capacity violations
within a route.
• GLSA outperforms the previous GRASP algorithm in minimizing the total route distance for
VRPBTW problems without customer precedence. GLSA does not consistently underperform
or outperform GRASP in minimizing the number of routes. This indicates that neither algo-
rithm should be used alone to solve such VRPBTW problems.
• GLSA generally underperforms the existing GA algorithm for VRPBTW problems with cus-
tomer precedence. However, GLSA occasionally does find a better solution than GA. After
additional development, GLSA might be a useful algorithm to run in conjunction with GA.

Acknowledgements

The authors would like to thank the anonymous referees for their helpful comments. This
research was supported by the Mack-Blackwell Transportation Center at the University of
Arkansas.

References

Aarts, C., Lenstra, J. (Eds.), 1997. Local Search in Combinatorial Optimization. John Wiley & Sons, New York.
Ball, M.O., Magnanti, T.L., Monma, C.L., Nemhauser, G.L. (Eds.), 1995. Network Routing. Elsevier Science,
Amsterdam.
Bodin, L.D., Golden, B.L., Assad, A., Ball, M.O., 1983. Routing and scheduling of vehicles and crews: the state of the
art. Computers & Operations Research 10 (2), 63–211.
Casco, D., Golden, B., Wasil, E., 1988. Vehicle routing with backhauls: models, algorithms and case studies. In:
Golden, B., Assad, A. (Eds.), Vehicle Routing: Methods and Studies. North-Holland, Amsterdam, pp. 127–147.
Chiang, W., Russell, R., 1997. A reactive tabu search metaheuristic for the vehicle routing problem with time windows.
INFORMS Journal on Computing 9 (4), 417–430.
Cheung, R.K., Hang, D.D., 2003. Multi-attribute label matching algorithms for vehicle routing problems with time
windows and backhauls. IIE Transactions 35 (3), 191–205.
Croes, A., 1958. A method for solving traveling salesman problems. Operations Research 5, 791–812.
Deif, I., Bodin, L.D., 1984. Extension of the Clarke and Wright algorithm for solving the vehicle routing problem with
backhauling. In: Didder, A. (Ed.), Proceedings of the Babson Conference on Software Uses in Transportation and
Logistic Management, Babson Park, pp. 75–96.
Duhamel, C., Potvin, J., Rousseau, J., 1997. A tabu search heuristic for the vehicle routing problem with backhauls and
time windows. Transportation Science 31 (1), 49–59.
144 Y. Zhong, M.H. Cole / Transportation Research Part E 41 (2005) 131–144

Fisher, M., Jaikumar, R., 1981. A generalized assignment heuristic for vehicle routing. Networks 11, 109–124.
Gelinas, S., Desrochers, M., Desrosiers, J., Solomon, M., 1995. A new branching strategy for time constrained routing
problems with application to backhauling. Annals of Operations Research 61, 91–109.
Goetschalckx, M., Jacobs-Blecha, C., 1989. The vehicle routing problem with backhauls. European Journal of
Operational Research 42, 39–51.
Golden, B., Assad, A. (Eds.), 1988. Vehicle Routing: Methods and Studies. Elsevier Science Publishers, North-Holland,
Amsterdam.
Golden, B., Baker, E., Alfaro, J., Schaffer, J., 1985. The vehicle routing problem with backhauling: two approaches. In:
Hammesfahr, R. (Ed.), Proceedings of the 21st Annual Meeting of SE TIMS, Myrtle Beach, South Carolina, pp. 90–
92.
Jacobs-Blecha, C., Goetschalckx, M., 1993. The vehicle routing problem with backhauls: properties and solution
algorithms. Technical report MHRC-TR-88-13, Georgia Institute of Technology, Atlanta, Georgia.
Kontoravdis, G., Bard, J., 1995. A GRASP for the vehicle routing problem with time windows. ORSA Journal on
Computing 7 (1), 10–23.
Potvin, J., Bengio, S., 1996. The vehicle routing problem with time windows, part II: genetic search. INFORMS Journal
on Computing 8 (2), 165–172.
Potvin, J., Duhamel, C., Guertin, F., 1996a. A genetic algorithm for vehicle routing with backhauling. Applied
Intelligence 6, 345–355.
Potvin, J., Kervahut, T., Garcia, B., Rousseau, J., 1996b. The vehicle routing problem with time windows, part I: tabu
search. INFORMS Journal on Computing 8 (2), 158–164.
Reeves, C. (Ed.), 1993. Modern Heuristic Techniques for Combinatorial Problems. Blackwell Scientific Publications,
Oxford.
Russell, R., 1995. Hybrid heuristics for the vehicle routing problem with time windows. Transportation Science 29 (2),
156–166.
Solomon, M., 1987. Algorithms for the vehicle routing and scheduling problem with time window constraints.
Operations Research 35, 254–265.
Toth, P., Vigo, D., 1996. A heuristic algorithm for the vehicle routing problem with backhauls. In: Bianco, L., Toth, P.
(Eds.), Advanced Methods in Transportation Analysis: Proceedings of the 2nd TRISTAN Conference. Springer,
Berlin.
Voudouris, C., Tsang, E., 1999. Guided local search and its application to the traveling salesman problem. European
Journal of Operational Research 113, 469–499.