OR Spectrum (2013) 35:457–478

DOI 10.1007/s00291-012-0284-5

REGULAR ARTICLE

Integrated planning of loaded and empty container

movements

Kris Braekers · An Caris · Gerrit K. Janssens

Published online: 8 February 2012

© Springer-Verlag 2012

Abstract Efficiently planning drayage operations is an important task for transpor-

tation companies since these operations constitute a large part of the cost of an inter-
modal transport. In this paper, a full truckload vehicle routing problem for transporting
loaded and empty containers in drayage operations is studied. For empty container
transports, either the origin or the destination is not predefined. The problem is for-
mulated as an asymmetric multiple vehicle Travelling Salesman Problem with Time
Windows (am-TSPTW). Two solution approaches are proposed: a sequential and an
integrated approach. For both approaches, a single- and a two-phase deterministic
annealing algorithm are presented. Results show that the proposed algorithms are able
to find good quality solutions in a small amount of computation time. The integrated
approach clearly outperforms the sequential one and the results confirm the advan-
tage of using a two-phase algorithm for vehicle routing problems with hierarchical
objectives. Finally, it is shown that the proposed integrated solution method improves
previous results on a similar problem.

Keywords Vehicle routing · Empty containers · Drayage · Deterministic annealing

1 Introduction

In the hinterland of a major seaport, drayage operations are concerned with the trans-
port of loaded and empty containers between shippers, consignees, inland container

K. Braekers (B) · A. Caris · G. K. Janssens

Transportation Research Institute, Hasselt University, Campus Diepenbeek,
Wetenschapspark gebouw 5, 3590 Diepenbeek, Belgium
e-mail: kris.braekers@uhasselt.be

A. Caris
Research Foundation Flanders (FWO), Egmontstraat 5, 1000 Brussels, Belgium

123
458 K. Braekers et al.

terminals and container terminals at the port. Efficiently planning these operations is
an important task for shipping lines and transportation companies since they represent
a considerable amount of the total cost of an intermodal transport (Smilowitz 2006).
Special attention should be paid to minimizing empty container movements since they
are costly non-revenue generating activities.
Traditionally, a sequential approach is used for the operational planning of loaded
and empty container movements. The problem is decomposed into two subproblems:
an allocation and a routing problem. First, an empty container allocation model is used
to determine the optimal repositioning of empty containers based on the locations of
demand and supply in the region. Such an allocation model minimizes the total distance
travelled by empty containers, without taking vehicle routing decisions into account.
Next, a routing model is used to create efficient vehicle routes performing both loaded
and empty container transport requests (Crainic et al. 1993; Huth and Mattfeld 2009).
The objective is to minimize the number of vehicles used, minimize the distance trav-
elled, minimize travelling time or a combination of these. Such vehicle routing models
are known to be very complex, especially when time windows are involved (Cordeau
et al. 2007). To solve problems of realistic size often meta-heuristics, such as tabu
search or simulated annealing, are used.
Palmgren et al. (2003) describe a similar decomposition approach for the log truck
scheduling problem in the forest industry. The authors solve a transportation problem
to find the possible flow between supply and demand points. When creating feasible
routes for their problem, the results of the transportation problem are used to limit the
number of demand points that can be reached from a certain supply point.
Recently, some efforts are made to integrate the allocation and routing subproblems
in drayage operations described above. Using an integrated approach, thus consider-
ing empty container allocation and vehicle routing decisions simultaneously, drayage
costs may be reduced. Since the origin or destination of empty container transports are
not determined in advance, the resulting problem is even more complex. Although sev-
eral papers have addressed this idea, the advantage of an integrated over a sequential
approach for planning drayage operations, has not been quantified.
In this paper, both a sequential and integrated approach are presented and compared
with each other for a full truckload drayage problem with time windows. For both
approaches, the problem is formulated as an asymmetric multiple vehicle Travelling
Salesman Problem with Time Windows (am-TSPTW) and a single- and a two-phase
deterministic annealing algorithm are proposed to solve the problem. Results clearly
show that the integrated approach outperforms the sequential one. Besides, results in
this paper confirm the advantage of using a two-phase approach for vehicle routing
problems with the hierarchical objective of first minimizing the number of vehicles
used and second total distance travelled. Finally, to assess the quality of the proposed
solution method, it is shown that our algorithm improves previous results on a similar
problem.
In Sect. 2, an overview of related literature is given. A detailed problem descrip-
tion, together with an overview of both solution approaches, is presented in Sect. 3.
In Sect. 4, the proposed deterministic annealing algorithms are described. The experi-
mental design and results are discussed in Sect. 5. Finally, conclusions are drawn and
future research opportunities are identified in Sect. 6.

123
Integrated planning of loaded and empty container movements 459

2 Literature review

In this section, related literature is presented. First, literature on empty container allo-
cation models and full truckload routing is discussed. Next, recent approaches to
integrate empty container allocation and vehicle routing decisions are described.

2.1 Empty container allocation models

The objective of an empty container allocation model is to determine the best distribu-
tion of empty containers, while satisfying both known and forecasted demand (Crainic
et al. 1993). Besides empty container demand and supply, repositioning empty con-
tainers in order to be able to satisfy empty container demand in future periods may
be considered. Container allocation models are described among others in Chang et
al. (2008), Chu (1995), Crainic et al. (1993), Di Francesco et al. (2006), Jula et al.
(2003) and Olivo et al. (2005). For a detailed review of these and other empty container
allocation models the reader is referred to Braekers et al. (2011).

2.2 Full truckload routing

Routing problems in drayage operations can be classified as full truckload pickup and
delivery problems (Erera and Smilowitz 2008). The routing problem considered in
this paper is a deterministic Full Truckload Pickup and Delivery Problem with Time
Windows (FT-PDPTW), which can be reduced to an asymmetric multiple vehicle
Travelling Salesman Problem with Time Windows (am-TSPTW).
Gronalt et al. (2003) develop four savings based heuristics for a FT-PDPTW. Goods
are transported between distribution centers or depots. Vehicles are based at different
depots and may perform several routes during the planning period. A FT-PDPTW with
multiple vehicle depots and additional weight constraints in the context of log truck
scheduling is studied by Gronalt and Hirsch (2007). Different variants of the tabu
search meta-heuristic are proposed to solve the problem. Imai et al. (2007) introduce
a full truckload pickup and delivery problem in the context of an intermodal terminal.
Caris and Janssens (2009) extend this problem to a FT-PDPTW by including time
window constraints at the customer locations. The problem is solved by a local search
heuristic. In a subsequent work, a deterministic annealing algorithm is proposed to
solve the problem (Caris and Janssens 2010). The effect of the introduction of an
appointment-based access control system at a port on full truckload drayage opera-
tions with time windows is studied by Namboothiri and Erera (2008). Mes et al. (2007,
2010) propose an agent-based approach for a dynamic version of the FT-PDPTW.
Jula et al. (2005) and Wang and Regan (2002) show that a FT-PDPTW may be
transformed to an asymmetric multiple vehicle Travelling Salesman Problem with
Time Windows (am-TSPTW) by collapsing each transport request into a single node.
Wang and Regan (2002) use a time window partitioning method to solve the problem.
The authors iteratively solve an under- and over-constrained version of the prob-
lem. Jula et al. (2005) present two exact approaches for solving small instances of the
am-TSPTW, using dynamic programming and genetic algorithms. An insertion
heuristic is proposed to solve large problem instances. Lower bounds on the

123
460 K. Braekers et al.

number of vehicles for the (a)m-TSPTW are presented by Desrosiers et al. (1988)
using Lagrangian relaxation, and by Mitrović-Minić and Krishnamurti (2006) using
precedence graphs.
The Multiple Depot Vehicle Scheduling Problem with Time Windows (MDVSPTW),
which is equivalent to a FT-PDPTW with multiple vehicle depots, is studied by Min-
gozzi et al. (1995), Desaulniers et al. (1998) and Hadjar and Soumis (2009). Currently,
problems up to 900 tasks and can be solved to optimality. However, since the problem
is applied in the context of urban bus scheduling, time windows are assumed to be
small (maximum 30 min) while in our problem time windows up to 4 h are considered.

2.3 Integrated approaches

Dejax and Crainic (1987) already suggested that the independent consideration of
container allocation and vehicle routing neglects possible synergies arising from an
integrated view on these problems. However, Crainic et al. (1993) stated that a sin-
gle mathematical model comprising container allocations and vehicle routing would
be computationally intractable. Due to the continuous improvement of Operations
Research techniques and computer capabilities, this opinion has changed. Recently, a
number of attempts have been made to integrate container (or trailer) allocation and
vehicle routing decisions.
Baldacci et al. (2006) study the Multiple disposal facilities and multiple inventory
locations Rollon–Rolloff Vehicle Routing Problem (M-RRVRP). The problem arises
in the sanitation industry where tractors move trailers between customer locations, dis-
posal facilities and inventory locations. Five types of trips are identified. For some trip
types, the origin or destination of an empty trailer is not predefined. A set partitioning
formulation is used and an exact solution method is proposed. Deidda et al. (2008)
propose a static, deterministic optimization model simultaneously addressing the allo-
cation of empty containers between shippers, consignees and a port and the design of
vehicle routes for empty containers. Loaded container transports are not considered.
Vehicles are located at a single depot at the port and have a capacity of two containers.
An exact algorithm is proposed. Huth and Mattfeld (2009) compare the results of a
sequential and an integrated decision making approach for allocating and routing swap
containers in a hub-and-spoke network. Vehicle routing decisions for both loaded and
empty containers with a vehicle capacity of two containers are considered. The allo-
cation problem is modelled as a multi-stage transportation problem while the routing
problem is modelled as a generalization of the pickup and delivery problem. A Large
Neighborhood Search (LNS) is used to solve the routing model in both the sequential
and integrated approach. Results show the advantage of an integrated approach for
this type of problem.
Some integrated approaches for full truckload (drayage) problems with time win-
dows have been proposed as well. Currie and Salhi (2004) propose a tabu search heu-
ristic for a FT-PDPTW with heterogeneous products and vehicles where the pickup
points of goods to be delivered to customers are not predefined. The objective is to
minimize total costs, including a fixed cost per vehicle used. Smilowitz (2006) stud-
ies the routing and scheduling of loaded and empty trailers or containers in drayage
operations. Trailer allocation decisions are made simultaneously with vehicle routing

123
Integrated planning of loaded and empty container movements 461

decisions by introducing flexible tasks for empty trailers demanded and supplied (ori-
gins resp. destinations are not predefined). Only allocations with a distance smaller
than a threshold are considered as possible executions for a flexible task. The objec-
tive of the model is to both minimize fleet size and travel time. The model is solved
by a branch-and-bound heuristic using column generation. This solution method is
improved in a subsequent work (Francis et al. 2007). Recently, dynamic versions of
this problem are studied by Escudero et al. (2011) and Zhang et al. (2011). Another
column generation approach embedded in a branch-and-bound framework for opti-
mizing drayage operations of trailers is proposed by Ileri et al. (2006). A heterogeneous
fleet of drivers with different start and end locations is assumed. The objective is to
minimize costs with company drivers having a different cost structure than third party
drivers. When travelling between certain types of tasks, intermediate stops at empty
container storage and supply locations are introduced to ensure that vehicles arrive at
the starting location of a task appropriately (empty or with an empty trailer). A similar
problem, in the context of container transportation, is investigated by Zhang et al.
(2009). Empty container allocations are integrated with routing decisions for vehicles
with a single container capacity. A single container terminal and several vehicle depots
with an empty container stock are considered. Vehicles do not need to return to their
starting depot. The objective is to minimize total travel time. It is shown that the prob-
lem can be formulated as a multiple vehicle Travelling Salesman Problem with Time
Windows (m-TSPTW) and multiple depots. A Reactive Tabu Search (RTS) algorithm
is proposed to solve the problem. Zhang et al. (2010) extend this problem to a multiple
depot, multiple terminal problem and solve it by a time window partitioning method.
Finally, Zhang et al. (2011) look at the single depot, single terminal problem where the
number of empty containers available at the depot is limited. Again an RTS algorithm
is proposed. It seems that solving this problem is much more complex than solving
the problems in Zhang et al. (2009, 2010) where there is no limit on the number of
empty container available at the depots.

3 Problem description

The problem in this paper is to create efficient vehicle routes performing all loaded and
empty container transport requests in a region during a single day. It is assumed that
a single vehicle depot and one or more container terminals are located in the region.
Both the container terminals and the vehicle depot are opened during the whole plan-
ning period. Empty containers can be stored at each container terminal and sufficient
empty containers are available at each terminal.
A loaded container transport represents a transport from a shipper to a container
terminal (outbound loaded container) or from a container terminal to a consignee
(inbound loaded container). For each container, the terminal to be used is predefined
so that for all loaded container transports the origin and destination are known in
advance. Time windows are imposed on these transport requests.
For empty container transports, either the origin or the destination is not defined in
advance. A shipper may request an empty container to be delivered before a specific
point in time. The origin of this empty container is irrelevant for the shipper and can

123
462 K. Braekers et al.

be chosen by the decision maker. On the other hand, a consignee will have an empty
container available after unloading an inbound loaded container. This container be-
comes available at a certain point in time and should be picked up before the end of
the day. The destination of the empty container is determined by the decision maker.
Empty containers can thus be transported from consignees to a container terminal,
from a container terminal to a shipper or directly from a consignee to a shipper.
A homogeneous fleet of vehicles with a single container capacity is assumed. All
vehicles start and end their route at the vehicle depot. When a vehicle arrives early at a
location, waiting is allowed at no cost. The service time to load and unload containers
is constant and the same for loaded and unloaded containers. A hierarchical objective
function is used. The primary objective is to minimize the number of vehicles used
while the secondary objective is to minimize total distance travelled.
The problem described above is similar to the ones studied in Zhang et al. (2009,
2010), although some differences exist. The objective is twofold in this paper (first
minimize vehicles, then minimize distance) while respectively only total travelling
time and total operation time are minimized in Zhang et al. (2009, 2010). Besides,
when a vehicle delivers a loaded container to a consignee, it does not have to wait
at this location until the container is unloaded and can be picked up. Instead, in this
paper it is assumed that the vehicle may leave for another task and an empty container
that becomes available at the consignee’s location may be picked up by any vehicle.
Finally, in this paper only a single vehicle depot is assumed and the empty container
stocks are located at the container terminals rather than at the vehicle depot.
In the following sections, the sequential and integrated solution approaches are
described in detail. To make this discussion more clear, a small example is shown
in Fig. 1 to demonstrate how a solution is found by the two solution approaches.
Figure 1a shows a small problem situation. The network consists of a vehicle depot,
two container terminals with an empty container stock, a single loaded container that
has to be transported from one of the terminals to a consignee, a single empty container
supply location and a single empty container demand location. No time windows and
a single vehicle are considered in this example. Parts b, c and d of Fig. 1 are discussed
throughout Sects. 3.1 and 3.2.

3.1 Sequential approach

When solving the problem sequentially, empty container allocations are determined
before vehicle routes are created. This may lead to a suboptimal solution but reduces
the complexity of the vehicle routing problem. The allocation model is discussed in
Sect. 3.1.1. This model results in a set of empty container transports that need to
be performed. In a second step, the routing problem is solved for loaded and empty
container transports together. This problem is described in Sect. 3.1.2.

3.1.1 Empty container allocation problem

Based on known demand and supply locations, the best distribution of empty con-
tainers is determined by an allocation model. The objective is to minimize the total
distance travelled by empty containers.

123
Integrated planning of loaded and empty container movements 463

(a) (b)

(c) (d)

Fig. 1 Illustration of both solution approaches

The problem is formulated as a transportation problem. The set of origins is com-

posed of the container terminals and the empty containers supplied by consignees.
Similarly, the set of destinations is composed of the container terminals and the empty
containers demanded by shippers. For each container terminal, supply at the origins
and demand at the destinations are set to respectively the number of empty containers
demanded and supplied in the region. For each possible allocation, the cost is repre-
sented by the distance between the origin and destination. The cost for allocations that
are not feasible because of time window violations is set to a very large value. Finally,
terminal–terminal allocations have a cost of zero since they do not represent actual
movements.
The transportation problem is solved optimally by the well-known Ford–Fulker-
son algorithm (Ford and Fulkerson 1956). The optimal allocations (consignee-shipper,
consignee-terminal and terminal-shipper) represent the empty container transports that
need to be performed. These transports now have a fixed start and end location. For
each transport, the time window for the start of service at the origin is determined
based on the time windows of the origin and destination.
Solving the empty container allocation problem for the example in Fig. 1 leads
to the optimal (least-distance) solution shown in part b of this figure. The empty

123
464 K. Braekers et al.

container supplied is transported to the container terminal on the right while the empty
container demanded is supplied from the container terminal on the left. Another option
is to perform a street turn by transporting the empty container supplied directly to the
empty container demand location, but then the distance traveled by the empty container
would be larger.

3.1.2 Routing problem

Once the empty container transport requests are known, the routing problem becomes
a Full Truckload Pickup and Delivery Problem with Time Windows (FT-PDPTW).
Since vehicles are assumed to have a single container capacity, the origin and destina-
tion of a transport request should be visited immediately after each other by the same
vehicle. Therefore, a transport request may be represented by a single node and the
problem can be formulated as an asymmetric multiple vehicle Travelling Salesman
Problem with Time Windows (am-TSPTW) (Jula et al. 2005; Wang and Regan 2002).
The following notation is used. Let G = (N , A) be a graph with node set N =
{0, 1, . . . , n} and arc set A = {(i, j)|i, j ∈ N , i = j}. The set of nodes consists of a
vehicle depot (index 0) and n tasks or transport requests to be performed. Each task
i ∈ N \{0} has:
– a start location,
– an end location,
– a distance di which is the distance to travel directly between the start and end
location,
– a duration si which is the sum of the loading time, the travel time from start to end
location and the unloading time,
– a time window [ai , bi ] between which the node should be reached by a vehicle in
order to be able to perform the task in time.
For the vehicle depot (i = 0), di and si are equal to zero and the time window is
equal to the total planning period [0, b0 ]. It is assumed that the distance and travel
time between two locations are proportional to the Euclidean distance between these
locations.
The parameters di j and ti j represent respectively the distance and time to travel
from the end location of task i to the start location of task j. The maximum number
of vehicles available is K and M is a very large number (at least b0 + max(i, j)∈A ti j ).
Binary decision variables xi j are used to determine whether any vehicle v ∈ V
travels directly from the end location of task i to the start location of task j. Continu-
ous variables ti represent the point in time at which a vehicle starts task i. The routing
problem is formulated as follows:
⎛ ⎞
  
lex min ⎝ x0i , di j xi j + di ⎠ (1)
i∈N (i, j)∈A i∈N
Subject to

xi j = 1 ∀i ∈ N , i  = 0 (2)
j∈N

123
Integrated planning of loaded and empty container movements 465


x0 j ≤ K (3)
j∈N
 
xi j = x ji ∀i ∈ N (4)
j∈N j∈N
ti + si + ti j ≤ t j + M(1 − xi j ) ∀(i, j) ∈ A, j  = 0 (5)
ti + si + ti0 ≤ b0 + M(1 − xi0 ) ∀i ∈ N (6)
ai ≤ ti ≤ bi ∀i ∈ N (7)
ti ≥ 0 ∀i ∈ N (8)
xi j ∈ {0, 1} ∀(i, j) ∈ A (9)

As discussed in Sect. 3, a hierarchical or lexicographic objective function  is used (1).

The primary objective is to minimize the number of vehicles used ( i∈N x0i ). The
secondary objective is to minimize total distance
 travelled, which consists of the dis-
tance travelled from each task
 to the next ( (i, j)∈A di j x i j ) and the distance travelled
to perform each task itself ( i∈N di ). Constraints (2), (3) and (4) are flow constraints.
Constraint (5) ensures that a vehicle cannot start a new task before finishing the previ-
ous task and travelling to the new one. Constraint (6) ensures that all vehicles return to
the vehicle depot before the end of the planning period. Time windows are represented
by constraint (7). Finally, constraints (8) and (9) make sure that both types of variables
only take on the appropriate values.
Figure 1b shows that after solving the empty container allocation problem for the
small example, three transport tasks need to be performed: one loaded container trans-
port and two empty container transports. Using the vehicle routing problem described
above leads to the optimal solution for the sequential approach which is shown in
Fig. 1c.

3.2 Integrated approach

When using an integrated approach, empty container allocations are not made before-
hand but simultaneously with vehicle routing decisions. This means that the origin
(destination) of an empty container demanded by a shipper (supplied by consignee)
is not fixed in advance.
The integrated problem can be formulated by creating a node for the vehicle depot,
for all loaded container transports and for all feasible empty container allocations.
Extra constraints should impose that exactly one allocation node for each consignee
and for each shipper is chosen. Such a formulation of the problem can be found in
Braekers et al. (2010). Solving this problem is however problematic since the number
of possible allocations and thus the number of nodes in the network becomes very
large for problems of a realistic size. Smilowitz (2006) uses a similar approach by
defining feasible allocations as possible executions of a flexible task. To overcome
the problem of the exponential growth of the network, a heuristic column generation
approach is proposed and the number of feasible allocations is restricted by imposing
a maximum distance.

123
466 K. Braekers et al.

Table 1 Node parameters

Nodes Distance di Duration si Time window [ai , bi ]

i ∈ NVD 0 0 [0, b0 ]
i ∈ NL Distance between start Loading time + travel time [ai , bi ]
and end location between start and end
locations + unloading time
i ∈ NS 0 Loading time [ai , b0 ]
i ∈ ND 0 Unloading time [0, bi ]

Table 2 Calculation of distance coefficients d̂i j .

j ∈ NVD ∪ NL j ∈ NS j ∈ ND

i ∈ NVD ∪ NL di j di j mine=1,...,r (die + dej )

i ∈ NS mine=1,...,r (die + dej ) mine=1,...,r (die + dej ) di j
i ∈ ND di j di j mine=1,...,r (die + dej )

Alternatively, the integrated problem can be formulated like the routing problem
for the sequential approach. This is done by introducing an intermediate stop at a
container terminal when travelling between certain types of nodes (Ileri et al. 2006;
Zhang et al. 2009). The node set N is composed of the vehicle depot (N V D , index 0),
a set of nodes for the loaded container transport (N L ), a set of nodes for the empty
container demand locations (N D ) and a set of nodes for the empty container supply
locations (N S ).
Again, to each node is assigned a distance di , a duration si and a time window
[ai , bi ]. For the vehicle depot and the nodes representing the loaded container trans-
ports, these values are identical as those for the sequential approach. For the empty
container supply and demand locations di is zero and si is equal to the (un)loading time.
Time windows are [0, bi ] for demand locations with bi the latest time the container
should be delivered at node i and [ai , b0 ] for supply locations with ai the time that the
container becomes available for pickup node i. An overview of these values is shown
in Table 1. The calculation of the distance d̂i j of travelling between two nodes i and j is
shown in Table 2, where r is the number of container terminals in the region. When an
intermediate stop at a container terminal needs to be made, the choice for the terminal
is based on a lowest distance basis. Travel times tˆi j between two nodes are calculated
similarly as the distances, but augmented with the container loading and unloading time
when making an intermediate stop at a container terminal or travelling directly from
an empty container supply to an empty container demand location. The other variables
(K , M, xi j , ti ) and the formulation of the problem are the same as for the sequential
approach, except that variables di j and ti j are replaced by respectively d̂i j and tˆi j .
Although the formulation is the same, the integrated problem is harder to solve
than the routing problem of the sequential approach. The reason is twofold. First, the
number of nodes slightly increases. Second, the nodes representing the empty con-
tainer demand and supply locations have much wider time windows than the nodes
representing empty container transport requests in the sequential approach.

123
Integrated planning of loaded and empty container movements 467

When the integrated approach is used to solve the small example in Fig. 1a, the
optimal vehicle route is determined without first deciding on the origin and destination
of the empty container demanded respectively supplied. Figure 1d shows the optimal
solution. Clearly the total distance traveled is less than for the solution of the sequential
approach (shown in Fig. 1c). The reason for this lower distance is that another empty
container allocation is chosen. In this case, an empty container is transported directly
from the empty container supply location to the empty container demand location.
Although the distance traveled by empty containers is larger for the integrated solu-
tion (since the optimal container allocation is not chosen), total distance traveled by
the vehicle is smaller than for the solution of the sequential approach.

4 Deterministic annealing algorithms

In the previous section, it is shown that for both solution approaches an am-TSPTW
needs to be solved. Solving such a problem exactly for realistic problem sizes is very
hard (Jula et al. 2005). Meta-heuristics can be used to find good and often optimal or
near-optimal solutions. In this paper, a deterministic annealing algorithm is proposed.
Deterministic annealing is a variant on the well-known simulated annealing meta-
heuristic. With deterministic annealing, also known as threshold accepting, a neigh-
boring solution worsening the objective function is accepted if the worsening is smaller
than a certain threshold (Dueck and Scheuer 1990). This threshold may be adapted
during the search. Deterministic annealing has been proven more effective than its
stochastic counterpart simulated annealing for several problems. Recently, determin-
istic annealing has been successfully implemented for a number of vehicle routing
problems (Bräysy et al. 2003, 2008; Caris and Janssens 2010; Nikolakopoulos and
Sarimveis 2007; Tarantilis et al. 2004).
In Sect. 4.1, the insertion heuristic used to find a feasible starting solution for the
algorithm is discussed. The local search operators and the deterministic annealing
scheme are respectively described in Sects. 4.2 and 4.3. Finally, the implementation
of the algorithm on the problem under study is discussed in Sect. 4.4.

4.1 Insertion heuristic

An initial solution for the problem is obtained by a parallel insertion heuristic similar
to the one proposed by Jula et al. (2005). Starting from a number of vehicles equal to
the lower bound, a node is randomly selected and inserted in its best possible position.
This operation is repeated until all nodes are inserted. If no insertion is possible, an
additional vehicle is added. The procedure is iterated a number of times to obtain a
good starting solution.

4.2 Local search operators

Several local search operators are used to improve the initial solution. Three operators
try to reduce the distance travelled. With the relocate operator, a node is removed
from its route and reinserted in another route or in another place in its original route.

123
468 K. Braekers et al.

The 2-Opt* operator (Potvin and Rousseau 1995) removes an arc from two routes and
recombines the resulting parts, that is: the first part of route one with the second part
of route two and vice versa. The exchange operator swaps a number of nodes between
two routes. Several variations of this operator are used, based on the number of nodes
that are swapped: exchange(1,1), exchange(2,1), exchange(2,2), exchange(3,2) and
exchange(3,3). Only one of these exchange operators is used in an iteration of the
algorithm. For the exchange(1,1) operator the insertion position of a node can also be
one place before or after the removal position of the other node. When two nodes from
a route are swapped to another route, the reverse insertion of these nodes is consid-
ered as well. For each operator, one route is selected randomly and all combinations
of nodes or arcs of this route with those of the other routes are considered. A first
improvement strategy is used.
Finally, two operators try to reduce the number of vehicles by reinserting all nodes
of respectively one or several routes into the other routes. The first operator tries to
insert all nodes of a randomly selected route into the other routes. The second operator
tries to insert all nodes of the p shortest routes into all other routes and p − 1 empty
routes. The parameter p is defined as a percentage of the number of routes in the
solution.

4.3 Deterministic annealing scheme

The proposed deterministic annealing scheme is based on the one in Bräysy et al.
(2008) and Caris and Janssens (2010). It is shown in Algorithm 1. The current best
solution Sbest is set to the best solution found by the insertion heuristic and the threshold
T is set to its maximum value Tmax . The deterministic annealing algorithm is iterated
n times. At each iteration, all local search operators are used in a random order. New
solutions with a lower number of vehicles as the current solution are always accepted.
New solutions with the same number of vehicles and total distance lower than the dis-
tance of the current solution plus the threshold T are accepted. If a solution is better
than the best solution found so far, this solution is set as best solution. When no new
best solution has been found, the threshold T is reduced by the threshold reduction
parameter T . Whenever T becomes negative, it is reset to r × Tmax , with r a random
number between zero and one. In case T becomes negative and no improvement has
been found for n imp iterations, the search is restarted from the best solution.

4.4 Implementation

In this section, the implementation of the deterministic annealing algorithm on the

problem described in Sect. 3 is discussed. For both the sequential and integrated
approach, a single- and a two-phase solution algorithm are presented.

4.4.1 Single-phase algorithms

The implementation of the single-phase algorithms is straightforward. An initial solu-

tion is found by the insertion heuristic and this solution is improved by the deterministic

123
Integrated planning of loaded and empty container movements 469

Algorithm 1 Implementation of the deterministic annealing algorithm

T = Tmax , ilast = 0 and Sbest = best solution of the insertion heuristic
for i = 1 → n do
for j = 1 → m do
Apply local search operator on S and accept or reject new solution S
if S is accepted then
S ← S
if f (S) < f (Sbest ) then
Sbest ← S
ilast ← 0
end if
end if
end for
if ilast > 0 then
ilast ← ilast + 1
T ← T − T
if T < 0 then
T ← r × Tmax
if ilast > n imp then
S ← Sbest
ilast ← 0
end if
end if
end if
end for

annealing algorithm for a predefined number of iterations. All local search operators
are used during the algorithm which means that the number of vehicles and total
distance are minimized simultaneously.

4.4.2 Two-phase algorithms

As pointed out by Bent and Van Hentenryck (2006) and Homberger and Gehring
(2005), simultaneous reduction of the number of vehicles and total distance travelled
in vehicle routing problems by a meta-heuristic controlling a neighborhood search,
may lead to an important shortcoming. The objective function often drives the search
towards solutions with a small distance. This complicates reaching solutions with a
low number of vehicles but higher distance, i.e. the search is mainly guided by the
secondary objective. To overcome this shortcoming, a two-phase solution method may
be used where the number of vehicles is minimized during the first phase and total dis-
tance is minimized during the second phase. Homberger and Gehring (2005) present
such a two-phase solution method for the vehicle routing problem with time windows.
A (μ, λ)-evolution strategy is used during the first phase, while a tabu search algorithm
is used during the second phase. Bent and Van Hentenryck (2006) use respectively a
simulated annealing algorithm and a Large Neighborhood Search (LNS) during their
two-phase algorithm for the pickup and delivery problem with time windows. Since
these two-phase methods provide very good results, a two-phase solution method for
the problem in this paper is proposed as well. During both phases the deterministic
annealing meta-heuristic is used.

123
470 K. Braekers et al.

First phase: reduce the number of vehicles Instead of using objective function (1), a
specific hierarchical objective function presented by Bent and Van Hentenryck (2006)
is used during the first phase of the algorithm. This objective function (10) guides the
search towards solutions with a small number of vehicles, while partially ignoring the
secondary objective of reducing total distance. Parameter lv represents the number of
nodes visited by vehicle v ∈ V .
⎛ ⎞
   
lex min ⎝ x0i , − lv2 , di j xi j + di ⎠ (10)
i∈N v∈V (i, j)∈A i∈N

The objective function (10) consists of three hierarchically (or lexicographically) struc-
tured objectives. The primary objective is to minimize the number of vehicles used
while the secondary objective is to maximize the sum of the squares of the number
of nodes in each route. Finally, minimizing total distance is the third objective. The
purpose of the second objective is to favour solutions with an unbalanced distribution
of nodes over the vehicles over solutions with an even distribution of nodes, i.e. a solu-
tion with a few long and a few short routes is preferred over a solution where all routes
have a length close to the average. The idea behind this objective is to remove nodes
from shorter routes and insert them into longer routes, thereby gradually reducing the
number of vehicles (Bent and Van Hentenryck 2006).
During each iteration of the deterministic annealing algorithm, all five types of
local search operators are applied in a random order. The two route reducing operators
have an effect on the primary objective, while the relocate, 2-Opt*, exchange(2,1)
and exchange(3,2) operators improve the secondary objective. The exchange(1,1),
exchange(2,2) and exchange(3,3) operators only have an effect on the third objective
and are mainly used to diversify the search. During this phase, the acceptance rule
of a neighboring solution is as follows. A new solution is accepted when it is better
as the current solution according to objective function (10) or when it has the same
number of vehicles and the worsening of the second objective value is smaller than
the threshold value T .
The implementation of the first phase of the sequential algorithm is again straight-
forward: the initial solution found by the insertion heuristic is improved by the deter-
ministic annealing algorithm for a fixed number of iterations. The implementation for
the integrated algorithm is more complex. The insertion heuristic and first phase of
the algorithm are not directly applied on the integrated problem. Instead, an initial
solution is found in the same way as for the sequential problem (by first determining
the optimal empty container allocations by the transportation problem) and during half
the number of iterations of the first phase, the number of vehicles is reduced while
keeping the empty container allocations fixed. Next, the best solution found so far is
transformed to a solution for the integrated problem by relaxing the optimal alloca-
tions. During the second half the number of iterations of the first phase, the number
of vehicles is reduced further.

Second phase: reduce total distance During the second phase, the best solution found
during the first phase is further improved with respect to total distance. The original

123
Integrated planning of loaded and empty container movements 471

objective function (1) is used. The two route reducing operators are not used since
reducing the number of vehicles is not considered anymore.

5 Experimental design and results

5.1 Experimental design

In order to test the robustness of the algorithm, a 24 factorial design is set up (Law
2007). Four problem characteristics are identified. For each characteristic a high (+)
and low (−) value is determined. The time window width for loaded containers (F1) is
a random number between 60 and 120 min (−) or between 120 and 240 min (+). The
number of container terminals (F2) is one (−) or three (+). For each instance the num-
ber of nodes (F3) is 100 (−) or 200 (+). These nodes exist of an even amount of loaded
container delivery locations, loaded container pickup locations, empty container sup-
ply locations and empty container demand locations. Finally, the (X,Y)-coordinates of
all nodes (F4) are randomly chosen between 0 and 25 km (−) or between 0 and 50 km
(+) on both axes. This results in 16 problem classes as shown in Table 3. For each
problem class, three random problem instances are generated.
It is assumed that a single vehicle depot is located in the center of the square region
and the location of the container terminals is the same for each instance. The planning
period equals 8 h. Service time for loading/unloading a container is 10 min.

5.2 Lower bound

Lower bounds on the sequential and on the integrated problem are found by a time
window partitioning method. Time window partitioning or discretization is introduced
by Wang and Regan (2002) for solving a m-TSPTW. The time window of each node
is partitioned into smaller parts, which are considered subnodes. Binary flow vari-
ables for links between nodes are replaced by binary flow variables for links between
subnodes. Exactly one subnode of each node should be visited and for each node a
vehicle must enter and leave the same subnode. An underconstrained version of the

Table 3 Overview of problem classes

Class F1 F2 F3 F4 Class F1 F2 F3 F4

1 − − − − 9 − − − +
2 + − − − 10 + − − +
3 − + − − 11 − + − +
4 + + − − 12 + + − +
5 − − + − 13 − − + +
6 + − + − 14 + − + +
7 − + + − 15 − + + +
8 + + + − 16 + + + +

123
472 K. Braekers et al.

problem is formulated by replacing constraints (5), (6), (7) and (8) in the formulation
of Sect. 3.1.2 with constraint (11). This underconstrained model results in a lower
bound for the original problem.

xi j (ai + si + ti j − b j ) ≤ 0 ∀(i, j) ∈ A (11)

Wang and Regan (2002) use an iterative approach by partitioning the time windows
into smaller parts at each iteration. Recently, Zhang et al. (2010) improved this method
by immediately looking for a good partitioning width. In this way the underconstrained
problem needs to be solved only once, reducing computation time to a large extent.
The authors found that instances with up to 800 subnodes could be solved efficiently.
In this paper, the method of Zhang et al. (2010) is slightly adapted to obtain a better
lower bound. The binary restrictions on the flow variables are relaxed. The resulting
LP-relaxation of the underconstrained model can be solved very efficiently. Therefore,
problems with up to 5,000 subnodes can be solved within 5 min. The lower bounds
generated in this way are tighter than those generated by the integer restricted problems
in a comparable amount of computation time.
Separate lower bounds are calculated for the sequential and integrated problem. To
obtain a lower bound on the number of vehicles, the objective function of the problem
is changed to minimizing the total time needed to perform all transport requests. The
total time needed is then divided by the length of the planning period to find a lower
bound on the number of vehicles (Koo et al. 2004). A lower bound on the total distance
is found by removing the fixed vehicle costs from the objective function. It should be
noted that lower bounds on the integrated problem are less tight than those of the
sequential problem. The reason is that the average time window width is much larger
for the integrated problem. Therefore the partitioning width was chosen somewhat
wider in order to ensure a lower bound is found within reasonable computation time.

5.3 Results

The best parameter values are found by testing the algorithms for a single problem
instance of each class. Best results are achieved with the following parameter val-
ues. The insertion heuristic is iterated 1,000 times to obtain a good starting solution.
The deterministic algorithm is run for 50,000 iterations (during each phase). For the
single-phase algorithms and the second phase of the two-phase algorithms, the maxi-
mum threshold value Tmax is set at 4 km for 25 km2 regions and at 8 km for 50 km2
regions. The threshold reduction parameter T is set at Tmax /2,500. During the first
phase of the two-phase algorithms the maximum threshold value Tmax is set at 6 for
the sequential algorithm and at 12 for the integrated algorithm while the threshold
reduction parameter T is set at Tmax /1,500. For all algorithms, the reset parameter
n imp is equal to five times the number of vehicles used in the current best solution. The
transportation problem and deterministic annealing algorithm are implemented in C
on a 2.1GHz Intel Core 2 laptop with 4GB RAM. The lower bounds are computed
using Lingo 10.0.

123
Integrated planning of loaded and empty container movements 473

Table 4 Overview of results

Value Sequential Integrated

Single-phase Two-phase Single-phase Two-phase

Average number of vehicles 10.46 10.04 10.33 9.99

Average gap (absolute) 1.00 0.58 1.08 0.74
Average distance (km) 1,813 1,837 1,805 1,815
Average gap (%) 2.78 3.74 4.74 5.33
Average computation time (s) 3.40 5.38 4.27 6.59

All four algorithms are tested on the 48 problem instances. For each instance, aver-
age results over five runs of the algorithm are obtained. Detailed results are shown in
Appendix A. An overview of results is shown in Table 4. The first two rows in Table 4
show the average number of vehicles used and the average absolute gap with the cor-
responding lower bound. The average distance travelled and the average relative gap
with the corresponding lower bound, are presented in rows three and four. In the last
row, average computation in seconds is shown. From Table 4, it can be concluded that
all algorithms are able to find good quality solutions in a small amount of time. A
comparison of the sequential and integrated approach should be based on real values
instead of the gaps with the lower bounds since different bounds are calculated for
both problems. Results show that on average the integrated approach offers better
results than the sequential approach, both in terms of the number of vehicles and total
distance. Comparing results of each problem instance individually shows that the inte-
grated approach almost always performs better (21%) or equally good (74%) in terms
of the number of vehicles used. When both approaches result in the same number of
vehicles, the integrated approach is able to find a better solution in terms of distance
in 95% of the cases.
It should be noted that although the computation time of the algorithms is very
small, the average computation time of the integrated approach is slightly higher
than that of the sequential approach. To further substantiate the conclusion that the
integrated approach performs best, the computational experiments are redone for the
sequential algorithms. This time, instead of a fixed number of iterations, the sequen-
tial algorithms are run for the same time as the computation time needed by the
integrated approach. However, the results of the sequential approach hardly improve
(same number of vehicles and average distances of 1812 and 1834 km, respectively
for the single- and two-phase method) and thus the integrated approach still provides
better results.
When comparing the results of the single- and two-phase algorithms, it is clear
that the two-phase method performs much better in terms of the number of vehicles.
Average distances are slightly higher for the two-phase algorithms than for the single-
phase algorithms which can be explained by the fact that the two objectives (minimize
vehicles and distance) are often conflicting. Since minimizing the number of vehicles
is prioritized over minimizing total distance, it can be concluded that best solutions are
obtained by the integrated two-phase algorithm. This corresponds with the findings of

123
474 K. Braekers et al.

Table 5 Comparison of results with Zhang et al. (2010)

TW width Zhang et al. (2010) Own results

Vehicles Gap (%) Time (s) Vehicles Gap (%) Time (s)

1h 47.6 1.20 103 47.1 1.13 3

2h 48.2 2.68 89 47.2 1.90 3
3h 43.2 3.96 158 40.9 2.69 3
4h 37.4 4.83 124 34.6 4.26 3
Average 44.1 3.17 118 42.5 2.49 3

Homberger and Gehring (2005) and Bent and Van Hentenryck (2006) that two-phase
methods work well on problems with a hierarchical objective function.

5.4 Comparison with previous results

To further assess the quality of the proposed algorithms in this paper, a comparison
is made with previous results obtained by a time window partitioning method on a
similar problem (Zhang et al. 2010). Since minimizing the number of vehicles used
is not an objective in Zhang et al. (2010), a comparison is made with the integrated
single-phase algorithm presented in this paper. The objective function of our algorithm
is changed to minimize the total operation time of the vehicles. To meet the assumption
of Zhang et al. (2010) that there are multiple vehicle depots with a limited number
of vehicles, the travel times between the vehicle depots and the nodes are initially set
to zero. After a solution has been found by the algorithm, an optimal allocation of
vehicles to the routes is made by solving a small transportation problem. Obviously,
the algorithm could be improved by addressing this decision during the search. For
example, a local search operator that swaps the starting depots of two routes could be
introduced.
In Table 5, average results obtained by the integrated single-phase algorithm on
the twenty instances of Zhang et al. (2010) are presented and are compared with the
results of the original paper. Average results for instances with the same time window
width are presented. For both methods, the average number of vehicles, the average
gap between the total operation time and the lower bound of Zhang et al. (2010) and the
computation time are shown. Results in Table 5 show that the integrated single-phase
algorithm presented in this paper outperforms the time window partitioning method
described by Zhang et al. (2010). Besides, computation time is small.

6 Conclusions and future research

Drayage operations often constitute a large part of total costs of an intermodal trans-
port. In this paper, a sequential and an integrated approach are proposed to plan loaded
and empty container drayage operations. For both approaches, a single- and a two-
phase deterministic annealing algorithm are presented to solve a full truckload routing

123
Integrated planning of loaded and empty container movements 475

problem with time windows. It can be concluded that for both solution approaches
good quality solutions can be obtained in a small amount of time. Results clearly
indicate that the integrated approach performs better than the sequential approach.
Besides, results in this paper confirm the advantage of using a two-phase approach for
vehicle routing problems with a hierarchical objective function. Finally, it is shown
that the proposed integrated solution method improves previous results on a similar
problem.
Future research could focus on a number of extensions of the problem addressed
in this paper. The number of empty containers available at each terminal could be
assumed to be limited. This strongly complicates the problem as is shown in Zhang et
al. (2011). Another extension may be the problem where for some containers and/or
vehicles, two containers can be transported simultaneously by a single vehicle. Finally,
the hierarchical objective function may be replaced by a multi-objective function. In
this way, the trade-off between the number of vehicles used and total distance travelled
could be explored.

Appendix A: Detailed results

Table 6.1 shows detailed results for the experiments described in Sect. 5.3.

Table 6.1 Detailed results

Instance Sequential Integrated

Single-phase Two-phase Single-phase Two-phase

Va Db V D V D V D

1.1 7 1,032 6.4 1,063 7 1,019 6 1,066

1.2 7 968 6.4 999 6.6 973 6 994
1.3 6 984 6 989 6.4 963 6 971
2.1 6 1,012 6 1,013 6 1,001 6 1,003
2.2 6 983 6 987 6 988 6 974
2.3 6 928 6 927 6 923 6 919
3.1 6 830 6 830 6 812 6 811
3.2 6 760 6 755 6 743 6 750
3.3 6 728 6 727 6 701 6 702
4.1 6 702 6 705 6 697 6 696
4.2 6 689 6 689 6 682 6 680
4.3 6 719 6 723 6 699 6 697

123
476 K. Braekers et al.

Table A.1 continued

Instance Sequential Integrated

Single-phase Two-phase Single-phase Two-phase

Va Db V D V D V D

5.1 12.8 1,789 12 1,820 12 1,803 12 1,800

5.2 12 1,837 12 1,833 12 1,827 12 1,823
5.3 12 1,654 11 1,712 12 1,651 11 1,697
6.1 12 1,645 11 1,685 12 1,659 11 1,680
6.2 12 1,661 11 1,696 11.6 1,660 11 1,671
6.3 12 1,795 11 1,837 12 1,793 11.2 1,809
7.1 11 1,436 11 1,424 11 1,403 11 1,403
7.2 11 1,315 11 1,315 11 1,283 10.8 1,299
7.3 11 1,302 10 1,342 11 1,301 10 1,322
8.1 11 1,361 11 1,363 11 1,346 11 1,339
8.2 11 1,306 10.4 1,346 11 1,292 10.4 1,324
8.3 10.4 1,175 10 1,185 10 1,163 10 1,162
9.1 9 1,936 9 1,939 9 1,935 9 1,930
9.2 9 1,986 9 1,986 9 1,977 9 1,959
9.3 9 1,826 8 1,884 8.2 1,836 8 1,851
10.1 9 1,943 8 1,987 9 1,941 8 1,966
10.2 9 1,845 8 1,926 8.6 1,865 8 1,891
10.3 8.4 1,894 8 1,896 8.6 1,881 8 1,883
11.1 8 1,609 8 1,613 8 1,610 8 1,607
11.2 8 1,683 8 1,683 8 1,666 8 1,669
11.3 9 1,505 8.6 1,567 8.8 1,516 8 1,582
12.1 8 1,478 7 1,555 8 1,455 7 1,528
12.2 8 1,564 8 1,581 8 1,553 8 1,556
12.3 7 1,354 7 1,362 7 1,354 7 1,355
13.1 17 3,381 16 3,435 16 3,357 15.4 3,354
13.2 17 4,068 17 4,074 17 4,047 17 4,045
13.3 17 3,501 16 3,537 16.4 3,503 16 3,519
14.1 16 3,345 15.2 3,436 16 3,344 15 3,408
14.2 17 3,595 16 3,647 16 3,637 16 3,622
14.3 16 3,231 15 3,305 15.8 3,279 15 3,265
15.1 15 2,948 14.8 2,947 15 2,908 15 2,921
15.2 15 2,810 14 2,855 14.8 2,803 14 2,824
15.3 16 2,731 15 2,770 15.2 2,722 14.8 2,744
16.1 14.4 2,802 14 2,820 14.6 2,730 14 2,739
16.2 14 2,681 14 2,678 14 2,652 14 2,643
16.3 14 2,706 14 2,713 14 2,681 14 2,678
a V: average number of vehicles used (over five runs)
b D: average total distance travelled (over five runs)

123
Integrated planning of loaded and empty container movements 477

References

Baldacci R, Bodin L, Mingozzi A (2006) The multiple disposal facilities and multiple inventory locations
rollon-rolloff vehicle routing problem. Comput Oper Res 33(9):2667–2702
Bent R, Van Hentenryck P (2006) A two-phase hybrid algorithm for pickup and delivery vehicle routing
problems with time windows. Comput Oper Res 33(4):875–893
Braekers K, Janssens GK, Caris A (2010) A deterministic annealing algorithm for simultaneous routing
of loaded and empty containers. In: Proceedings of the Industrial Simulation Conference, Budapest,
Hungary, pp 172–176
Braekers K, Janssens GK, Caris A (2011) Challenges in managing empty container movements at multiple
planning levels. Transport Rev 31(6):681–708
Bräysy O, Berger J, Barkaoui M, Dullaert W (2003) A threshold accepting metaheuristic for the vehicle
routing problem with time windows. Cent Eur J Oper Res 11(4):369–387
Bräysy O, Dullaert W, Hasle G, Mester D, Gendreau M (2008) An effective multistart deterministic anneal-
ing metaheuristic for the fleet size and mix vehicle routing problem with time windows. Transport Sci
42(3):371–386
Caris A, Janssens GK (2009) A local search heuristic for the pre- and end-haulage of intermodal container
terminals. Comput Oper Res 36(10):2763–2772
Caris A, Janssens GK (2010) A deterministic annealing algorithm for the pre- and end-haulage of intermodal
container terminals. Int J Comput Aided Eng Tech 2(4):340–355
Chang H, Jula H, Chassiakos A, Ioannou P (2008) A heuristic solution for the empty container substitution
problem. Transport Res E-Log 44(2):203–216
Chu Q (1995) Dynamic and stochastic models for container allocation. PhD thesis, Department of Ocean
Engineering, Massachusetts Institute of Technology
Cordeau JF, Laporte G, Savelsbergh MWP, Vigo D (2007) Vehicle routing. In: Barnhart C, Laporte G
(eds) Transportation, Handbooks in Operations Research and Management Science, vol 14. Elsevier,
Amsterdam, pp 367–428
Crainic TG, Gendreau M, Dejax P (1993) Dynamic and stochastic models for the allocation of empty
containers. Oper Res 41(1):102–126
Currie RH, Salhi S (2004) A tabu search heuristic for a full-load, multi-terminal, vehicle scheduling problem
with backhauling and time windows. J Math Model Algorithm 3(3):225–243
Deidda L, Di Francesco M, Olivo A, Zuddas P (2008) Implementing the street-turn strategy by an optimi-
zation model. Marit Pol Manag 35(5):503–516
Dejax PJ, Crainic TG (1987) A review of empty flows and fleet management models in freight transporta-
tion. Transport Sci 21(4):227–248
Desaulniers G, Lavigne J, Soumis F (1998) Multi-depot vehicle scheduling problems with time windows
and waiting costs. Eur J Oper Res 111(3):479–494
Desrosiers J, Sauvé M, Soumis F (1988) Lagrangian relaxiation methods for solving the minimum fleet
size multiple traveling salesman problem with time windows. Manage Sci 34(8):1005–1022
Di Francesco M, Manca A, Olivo A, Zuddas P (2006) Optimal management of heterogeneous fleets of
empty containers. In: Proceedings of the International Conference on Information Systems, Logistics
and Supply Chain. Lyon, France, pp 922–931
Dueck G, Scheuer T (1990) Threshold accepting: a general purpose optimization algorithm appearing supe-
rior to simulated annealing. J Comput Phys 90(1):161–175
Erera AL, Smilowitz K (2008) Intermodal drayage routing and scheduling. In: Ioannou P (ed) Intelligent
freight transportation, automation and control engineering series. CRC Press, Boca Raton, pp 171–188
Escudero A, Muñuzuri J, Arango C, Onieva L (2011) A satellite navigation system to improve the manage-
ment of intermodal drayage. Adv Eng Inform 25(3):427–434
Ford LR, Fulkerson DR (1956) Solving the transportation problem. Manage Sci 3(1):24–32
Francis P, Zhang G, Smilowitz K (2007) Improved modeling and solution methods for the multi-resource
routing problem. Eur J Oper Res 180(3):1045–1059
Gronalt M, Hirsch P (2007) Log-truck scheduling with a tabu search strategy. In: Doerner K, Gendreau M,
Greistorfer P, Guthjahr W, Hartl R, Reimann M (eds) Metaheuristics—progress in complex systems
optimization. Springer, New York, pp 64–88
Gronalt M, Hartl RF, Reimann M (2003) New savings based algorithms for time constrained pickup and
delivery of full truckloads. Eur J Oper Res 151(3):520–535

123
478 K. Braekers et al.

Hadjar A, Soumis F (2009) Dynamic window reduction for the multiple depot vehicle scheduling problem
with time windows. Comput Oper Res 36(7):2160–2172
Homberger J, Gehring H (2005) A two-phase hybrid metaheuristic for the vehicle routing problem with
time windows. Eur J Oper Res 162(1):220–238
Huth T, Mattfeld DC (2009) Integration of vehicle routing and resource allocation in a dynamic logistics
network. Transport Res C-Emer 17(2):149–162
Ileri Y, Bazaraa M, Gifford T, Nemhauser G, Sokol J, Wikum E (2006) An optimization approach for
planning daily drayage operations. Cent Eur J Oper Res 14(2):141–156
Imai A, Nishimura E, Current J (2007) A Lagrangian relaxation-based heuristic for the vehicle routing with
full container load. Eur J Oper Res 176(1):87–105
Jula H, Chassiakos A, Ioannou P (2003) Empty container interchange report: methods for modeling and
routing of empty containers in the Los Angeles and Long Beach port area. Final report, Center for the
Commercial Deployment of Transportation Technologies. California State University, Long Beach
Jula H, Dessouky M, Ioannou P, Chassiakos A (2005) Container movement by trucks in metropolitan
networks: modeling and optimization. Transport Res E-Log 41(3):235–259
Koo PH, Lee WS, Jang DW (2004) Fleet sizing and vehicle routing for container transportation in a static
environment. OR Spectrum 26(2):193–209
Law A (2007) Simulation modeling and analysis. 4th edn. McGraw-Hill, New York
Mes M, van der Heijden M, van Harten A (2007) Comparison of agent-based scheduling to look-ahead
heuristics for real-time transportation problems. Eur J Oper Res 181(1):59–75
Mes M, van der Heijden M, Schuur P (2010) Look-ahead strategies for dynamic pickup and delivery prob-
lems. OR Spectr 32(2):395–421
Mingozzi A, Bianco L, Ricciardelli S (1995) An exact algorithm for combining vehicle trips. In: Daduna
J, Branco I, Paixao J (eds) Computer-aided Transit Scheduling. Lecture Notes in Economics and
Mathematical Systems, vol 430, pp 145–172. Springer, Berlin
Mitrović-Minić S, Krishnamurti R (2006) The multiple tsp with time windows: vehicle bounds based on
precedence graphs. Oper Res Lett 34(1):111–120
Namboothiri R, Erera AL (2008) Planning local container drayage operations given a port access appoint-
ment system. Transport Res E-Log 44(2):185–202
Nikolakopoulos A, Sarimveis H (2007) A threshold accepting heuristic with intense local search for the
solution of special instances of the traveling salesman problem. Eur J Oper Res 177(3):1911–1929
Olivo A, Zuddas P, Di Francesco M, Manca A (2005) An operational model for empty container manage-
ment. Marit Pol Manag 7(3):199–222
Palmgren M, Rönnqvist M, Värbrand P (2003) A solution approach for log truck scheduling based on
composite pricing and branch and bound. Int Trans Oper Res 10(5):433–447
Potvin JY, Rousseau JM (1995) An exchange heuristic for routeing problems with time windows. J Oper
Res Soc 46(12):1433–1446
Smilowitz K (2006) Multi-resource routing with flexible tasks: an application in drayage operations. IIE
Trans 38(7):568–577
Tarantilis C, Kiranoudis C, Vassiliadis V (2004) A threshold accepting metaheuristic for the heterogeneous
fixed fleet vehicle routing problem. Eur J Oper Res 152(1):148–158
Wang X, Regan AC (2002) Local truckload pickup and delivery with hard time window constraints. Trans-
port Res B-Meth 36(2):97–112
Zhang G, Smilowitz K, Erera A (2011) Dynamic planning for urban drayage operations. Transport Res
E-Log 47(5):764–777
Zhang R, Yun WY, Moon IK (2009) A reactive tabu search algorithm for the multi-depot container truck
transportation problem. Transport Res E-Log 45(6):904–914
Zhang R, Yun WY, Kopfer H (2010) Heuristic-based truck scheduling for inland container transportation.
OR Spectr 32(3):787–808
Zhang R, Yun WY, Moon IK (2011) Modeling and optimization of a container drayage problem with
resource constraints. Int J Prod Econ 133(1):351–359

123