Dynamic and Stochastic Models for the Allocation of Empty Containers

Author(s): Teodor Gabriel Crainic, Michel Gendreau and Pierre Dejax

Source: Operations Research , Jan. - Feb., 1993, Vol. 41, No. 1, Special Issue on Stochastic
and Dynamic Models in Transportation (Jan. - Feb., 1993), pp. 102-126
Published by: INFORMS

Stable URL: https://www.jstor.org/stable/171947

REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/171947?seq=1&cid=pdf-
reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Operations
Research

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 DYNAMIC AND STOCHASTIC MODELS FOR THE ALLOCATION
OF EMPTY CONTAINERS

TEODOR GABRIEL CRAINIC

UTnii ersite dui Qiuebec a M,Uontreal, M,Uontreal, Qucbec, Canacda

MICHEL GENDREAU
UTniversitcde MIontreal, AIontreal, Quebec, Canacda

PIERRE DEJAX
Ecole Cent rale Paris, Chatena.'-AIalabri', France
(Received September 1990; revisions received December 1991, May 1992, September 1992; accepted October 1992)

The empty container allocation problem occurs in the context of the management of the land distribution and
transportation operations of international maritime shipping companies. It involves dispatching empty containers of
various types in response to requests by export customers and repositioning other containers to storage depots or ports
in anticipation of future demands. We describe the problem and identify its basic structure and main characteristics. We
then introduce two dynamic deterministic formulations for the single and multicommodity cases, which offer a general
modeling framework for this class of problems, and which account for its specific characteristics: the space and time
dependency of events, substitutions among container types, relationships with partner companies, imports and exports,
massive equilibration flows, etc. Finally, we provide a mathematical formulation for handling, in the single commodity
case, the uncertainty of demand and supply data that is characteristic of container allocation and distribution problems.
Various modeling choices, data requirements, and algorithmic considerations related to the implementation of the
models are also discussed.

Since the 1970s, freight transportation activities To our knowledge, there does not exist any other
have seen many fundamental changes both in model specifically aimed at the empty container allo-
terms of technology and organization. In particular, cation problem that we study. In a recent survey of
one must note the increasing use of multimodal trans- the literature on fleet management models in freight
portation routes and of containerized transportation transportation, Dejax and Crainic (1987) noted that
practices. This is particularly true in the area of inter- relatively little effort was directed toward developing
national maritime shipping, where the maritime car- models aimed specifically at container transportation
rier normally owns the containers and ensures their issues, with the bulk of the work in the area being
multimodal transportation from the initial origin to dedicated to the maritime aspects of the problem.
their final consignee. Thus, Dejax and Crainic quote a few authors who
We are particularly interested in the planning of the addressed the problem of the empty container distri-
land transportation part of the international maritime bution as that of allocating containers available in
shipping of containers. This is an extremely complex surplus at a terminal, from earlier loaded shipments,
activity, especially if one aims to simultaneously to demanding terminals (e.g., a customer's warehouse,
optimize the cost and service aspects of the company's a rail yard, or a port), in preparation for subsequent
operations in a competitive environment. A funda- loaded movements. Antonisse (1988) is a recent exam-
mental component of this problem is concerned with ple of such efforts for the worldwide maritime con-
the allocation of empty containers to customers, in tainer fleet management problem.
preparation of future loaded export movements, and The state-of-the-art described by Dejax and Crainic
the repositioning of other empty containers to storage may be explained, at least partially, by the combined
depots or ports to be used to fulfill the next periods' effect of the relative complexity of the container land
orders (both expected and forecast). transportation issues and the opinion, prevalent until
Suibject classifications: Networks/graphs, mLllticommodity: transportation of empty containers. Transportation, freight: allocation of empty containers.
Transportation, network models: stochastic and dynamic flow models.
Area of review: DISTRIBUTION, TRANSPORTATION AND LOGISTICS (SPECIAL ISSUE ON STOCHASTIC AND DYNAMIC MODELS IN TRANSPORTATION).

Operations Research 00030-364X/93/4101-0102 $01.25

Vol. 4 1, No. 1, JanUary-FebrUary 1993 102 ?0 1993 Operations Research Society of America

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 Alodels for Allocating Empty Containers / 103

very recently in the industry, that the impact of land 1. PROBLEM CONTEXT
operations on the company's overall performance was Our research is motivated by the analysis of typical
marginal. This perception is, however, gradually land transportation problems encountered by inter-
evolving and the land portion of the container journey national shipping companies that operate maritime
is receiving increasing attention. This is especially true
and land networks of considerable, even continental,
in today's increasingly competitive environment scale. Crainic, Gendreau and Dejax (1990) describe
where growing customer demand is associated with the various operations and planning issues involved,
the search for cost efficient, high quality service oper-and present the methodological framework that we
ations. This view is similar to those of, among others, propose to address them. To facilitate the understand-
Jarke (1982), who reviews container management ing of the allocation model, this section is dedicated
problems and insists on the necessity for efficient to a brief overview of these issues and design
information and decision support systems for con- principles.
tainer transportation logistics, and Antush (1989), Arriving ships carry containers, which generally
who emphasizes the need for developing new infor- come in two sizes and about ten major types each,
mation systems and planning methodologies for con- loaded with imported goods, as well as empty con-
tainer transportation. tainers returning from previous exports. Loaded con-
The goal of this paper is to propose models for tainers must further be transported to their final
empty container allocation in a land distribution and destination, usually an industrial firm or a distribution
transportation system that address the complexity of center located somewhece on the continent. Ground
the specific problems which arise in this area, partic- transportation may be performed by using rail, truck,
ularly the space and time dependency of events, the barges and mixed modes. This latter option usually
uncertainty of supply and demand data, plus specific consists of an initial railway shipment to an inland
operational characteristics, such as substitutions railway terminal, followed by a final movement by
among container types, relationships with partner truck to the customer. The empty containers that are
companies, imports and exports, and massive equili- imported may either stay at the port for a while, or be
bration flows. The focus of the paper is on modeling immediately dispatched wherever they are needed for
issues; its main contribution is threefold: identification subsequent exports.
of the basic elements and structure of the container Once unloaded, empty containers have to be picked
distribution problem; development of a multicom- up at the customer's site and moved away. They may
modity model that integrates substitutions of con- either return to the port of origin, or be transported
tainer types; and to provide a mathematical to any other depot: port or inland terminal. This
formulation for handling the dynamic aspects and the
return shipment may, but more often than otherwise
uncertainties characteristic of container allocation and will not, be made using the same vehicle (truck, rail-
distribution problems. way car, etc.) that has been used for the transportation
The paper is organized as follows. In the first section, from the port to the customer. Similarly, exporting
we give a brief description of the issues related to firms need empty containers to be delivered to them,
planning container distribution and transportation either from the future export port, or from any other
systems, as well as of the general methodological location. Due to complex organizational considera-
framework we propose to address them. The land tions, few (Dejax, Crainic and Delorme 1988 report
container allocation problem, its main components some 3% for the case they studied) empty container
and characteristics are described in Section 2. movements are made directly from import to export
Section 3 is dedicated to the development of the customers. After loading, containers are transported
mathematical and network formulations of the deter- to the export port and are loaded on ships together
ministic, single commodity version of the model, with empty containers sent abroad to cope with the
while the multicommodity version is presented in worldwide imbalance in the supply/demand of certain
Section 4. The dynamic, stochastic single commodity container types.
formulation is treated in Section 5. In Section 6 we Note the importance of empty container move-
compare our modeling framework with the work of ments. First, every commercial (profitable) movement
other authors, while Section 7 is dedicated to consid- of a loaded container almost automatically generates
erations related to the practical implementation of the a nonprofitable empty container movement: Once
models. Our conclusions indicate future research loaded containers are delivered to a customer's site
directions in this area. and are unloaded, they normally have to be moved

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
104 / CRAINIC, GENDREAU AND DEJAX

away empty, while empty containers have to be deliv- Long-Term Forecasts

Strategic/Tactical Model
ered to exporting customers. Moreover, significant
regional imbalances between imports and exports
within the area served by the company (their ampli-
tude may vary according to the container type and
Depot Selection
the geographical subdivisions considered), as well as
Empty Balancing [_MIS|
opportunities for cost reductions due to mass trans- Customer to Depot Assignment

portation of containers over relatively long distances

(by railway, for example) result in land movements of
empty containers directly between depots: These are v r , _ Short-Term Forecasts

called balancingJlows. Empty Allocation Model .

Known Supplies/Demands

Thus, by their impact on the amplitude of the Ship Schedules

resource requirements and on the quality of service Costing Rules

offered to the company's customers, empty move-

Dispatch List of Empties
ments play a central role in the management of the
land distribution and transportation of containers. A
few figures illustrate and support this point (Dejax,
Loaded Movement Requests
Crainic and Delorme). A major European container
Routing Model Substitution Rules
shipping company operates worldwide maritime ship- Fleet Characteristics
Cost Formulas
ping lines to and from some twenty European ports,
while its land distribution and transportation network Container Routes
Vehicle Itineraries
mainly covers France, Germany, Holland, Belgium
and Luxembourg. In 1986, its operations involved
over 300,000 land container movements, which rep- Figure 1. Outline of the overall planning approach.
resented an estimated total distribution and transpor-
tation cost of some U.S. $50,000,000. Of these, 40%
were empty movements! approach. The topmost level concerns the strategic/
Several key questions have to be addressed by con- tactical planning of operations. The model is formu-
tainer shipping companies for the organization of their lated as a multimode, multicommodity location/
land distribution and transportation operations, ques- distribution problem with interdepot balancing
tions which clearly belong to several planning levels: requirements (Crainic, Dejax, and Delorme 1989).
strategic (e.g., the selection of inland terminals and Several specifically tailored algorithms (see Crainic,
the determination of their logistic zones), tactical (e.g., Gendreau and Dejax 1990 for references) have been
the evaluation of balancing flows), and operational developed, which efficiently make full use of the par-
(e.g., the dynamic allocation of containers). They can- ticular characteristics of the formulation. The main
not all be represented by a single model. Moreover, output of this phase is the strategic/tactical transpor-
they do not require the same data, nor the same tation plan, which specifies the network configuration
personnel and resources for their planning and exe- for the current planning period: the depots that are to
cution. Hierarchical relations exist among them and be used, the allocation of customers and customer
are reflected in the flow of information and decisions. zones to depots, and the main interdepot balancing
Yet, the modeling effort is significantly facilitated by flows. The allocation of customers to depots is speci-
the fact that container shipping companies do not fied by container type and direction of movement,
often need to build expensive inland terminals: They from customer to depot and from depot to customer.
may use already existing terminals, such as rail yards, In general, a customer may be assigned to several
for a much lower cost than the cost of building their depots, for each type of container and each direction
own facilities. This allows them to periodically re- of movement. Then, the strategic/tactical plan indi-
evaluate the decisions concerning the use of inland cates the relative importance (proportions) of these
terminals and the corresponding allocation of cus- flows.
tomers to depots. An integrated, multilevel method- Balancing traffic is also indicated by container type
ology, which reflects the observed hierarchy in the and direction of movement. Generally, however, these
decision process and the flow of information, may the results do not represent decisions to be directly carried
be designed to attempt to answer these questions. out in actual operations. Rather, being based upon
Figure 1 displays a schematic representation of this known customer agreements and medium-term

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 Models for Allocating Empty Containers / 105

demand and supply forecasts, they act as indicators of requires that a wide gamut of information be retrieved
the magnitude of the balancing operations required from the company's information system. In particular,
over the following periods, and may be used to nego- data on demand are to be given, including long and
tiate transportation tariffs (with rail companies, gen- short-term forecasts, as well as known supplies and
erally) for regular, consolidated movements of empty demands for the forthcoming periods. Equally essen-
containers (e.g., unit trains). These agreements will tial are the rules that specify how transportation means
then constrain the day-to-day operations. are chosen and priced, plus the company's own
The two other models correspond to the level of the accounting criteria concerning the various costs
operational planning of the company's activities. At related to owning and using containers. In the follow-
this level, one has to make sure that demand is satis- ing sections, we further discuss these information
fied, and the most effective routes and means of needs in relation to the requirements of the allocation
transportation are selected and used to achieve the model.
best possible economic and service performance in a
dynamic and stochastic environment. Ideally, one
2. PROBLEM DESCRIPTION
would like to develop a single mathematical model to
optimize the short-term land operations of the com- Several decisions are interwoven into the empty con-
pany to fully account for the interactions between the tainer allocation process. These decisions are mainly
various decisions to be made. Given the intrinsic concerned with wvhen to initiate a transportation oper-
complexity of the problem at-hand and the current ation, as well as with from wAhere to originate, or wvhere
state-of-the-art in OR, this is not feasible. This to terminate, the operation. (At this planning stage,
explains our decision to break down this large prob- howv to perform the movement-mode, route,
lem, and to develop two optimization models. etc.-is usually implicit in the spatial and temporal
The allocation model aims to determine the best definition of the movement.) Besides these movement
distribution of empty containers that satisfies both decisions, one is also concerned with operations that
known and forecast customer demands. To allow for affect the availability of containers for allocation:
the high level of uncertainty inherent in the real-time stock levels at depots, substitutions among containers
operation of a transportation system and for the of various types, additional resources (e.g., new or
impact of current decisions in the future, it should be borrowed containers), etc. These decisions require a
a stochastic, dynamic model that covers a planning great deal of information, including a variety of con-
horizon of one to two weeks. The objective of this siderations regarding the company's distribution and
paper is to present formulations for the allocation transportation system, the levels of supplies and
model. demands, the company's operating policies and rela-
The routing model strives to minimize the overall tions with its environment (competition, partner ship-
transportation cost of the loaded and empty containers ping companies, official rules and regulations,
from their origin to their destination. This is a com- agreements with carriers, etc.). Here, we describe the
plex and difficult problem, because it involves not main decisions, system components, and data consid-
only the determination of the best (multimodal) con- erations which make up the empty container alloca-
tainer itineraries, but also the construction of the tion process. The emphasis is on the elements that
routes for the vehicles required to transport them (for most strongly characterize (and increase the complex-
one or several modes). It is worthy to note that, ity of) the problem, and hence which have significantly
because the container shipping company does not influenced the design of the models.
own, in any significant measure, the required land The physical network which corresponds to the
transportation means, substantial savings can be company's distribution and transportation system is
achieved by using a routing model. This is particularly made up of ports, depots and customers, as well as
true for truck transportation where most savings are the transportation links connecting these elements.
achieved by generating multistop "circuit" routes (and Such a network is schematically drawn in Figure 2.
by suitably matching requested container movements In a land container distribution and transportation
to these routes) for which tariff reductions may be system, ports play a special role. They are the main
negotiated. The output of the routing model is a list entry and exit points of the system because most of
of movement orders which completely describe the the exchanges between the land transportation system
loaded and empty movements to be executed during and the international shipping network which con-
the next period (Dejax et al. 1992). tains it take place in ports. Furthermore, container
As illustrated in Figure 1, the use of these models shipping companies are primarily maritime industries-

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
106 / CRAINIC, GENDREAU AND DEJAX

h ~~~~~~~~~~~~~~~~~~~Pool

Po rt

Depot

XD Demand
customer

Supply S
customer

Figure 2. Possible container movements in a land transportation system.

hence, they operate significant parts of their land We distinguish between demand and supply cits-
network from ports, where they also maintain impor- tomners (although some may be both). A demand
tant stocks of empty containers. Consequently, customer requires that empty containers be delivered
we have decided to specifically include ports in the at its site, while a supply customer has empty con-
formulation. tainers, from a previously loaded import shipment,
Inland depots complete the backbone of the land that have to be taken away. Over a given planning
distribution system. Physically, these depots are often horizon, each customer may generate several supply
located in facilities offered by other transportation and demand requests, each of these is characterized
modes (e.g., rail yards); the container shipping com- by the number of containers requested, their
pany only rents the space it needs. Consequently, the type, temporal conditions (e.g., the latest date of deliv-
company may adjust its network configuration and ery), etc.
operations to the demand variations. Yet, for a given The location of a customer's site is known. For each
planning horizon, the port and depot network config- demand customer, the strategic/tactical plan has also
uration (depots j, k and port h of Figure 2) is fixed. determined the depots that may satisfy its requests for
Depots, both inland and in ports, serve as primary each type of container. In Figure 2, for example,
storage and transit locations for empty containers demand customer i may be served from both depots]j
(they may also serve as intermodal transfer points for and k, while the requests of customer i may only be
multimodal transportation of loaded containers; this satisfied out of depot k. For each request of a demand
aspect is, however, beyond the scope of this paper): customer, one also generally knows the delivery wvin-
Stocks of empty containers are built, demands for dowv within which containers have to be delivered. The
empties are satisfied by allocating containers out of delivery windows are mainly determined by the due
these facilities, empty containers available at cus- date in ports (determined by the ship schedules) and
tomers' sites are shipped to depots, and balancing by estimations of the time necessary to load and to
movements are initiated from there. The warehousing transport the containers. It may be worth noting that,
function and decisions are rather straightforward. To at the level of planning considered herein, the defini-
describe the movement decisions, however, we need tion of delivery windows implies that backorders are
first to introduce the customers of the system. not allowed in the system. Thus, the allocation

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 AModels for Allocating Empty Containers / 107

decision corresponds to determining how to satisfy tactical plan may be viewed as a set of operational
each customer demand request: what quantities, from policies in an ideal situation. In practice, however,
what depot, and when to initiate the movement. one has to deal with the uncertainties inherent in real
Supply customers may either send the empty con- operations, and has therefore to depart from this
tainers to one (or several) of their assigned depot(s) target. Hence, the exact amplitude of the interdepot
(depot k for customer s of Figure 2), or back to the balancing flows is a decision that belongs within the
port that originated the loaded movement (link (s, h) scope of the allocation model.
of Figure 2). This is an important issue, because of its Differences between supplies and demands of con-
influence on the inventory levels at depots and on the tainers of given types exist not only at the national
subsequent allocation decisions. Thus, the input of and regional levels, but also at the level of the inter-
loaded containers into the system and the moment national trade. Therefore, one notes the existence of
when these containers are again available to the com- external movements of enmpties, associated with the
pany have a significant impact on routing issues and input/output function of ports and represented by the
on the future supply of empty containers at depots. arrival (departure) of empty containers from (to) out-
The dyvnanmic and stochastic characteristics of the side the land system. Ships perform these movements,
problem are mostly evident in the definition of the together with the import and export of loaded con-
supplies and demands of the system. The empty con- tainers. At the decision level we consider, one has, for
tainer supply consists of all the containers which are all practical purposes, no possibility to modify the
made available for allocation at any given time. The ships' schedules. Hence, these schedules are assumed
supply for each type of container is made up of the to be known, as are the manifests of the ships speci-
stocks at depots, the containers present at supply fying the origin, destination, type and number of each
customer sites, and those entering the system via shipment of (loaded or empty) containers. Conse-
arriving ships in ports. Note, however, that before quently, in our formulation, ship arrivals and depar-
being available for allocation, most containers have to tures correspond to temporal entry and exit points of
be moved: from supply customers to some depot, or the system, determining when certain supplies become
between depots. Furthermore, not all customer sup- available, and the latest date when some demands
plies become available at the same moment, even if may be satisfied. We do not track containers outside
the loaded import containers are delivered from the the land transportation system.
same ship. This is mainly due to uneven transporta- Consideration also has to be given to the external
tion and unloading times. aftairs activities which impact upon the availability of
The empt.y container demand is indicated for empty containers, such as possible exchanges with
each customer, for each type of container, and partner container shipping companies, renting of
for each period of the planning horizon. In fact, the empty containers, and acquisition of new ones. The
same customer may require containers, of the same or wreckage of containers, either definitely (when they
of different types, at several periods. Therefore, one are scrapped and go out of the system) or temporarily
needs to distinguish between the physical customer (an estimation of the expected repair time is then
(its site) and its various demand requests. Most of needed), also falls into this category. We schematically
these demands correspond to identified customers, illustrate these possibilities by the connections
with known sites, and can be considered explicitly in between depots] and k (the link from port h is omitted
the allocation model. However, while some of these for clarity's sake) and the node pool of the network in
demands are completely determined ("sure" orders), Figure 2. Probability distributions based on historical
those of some other customers are known with less data may be used to represent the amplitude of these
certainty, especially as the period where the demand activities.
is expected is farther down the temporal axis. Finally, An important characteristic of the problem, and
it is also known that, in each region and for each type one which adds considerably to its degree of complex-
of container, there is a nonzero probability that an ity, is the possibility to substitute one type of container
unforecast demand will pop up in some future period. for another. The container substituttion rules are spe-
We associate this type of demand with the depots. cific for each pair of container types, and are not
As previously described, the strategic/tactical plan necessarily symmetric, neither in type (type A may be
mainly specifies the network configuration for the substituted for type B, but the reverse is not permit-
current planning period, and the restrictions imposed ted), nor in number (one type A container may be
upon the interdepot balancing flows by the agreements substituted for a container of type B, but it takes two
between carriers and the company. The strategic/ of these to replace one type A). Moreover, these rules

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
108 / CRAINIC, GENDREAU ANi) DEJAX

may have to be adjusted according to the loading/ dynamic, so this characteristic is not included in the
unloading facilities available at the customer's site section titles.
(one cannot substitute a 40-foot container for two
20-foot ones if the dock is not long enough), and the
3. DETERMINISTIC SINGLE COMMODITY
specific commodity that is to be transported (one
MODEL
cannot substitute two 20-foot containers for a 40-foot
one if anything over 20 feet long has to be trans- This section presents the modeling of the problem as
ported). The handling of substitutions is further com- a deterministic dynamic, single commodity formula-
plicated by the emergence of containers of dimensions tion. It is an idealized situation which nevertheless
that differ from the "traditional" 40 or 20 feet. allows a good presentation of the notation and math-
To conclude this problem description, we summa- ematical formulation. Most concepts and notation
rize the dynamic aspects that characterize the empty naturally generalize to the other models. Note that if
container allocation problem: the temporal character- this formulation is to be used per se, a number of
istics of the problem (of various transportation activ- additional questions should be addressed, including
ities, modes and itineraries, port and terminal the treatment of the end of the horizon. We discuss
operations and delays, etc.); the need to reposition this issue in Section 7.
(and substitute) containers to be able to satisfy future
3.1. The Mathematical Model
demands; that demands and supplies "appear" at var-
ious points in time and vary with time (even for the For a given planning horizon, the configuration of the
same "physical" customer); the existence of delivery port and depot network that is to be used by the
windows for demand customers; and the stocks of company is invariant with the specific period (day).
containers at depots used as buffers in the face of On the other hand, the set of active customers (i.e.,
varying supplies and demands. Concerning the sto- customers with outstanding requests for empties or
chastic aspect of the problem, note that it derives from with available empty containers) varies according to
the uncertainties of future supplies and demands. the time period. Furthermore, requests by the same
Hence, even though only immediate movements and firm with different due dates have to be treated sepa-
decisions are of interest and are transmitted to the rately. The set of identified customner requests is thus
routing module, a much longer planning horizon is specifically defined for each time period, according to
required to properly represent and assess their impact. the strategic/tactical plan and the forecast demand for
Finally, note that the substitution mechanism, the known customer firms. Note that these definitions
stocks at depots, and the possibility to access con- imply that two requests of the same customer are
tainers outside the system (buy, rent, etc.), form a very treated as two different items by the system. Let
rich recourse structure for the problem.
T: the length of the planning horizon, t = 1,
The modeling structure that we propose for this
problem may be described as a dynamic netwvork
H: the set of ports (harbors)-entry/exit points of the
model over a planning horizon of n days or periods,
system;
which is applied in a rollin1g horizon framework. To
D: the set of inland depots (does not include the port
make for a clearer and smoother presentation, we
depots);
introduce three formulations. First, a deterministic
I': the set of identified denman1d custonmer requests, to
model of the single commodity case (each container
be delivered at the latest in period t, t = 1,
type corresponds to a commodity), followed by the
2, ...,9 T;
deterministic formulation of the general multicom-
Ai: the delivery wiindowv of customer i, i E IP:
modity problem where container substitutions are
if Ai = 0, one has to deliver in period t
considered. Finally, we initiate the exploration of how
exclusively;
to handle uncertainties in the context of container
if Ai > 0, one has the option of delivering in all
distribution by introducing a stochastic formulation
pefiods, t, t - 1, . . . , t - Ai;
of the single commodity case. Given the current state-
S: the set of supply customer requests to be picked
of-the-art in the area, as well as the length of this
up in period 1, I = 1, 2, .. . , T.
paper, the stochastic formulation of the multicom-
modity problem clearly is beyond the scope here (for Note that we assign ports and nonport (inland)
a possible approach to this problem, see Crainic, depots to different sets; this makes for a cleaner nota-
Gendreau and Dejax 1991). All formulations are tion, especially for versions of the model presented

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 Models/fbr Allocating Empty Containers / 109

later in the paper. For plain descriptions, however, necessary time to journey between the two points. We
we will use the term depot to indicate any facility, assume all transit times to be independent of their
port or inland depot, where the warehousing function starting (t) and arriving (t') periods. We need the
is performed. following additional notation:
The network operations are driven by the amount
I:. the set of all demand customer requests i which
of empty container supply and demand associated
may be satisfied from depot j by a movement
with customers, depots and ports, for each one of the
which starts in period t and arrives in period t +
periods of the planning horizon. The supplies and
rTj; i.e., customer i belongs to Ij' if t' - Ai < t +
demands of empty containers associated with ports
rjj t' foriE It';jE D U H, t= 1,2,..., T;
reflect that, in general, any transportation system is
J': the set of all depots j which, in period t', may
embedded into a larger one: regions into states, states
be used to satisfy the demand customer request
into the world, etc. In particular, these values reflect
i, i c I'; the shipment starts in period t' and
the differences between supplies and demands at the
arrives in period I' + Tj, such that t - a, < t' +
level of the international trade. Let us then define:
Tji < t; _' C D U H, t' = 1, 2, . . ., T-
X'. the demand of the identified customer i, who J': the set of all depots j where supply (customer) s,
requires that containers be delivered at its site in s c S' may send in period t the empty containers
period t at the latest; i E I', t = 1, 2, . . .T; present on-site; in particular, the port of origin
X,: the export demand at portj in period t; it repre- of the loaded movement which generates the
sents the empty container requests which current empty supply, noted j(s), belongs to J.;
originate outside the study area; j E H, t = 1, J' C D U HII s C S', t = 1, 2, . . ., T;
2, ...,I T; SJ: the set of all supply customers s, s E S', t' =
Y',: the supply of containers that become available in t - Tr, from which depotj may receive containers
period t at supply customer s; s E S, t 1, in periodt; je DU H, t =1, 2, ..., T.
2,..., T;
We can now formally define the following decision
Yj: the import supply at port j in period t; it repre-
variables:
sents the input of empty containers from outside
the study area; j E H, t = 1, 2, . . ., T. v,i: the volume of containers allocated in period t
from depot jto customer i. The shipment arrives
Generally speaking, the decision variables represent
in period t'=t+r; jEDUH,iEI,t =1,
container movements through time and space. Those
2, ...,~ T;
movements that involve customers are identified by
v',: the volume of empty containers picked up in
the letter v, while wv is used for all others.
period t at supply customer s, and which arrive
In all formality, the variable which corresponds to
at depot j in period t + TIJ; s CS', j J',, t = 1,
the allocation decision from depot to customer is:
2, T;
v,': the volume of containers allocated from depot j s14,: the stock of empty containers available at depot
to customer i in period 1, and arriving at destina- j at the end of period t (the inventory flow from
tion in period 1'; i E 1j, j E DUll, t = 1, depot j in period t to depot j in period t + 1);
2, . .. I T, jC D U H, I = 1, 2., T;
ji: the flow of empty containers shipped from depot
to clearly specify the spatial-temporal occurrence that
j to depot k; shipment starts in period t and is
represents the identified customer. However, once the
completed in period t + Tr,;j E D U H, k E D U
starting time period of a movement is specified, the
H, t= 1,2,... T.
arriving time is completely determined by the corre-
sponding transit time, and hence it is not a decision. We are now ready to give the constraints of the
The same comments apply to all the decision variables deterministic, single commodity model.
defined therein, and we can thus use a simpler
3.1.1. Identified Customer Demand
notation.
The demand of an identified c stomer may be satis-
We use the notation T,P to identify the tr-ansit times
of the various movements, where the origin q and fiedthe from all the depots to w hich the customer is
destination 1i may represent a depot, a customer, or a
linked, by shipments which arrix e in any of the periods
that fall in the time interval defined by the customer's
port. These delays include the time required for load-
ing at origin and unloading at destination, plus the delivery window. In some cases, part of the identified

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
110 / CRAINIC, GENDREAU AND DEJAX

demand may already have been satisfied by shipments or newly bought ones) at depot j in period t; j E
initiated in periods prior to the beginning of the DUH,t= 1,2,..., T.
planning horizon, because either the customer's deliv-
ery window starts before this time, or these shipments We denote by it,,` the stock available at
have yet to arrive at the customer's site. As these D U H, at the beginning of the planning horizon. As
shipments correspond to known decisions, a constant previously, we use a constant to adjust the depot
value K' adjusts the identified demand for the plan- supply for the containers shipped to the depot before
ning horizon (of course, the definition of the set J' is the beginning of the current planning horizon:
modified accordingly):
K,: the volume of containers sent to depot j before
K': the volume of containers sent to customer i, i E I' the beginning of the current planning horizon
before the beginning of the current planning and that arrive in period t; j E D U H, t = 1,
horizon. 2, . .. , T.

The constraint then becomes:

Then

VI = xi - X'K'
11 1X ( 1 + VI,)
for all i E I', t = 1, 2,..., T. (1) / F k/ D)UII|Tkj=/-/ . SI

Note that, to ensure feasibility, we assume that for + b* + K'- W E V,

each identified customer demand i, iEI', not all the kFDUII . j

sets J' are empty. In operational terms, this implies

for all J D, t = 1,2, .., T. (3)
that there is always a depot sufficiently close by to
satisfy the demand of an identified customer. This
3.1.4. Port Supply, Demand and Stock
corresponds to the fact that the sales department of
the company should not accept an order which is
temporarily infeasible. 11X, 11X,. + E k( E t,+ E ,,

3.1.2. Customer Supply I?IA [l)UIIK=-

Empty containers, which are available at a customer

site, have to be moved out, and may be sent to the
II K [i fi4~ '~ / 11
poIt where the initial loaded movement originated, or v,i - XI + Y' + e'

to one (or several) of the depots to which the customer iE',

is assigned in the strategic/tactical plan:

forall jeH, t= 1, 2, ...,T, (4)
E l vs= Y' for all sE S', t= 1,2 ...,I T. (2) where

e': the unsatisfied export demand of empty containers

3.1.3. Stocks at Nonport Depots in port j, in period t (e' - 0); j C H, t = 1,
2, . .. , T.
For each depot, the stock of empty containers is
computed at the end of each time period. Out of the
The constraints that determine the stock levels at
depot are the volumes shipped to the other depots, to
poIt depots are very similar to those controlling stocks
poIts and to customers. The volume input during the
at nonport depots, with the exception that the net
period is made up of container traffic that left in
import/export demand at the poIt has to be met. We
previous periods from the other depots, the poIts and
identify the slack variable as the unsatisfied demand,
the supply customers, and that is arriving to the depot
mainly to account for the cost of not meeting the
in the current period, according to the corresponding
export demand, in some cases. This is the only
transportation times. This volume is added up to the
demand we assume possible to reject. Note that the
stock available at the beginning of the period, and to
presence of both the bh and e, variables, with
the containers which are brought in from outside the
the corresponding costs, allows one to perform a num-
system:
ber of interesting analyses on the tradeoffs between
bJ: the volume of containers borrowed (from partner the cost of not meeting demand versus the cost of
companies; it may also represent rented containers, bringing in new containers into the system.

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 Mocdels for Allocating Empty Containers / 111

3.1.5. Balancing Flows The total cost of operating the land distribution and
transpoitation system may then be written as:
V A _1< Uki

for all j E D U H, k E D U H,
E 4 E ( ECJiVi ~+ E Ci'kli Jk +6CJ1 Jj+ Pihi
i=1,2,..., T .E jDUH iE-I kEDUH

t = 1 ,2, ..., T, (5)

where
+ > jejC + > > C (6)
jEH sE^S jEJef J

,/(U,k): the lower (upper) bound on balancing move-

ment of containers initiated in period t and the model minimizes (6), subject to (1)-(5) and
from depot j to depot k; j, k E D U H, t = the
1, usual nonnegativity restrictions on the decision
2, . .. I T. variables.

These equations ensure that balancing movements 3.2. The Network Representation
are carried out according to the rules of the company,
The problem we study displays a natural network
especially its agreements with carriers. In particular,
structure, which is present in the deterministic model
the 1,k and U,I bounds are exogeneous to the model
of this section, as well as in the more complex for-
and reflect these policies and agreements.
mulations that follow. As a basis for these formula-
To write the objective function of the model, we
tions, we introduce the minimum cost network flow
need to define the cost criteria associated with each
model that corresponds to the single product deter-
one of the transportation and warehousing activities
ministic case. The transformation involves the con-
defined in the system. In all generality, the cost func-
struction of the network and the specification for each
tion associated with a specific activity (e.g., moving
arc of the flow variables, the corresponding flow upper
containers, in a specific time period, between a given
and lower bounds, and the cost function. The usual
customer and a paiticular port) could depend upon
flow conservation constraints may then be inferred
the values of all flow variables in the model. Yet, in
directly. However, when the network variables and
practice, such a broad definition is not required
constraints correspond rather straightforwardly to
because most empty traffic that involves customers
those of the previous model, we do not explicitly
takes place over relatively short distances and moves
rewrite them.
only a few containers at one time (this type of move-
The formulation is based on a multisource, multi-
ment is almost always truck based). We then assume
sink network G = (N, A), whose spatial topology
that for this type of activity, the cost function depends
conceptually unfolds in two-dimensional space: time
only on the intensity of the activity. Furthermore, to
spans the x axis, while space (location) unrolls on the
keep the presentation simple, we assume that all func-
tions are linear. We define the following notation: v axis. The time axis is divided into periods, and for
each period a network representation of the actual
c,i: the unit transportation cost from depot j to distribution and transpoitation system of the com-
demand customer i in period t j E D U H, I E pany unfolds on the space axis. Although the sets of
I', t =1, 2, ..,T; poits and depots used by the company do not vary
c',: the unit transportation cost from supply cus- within the planning horizon, this representation is
tomer s to depot j in period t; s E S',j E D U H, specific to each paiticular period, because the sets of
t=i 12,... identified supply and demand customer requests are
c,A. the unit transportation cost from
usually differentdepotj tooccurrence.
for each time depotDynamic
k in period t; j e D U H, k e D U H, t = 1, arcs link geographical representations along the
2,.. T; time axis, and these arcs may span more than
c,: the cost of holding one container at depot j in one time period, as we will show below.
period t; j E D U H, t = 1, 2..., T; The node set N encompasses elements which rep-
c: the cost of bringing in one container from outside resent the components of the distribution system:
the system at depot j in period t; j E D U H, t = poits, depots, and customers. The sets of nodes that
1, 2,...,I T; represent identified demand and supply customer
,: the unit penalty cost for not satisfying the exter- requests are specific to each time period; hence, we
nal demand of empty containers at poIt j in use the P and S' notation previously defined. The
periodt j&H,t=1,2,...,T. poits and depots used by the company are the same

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
112 / CRAINIC, GENDREAU AND DEJAX

for each specific period; hence a replication of these a': the dummy node that defines arcs which capture
elements (noted H' and D', respectively) is required the unsatisfied external demand at ports.
for each time occurrence of the spatial representation.
The definition and inclusion of the links of the
Note that in this representation, a large number of
network is mainly performed according to the com-
nodes appear as sources or sinks to account for the
pany's operating policies, the strategic/tactical plan,
containers that enter and exit the system; a limited
the ship schedules, and the temporal characteristics of
number of the X, Y and K constants associated with
supplies and demands. The plan specifies to which
such events are indicated in Figure 3.
depot(s) each customer may be linked, and the per-
Individual nodes are identified by the juxtaposition
mitted depot-to-depot movements. The ship schedules
of the letter identifying the network element (depot,
directly determine when exchanges with the outside
port, demand, or supply customer request) and the
world occur and influence, through the forecasting
time period. Thus, in the example shown in Figure 3,
process, when supplies and demands appear in the
k(t - 2) and i(t) correspond to the replication of depot
system. Additional arcs are included to model ware-
k in time period t - 2, and to demand customer
housing decisions, and to account for specific preoc-
request i, i E I', respectively.
cupations of day-to-day operations.
Finally, we add two super-source nodes to the set N
We illustrate the network requirements by using the
to account for the previously defined relationships
simple case of Figure 2. The resulting network is
with the external world:
displayed in Figure 3 for three time periods: one
a: the source for the new (borrowed from partner harbor and two inland depots, each replicated three
companies, rented, bought, etc.) containers intro- times, plus two identified demand customer request
duced into the system; nodes, i(t) and i(t - 1), and one identified supply

Space

;\ ~~~~~~~~K t-2X Kt /> ( _ _ _ -2) j(t-1)

At j(t) - K -
cx~~~~~~~~~~~~~~~~~~~~~~~~~

h(t-2)kt-) (- kt

h t-2 h t-1 i

t y t-2 -
x ~ ~ I

t-2 t-1 t Time

X Port j Depot C Demand X Supply
Customer Customer

Figure 3. Outline of the time-space network representation.

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 Models for Allocating Empty Containers / 113

customer request node s(t - 2). Note, however, that demand for empty containers that the system cannot
to keep the presentation clear, we have drawn only satisfy directly. Such arcs are introduced into the
representative arcs, instead of the complete network. network only in periods when external demands are
(When no risk of ambiguity exists, in the following we specified at port nodes. Arcs of the type (a, h(t)) and
only define the notation for period t.) (a, j(t)) (only four such arcs are drawn in Figure 3)
The arcs (h(t), k(t)), and (k(t), h(t)) are shown to capture the recourse to external sources of empty
illustrate the possibility of limited exchanges between containers in order to satisfy the total demand.
neighboring depots. These exchanges require small Although in theory the upper flow limit on these arcs
transportation times and involve relatively limited should be equal to infinity (to ensure feasibility), in
volumes of containers. Such movements may be ini- practice some sensible and practical finite value
tiated at any moment, for example, when the trans- bounds exist. Suitable total supplies have to be defined
portation cost from depot h to depot k plus the at nodes a and a'.
distribution cost starting from k is lower than the The arcs (s(t - 2), h(t - 2)) and (s(t - 2), k(t))
holding cost at h plus the corresponding distribution implement the requirement that empty containers
cost. Although the flow variables are similarly identi- have to be picked up at a (identified supply) custom-
fied, this type of movement is different from the er's location to be transported either to the port of
massive interdepot balancing traffic, represented origin of the initial loaded movement, or to one (or
by arcs spanning more than one period, such as several) of the designated depots. Note that arcs
arc (j(t - 2), h(t)) of Figure 3. between customers and depots span as many periods
The depot to customer arcs of type (j(t'), i(t)) as required by their respective transportation times.
illustrate that a customer demand may be met from Also, note that one may quite straightforwardly model
several close or not so close depots, and that the the possibility that containers may be left at customers'
requested containers may arrive at any moment dur- sites for later removal: One simply replicates the cor-
ing the specified delivery window. Thus, for example, responding s(t) nodes the number of periods the con-
the demand of customer i, which may be served from tainers may be left there, and adds customer holding
depots j and k, has to be delivered in period t with arcs (s(t), s(t + 1)) that capture the effect of keeping
a delivery window Ai = 2. Depot j may deliver in inventories at customers' sites. Of course, appropriate
less than a period; hence the three possible arcs, cost functions have to be defined because, otherwise,
(j(t - 2), i(t)), (,j(t - 1), i(t)), and (j(t), i(t)), are significant volumes of containers might be left at
included into the network. Depot k is farther away customers' locations to avoid inventory costs at
and thus cannot make a timely delivery by allocating depots. This approach assumes that one does not need
in period t; hence, link (k(t), i(t)) has not been to pick up all containers present at a given location at
included. On the other hand, node 1(t - 1) represents the same time.
the case of a customer with no delivery window Finally, Figure 3 illustrates the demands and sup-
(A, = 0) and only one depot that is sufficiently plies associated with the nodes of the network. Note
close to be able to satisfy the request (thus implying that port supplies and demands appear only at those
K&-' = 0). We thus include into the dynamic network periods when ships are scheduled. Otherwise, the sup-
a (j(t'), i(t)) depot to customer arc only if the corre- plies and demands of the network model are the same,
sponding transportation operation is permitted to and appear at the same periods as those of the previous
formulation. In particular, supplies and demands are
meet the demand in time, i.e., if t - Ai < t' + Tji < t.
Note that because the arc originates in the time period adjusted to account for the activities of previous
when the transportation operation is initiated from periods, when necessary.
the depot and ends in the latest delivery time period
for the customer, the number of periods the link spans
4. DETERMINISTIC MULTICOMMODITY
may be greater than the actual transportation time.
FORMULATION
We also illustrate in Figure 3 some of the holding
arcs of type (j(t), j(t + 1)), which model the manage- The possibility to substitute one type of container for
ment of inventories of empty containers at depots, another increases the flexibility of the distribution
as well as the two super-sources a and a'. This system and enhances its ability to adequately respond
enables us to illustrate the arcs used to model the to the customers' requests. Hence, a more general
various relationships with the outside world. Arc model of the container distribution problem is a
(a', h(t - 2)) supports the flow, represented by the multicommodity formulation, where each container
slack variable e5'2 , required to meet the external type is represented by a different commodity, and

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
114 / CRAINIC, GENDREAU AND DEJAX

where partial substitutions are allowed among the demands that have to be satisfied by 40' containers
various container types. and those which may be accommodated with either
Let P be the set of container types (products) with 40' or 20' ones.
p E P. Then, in general, the notation of the preceding To adequately represent the substitution mecha-
section generalizes rather straightforwardly to the nism and its consequences, we distinguish the con-
multicommodity case by inserting the product index tainers which physically are of a given type p, from
p before the time index t. For example, the identified the equivalent containers of type p which are to be
customer demand becomes: distributed to customers. This is tantamount to differ-
entiating the two main depot functions: managing
XI': the demand of identified customer i, who requires
stocks of containers of various types and allocating
that containers of type p be delivered at its site
containers to satisfy customer demands. As a model-
in period t at the latest; i E P", p E P, t = 1,
ing principle, substitutions occur only at depots. Fur-
2, . . . , T,
thermore, stocks of containers at depots, all interdepot
while the corresponding node set is written P"'. For traffic, as well as all exchanges with the outside world,
the sake of brevity, we do not explicitly redefine all are of true containers of the given type, while demands
the variables, constants and sets, nor do we write all are satisfied by using equivalent containers.
the equations previously described. We indicate here- In the network formulation, this approach is imple-
after only those elements that change or are new due mented by associating two nodes to each depot and
to the multiproduct nature of the problem. port:
The network G = (N, A) that underlies the deter-
IY'(H"'): the subset of nodes representing the repli-
ministic formulation is now a three-dimensional struc-
cation for container type p and period t of
ture: tiune on the x axis, space (location) on the y,? axis,
the physical locations (depots and ports)
and pr-oducts (container- ty,pes) on the axis. The
where stocks of empty containers are
product axis has as many units as the number of
managed; this corresponds to the ware-
container types defined in the problem. The network
housing function of the depot; p E P, t =
representation of the actual distribution and transpor-
i, 2,... I T;
tation system of the company, which unfolds on the
Y1 (H'"): the subset of nodes representing the repli-
space axis, is now specific to each particular (time,
cation for product p and period t of the
pr-oduct) occurrence, because the sets of supply and
source depots (ports), from where equiva-
identified demand customer requests usually vary with
lent empty containers of type p are allocated
the container type as well as the time period. Note
to customers; this corresponds to the allo-
that the sets of ports and depots used by the company
cation function of the depot; p E P, t = 1,
continue to represent the invariant part of the net-
2, . . . , T.
work. Dynamic arcs link geographical representations
along the time axis, while substitution arcs link the This definition of the network structure ensures that
various commodity networks for a given time period. substituted containers are used directly to satisfy
Substitution /actors are used to represent some of demands and do not end up in inventories at depots.
the technical restrictions related to using a container The customer allocation decision variable then
of a type different from the one required by the becomes:
customer. For any given pair (p, r) of products, the
substitution factor a,)r represents that more than one vPtj: the volume of equivalent containers of type p
container of type p might be required to replace one allocated in period t from depot j to customer i;
container of type r. It should be stressed that these j E D U H, i E IP', p E P, t = 1, 2, T;
factors do not correspond to direct volume relation-
while new decision variables have to be defined as:
ships between container types; rather they are the
inverse of some integer number: v',: the total volume of containers of type p in stock
at depot j, which will be allocated in period t as
a,p = (the number of type p containers needed to
equivalent containers of type p; j E D U H, p E
replace one type r container)-'.
P, t = 1, 2, . .. T;
This approach is general although its implementation iv,!': the number of empty containers of type p
may require a detailed definition of container types, substituted for containers of type r at depot j
and the specification of demands in appropriate units. in period t; j E D U H, p E P, r E P, I = 1,
For instance, one may have to differentiate between 2 . . . , T.

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 Alodels for Al,41ocating Empty Containers / 115

Space _ Pt

-j p (t-h1 \ jp(t)

vPt

j'p(t-1 ) j(

hp(t-1), II :p )
/ 5P(t-1)
4 Wkb ~ ~~~~~~~~~~~~~~it XPt Kt

pVt kp(t
sh

// ~ ~~~~P(t-2) ~~~~~k'p (t- 1) k'p(t)

sp(t-2)

jp(t-1

p(t-1)

t-2 t-1 t Time

> Port L] Depot Demand Supply

Customer Customer

Figure 4. Outline of a time-space container type layer.

Figures 4 and 5 illustrate these concepts. For clarity since substitute containers may have been used for
reasons, Figure 4 only displays part of the subnetwork the original loaded movement, technically, these flows
that corresponds to the space-time layer of product p. could be interpreted as representing traffic of equiva-
It also shows a few of the corresponding decision lent containers. This approach would rapidly put the
variables. The example network of Section 3 is also system into a chaotic state, as equiivalenit and ph.ysical
used here. Nodes are noted by the juxtaposition of the containers would mix into the depot inventories. An
letter that identifies the network component, the prod- elegant way of addressing this issue is based on the
uct, and the time period. observation that, since supply customers requests are
As illustrated in Figure 4, for each container type generated from actual movement orders for loaded
layer in the network and for each period, all flows containers, one knows exactly the container types
between warehousing nodes (depots and ports in H"' involved when these requests materialize. Hence, at
and D", respectively) represent movements of true generation time, the supplies at a customer's site may
containers. The flow on arc ((hp(t - 1), (kp(t - 1)) is be transformed into as many different supply cus-
a case in point. True containers also flow on the arcs tomer requests as the number of different container
from each warehousing depot to its corresponding types present at the customer's site. Consequently, the
allocation one, such as arc (( jp(t),j'p(t)). On the other supply customer to depot flows may be assumed to
hand, flows that originate from the allocation depot represent traffic of true, physical containers.
nodes, and that end up at demand customer request Permitted substitution activities are represented by
nodes, e.g., on arc (j'p(t), ip(t)), should be interpreted arcs linking different container type layers of the net-
as movements of equivalent containers. work. At a depot j, substitution arcs ( jp(t),j'r(t)) start
Movements that originate from supply customer at the warehousing node of one product (e.g., p) and
request nodes (e.g., on arcs (sp(t - 2), kp(t)) and point to the allocation node of another (e.g., r). Flow
(sp(t - 2), hp(t - 1))) warrant a special note. Indeed, variables i<" measure the volume of substitutions,

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
116 / CRAINIC, GENDREAU AND DEJAX

Container the containers that are used for substitutions

types

Jt = (A7) + EI( E v ?' + E '0

J/~~~~ '/ J1ttuS|r.=- E
+Mft + KlI)l E 7 l;yA-JI- E 't; (7)

vP! / A
k CI) Uh HrEP

for allj E D, p E P, t = 1, 2, ...,T, and

/00/ Prit5 a
1 SSi kC1)UZ/ | rki=1-I ' .E

for ale'l} P - ) + E l E + E Uv)

+ I-Yj) + Kj.1- ,v/\-VI2

keIDUII
_XX + j + ej (8)
I-P

for all j E H, p E P, t = 1, 2. T, respect

while a new set of constraints has to be defined to
represent the depot allocation balance.

4.1. Depot Allocation Balance

VJ/ + E a,,, J/ vJ' 0

for all JE D U H, pEP, t = 1, 2, ... T. (9)

modity layers at depotj in period t.
Most of the cost definitions of the previous sections
readily generalize to the multiproduct case, especially
since we assume linear cost functions. Define
which are governed by the substitution (gain) factors
a,,,. Upper bounds B,2" limit the amount of flow trans- c'l1: the substitution cost at depot j in period t of a
ferring from one commodity layer to another, thus container of type r by a container of type p;
modeling the actual restrictions present in company jEDUH, pEP rEP,t= 1, 2, ..., T.
operating policies.
The deterministic multiproduct model then
Figure 5 shows a number of such arcs which illus-
minimizes
trate some of the possible substitution cases. Note that
the jp(t) and j'p(t) nodes of F:igure 5 are the same as
those of Figure 4, while the three other node pairs
LJEDUII ~~I Cj7'jk,j+
represent the warehousing and allocation activities in
period t for three other products (1, r and g) at the t=12...Tp-P_jDul c i'p' , kED/?l+ ' U' 1IV'

same depot j. Figure 5 also displays a number of arc J1+ 'ii , Jil
+ Cl4 14Y,! JZT JI8 II
rP /
decision variables and substitution factors.
Three cases are illustrated. On the one hand, con-
tainers of types p and r may be substituted one for the + E c'4"e'`+ E ScivI/j (10)
J.clt ,,SEM'j,9, C
other; hence both arcs (jp(t), j'r(t)) and (jr(t), j'p(t))
exist. On the other hand, containers of types p and r subject to (1), (2), (7), (8), (9), and (5), to the upper
may be substituted for the type 1, but the converse bounds on the substitution variables, and the usual
operation is not allowed. Finally, no substitutions are nonnegativity restrictions on the decision variables.
permitted between containers of type g and any of the To solve this large, dynamic, linear, multicommod-
types p, r or 1, and no distribution arcs are defined. ity, minimum cost, generalized network flow problem,
Relatively few modifications have to be performed one could apply a standard network flow with gains
to the constraints of the single commodity model algorithm to the multilayer representation of the net-
besides the insertion of the product index p. Indeed, work, thus solving the problem as a single commodity
only (3) and (4) have to be modified to account for one. However, this approach does not account for the

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 ModelsJbrAllocatingEEmptyContainers / 117

special structure of the network. More precisely, it The internal stochastic demand is made up of uncer-
does not acknowledge that, were it not for the substi- tain requests associated with given customers. As in
tution arcs, the problem would decompose naturally the previous sections, each of these requests should
in a series of much smaller pure network problems. specify the customer location, the number of con-
In this context, decomposition methods that seek to tainers demanded, and the delivery window. Note
exploit this near-separability of the problem seem that, aside from "real" customers of the company,
quite attractive, and we intend to follow this approach. other events which have an impact on the stock of
containers, such as accidents temporarily or perma-
nently damaging containers, or requests from partner
5. STOCHASTIC SINGLE COMMODITY MODEL companies, are easy to model as demand customer
requests within this framework.
This section is dedicated to the development of a
In a similar fashion, the internal stochastic supply
formulation of the dynamic single commodity case
is made up of uncertain requests for picking up avail-
that takes into account the stochastic nature of some
able empty containers at customer sites in future
of its major components. We aim at a formulation
periods. Events such as returns of repaired or lent
that is both analytically tractable and numerically
containers may also be modeled as supply customer
solvable, and that parallels as closely as possible the
requests. It is important to note, however, that for
deterministic model of Section 3.
supply customer requests the uncertainty derives, in
In previous sections, we assumed all demands and
most cases, not from the randomness of the amount
supplies to be perfectly known throughout the plan-
of containers available for pickup, but rather from the
ning horizon. We now relax this assumption and
"when does that supply become available" type of
distinguish between two tvpes of demands and sup-
question. Indeed, most supply customer requests are
plies: on the one hand, deterministic demands and
generated based on the knowledge, gained from pre-
supplies resulting from firm requests made prior to
cise import port to customer movement orders, that a
the beginning of the planning horizon and, on the
given number of loaded containers are present, or are
other hand, stochastic demands and supplies corre-
arriving, at a customer's location. Hence, while the
sponding to forecast but uncertain future requests.
volume of supply available at a customer's location is
In the dynamic context of a planning horizon of
well known, the moment when the empty containers
T periods, we also assume that the first period of the
will be available becomes a random event. Indeed, if
horizon, for which decisions have to be implemented
loaded containers are delivered today, this does not
immediately, is fully deterministic. The uncertainties
necessarily mean that the containers will be emptied
thus appear only in periods 2 to T, and in a rolling
by, say, the following day. It may be tomorrow, or the
horizon mode of operation their main role is to serve
day after, or the day after that. One may infer, from
to ponder today's decisions by their estimated effect
historical data, the amount of variation particular to
on future ones. It is also important to note that, in
a given customer, variation that may increase when
this section, the durations of the various transporta-
the loaded containers have not yet arrived at their
tion activities are still assumed to be deterministic
destination.
parameters.
Finally, the external supplies and demands of empty
The remainder of the section is organized as follows.
containers associated with ports in future periods may
We first discuss in more detail the stochastic demands
be also viewed as random variables which reflect, in
and supplies. We then briefly review possible modeling
particular, the differences between supplies and
approaches to the problem, before presenting a tract-
demands at the level of the international trade.
able, two-stage restricted recourse model for it. This
is followed by a complete mathematical formulation
5.2 Possible Modeling Approaches
of the model, and by a discussion of algorithmic issues
relative to its solution. The deterministic model of Section 3 describes fairly
extensively the complex relationships existing between
5.1. Stochastic Demands and Supplies
the various decisions that have to be made when
The stochastic elements we consider are associated addressing the problem of optimally allocating con-
with the internal customer demands and supplies, as tainers over a large network. Ideally, we would like to
well as with the demands and supplies of empty con- come up with a stochastic model that accounts for the
tainers from outside the system which occur at exit uncertain nature of supplies and demands, while
and entry points, at ports in particular. retaining as much as possible of the interactions

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
118 / CRAINIC, GENDREAU AND DEJAX

between the decision variables. At the same time, we problem with (T - 1) periods for each scenario
would also like this model to be computationally considered.
tractable for large sized networks. As we shall see If we want to achieve a tractable model, we must
further on, these objectives are unfortunately contra- therefore accept much stricter restrictions with regard
dictory. The purpose of this section is to examine a to the recourse strategy. Hence, the approach that we
series of possible formulations in which different have decided to use is the following: All the "real"
tradeoffs are made between the complexity of the decisions are made at the beginning of the planning
interactions and the tractability of the models. horizon on the basis of the information available at
We first consider a direct adaptation of the model that time, and we let the system under study simply
of Section 3. Without going into all the intricacies of react in a prespecified fashion to adjust these decisions
defining exactly when the random variables that to the random events occurring in later periods
describe stochastic demands and supplies are realized, as these unfold (decisions in these periods are not
let us simply suppose that new information on these re-optimized). More formally, the original problem is
becomes available at the beginning of each period. divided into two stages. In the first stage, we consider
Under such an assumption, it should be possible to a deterministic optimization problem that encompas-
revise the allocation of containers at the beginning of ses the decisions made at the beginning of the planning
each period, and the natural way to state the problem horizon, in particular, the first-period decisions. In the
is as a T-stage stochastic program wvith fill netvwork second stage, we determine, for given values of the
recouirse (this indicates that all future decisions are re- first-stage variables and the random parameters, the
optimized in each period). Given the complex rela- impact on the overall objective of the corrective
tionships that link the decision variables in different actions undertaken in reaction to random events.
periods, the simplest way to write this model is by Such an approach obviously limits the range of actions
resorting to a scenario formulation, in which each in future periods, but it allows us to account for
scenario corresponds to a specific joint realization of uncertainties with a reasonable computational effort.
the random variables. Such a formulation yields a It should also be pointed out that if the model is
large mathematical program in which all decision properly defined, it should provide good conservative
variables and constraints have to be replicated for estimates of the costs to be incurred in the later periods
each scenario, while the replications of the decision (one could certainly do better on the average by re-
variables are linked by a set of nonanticipativitv con-optimizing decisions on the basis of the new infor-
straints which ensure that the decisions made in any mation which becomes available as time goes on).
stage t depend only on the values of the random
variables realized up to that time (see Wets 1974, and
5.3. The Two-Stage Restricted Recourse Model
Rockafellar and Wets 1991 for further details). Unfor-
tunately, the size of most container land distribution To fully specify the model, we must spell out how the
networks is such that the number of possible joint random parameters will be defined, the set of first-
realizations of the random demands and supplies is stage decision variables, and how the container allo-
extremely large. This implies that, even if only a small cation system will handle random events. Let us first
fraction of these possible realizations were sampled as examine in this context, the stochastic elements pre-
scenarios, the resulting mathematical program would viously described: uncertain customer and export
still be numerically intractable. supplies and demands.
One way to somewhat simplify the previous model With regard to uncertain customer requests, we face
is to assume that all the information regarding uncer- a difficult problem because these may, in some cases,
tain supplies and demands becomes available at the be satisfied from more than one depot. When such a
beginning of the second period (all random variables request is realized, we should thus have to decide by
are thus realized at that time). This yields a twvo-stage which depots it will be handled. In the case of
stochastic program wvith full netvwork recouirse in which demands, this is further compounded by the necessity
periods 2-T are lumped together in the second stage. to make a decision on the timing of the container
Such a simplification, however, does not significantly movements to the customer location, because of the
reduce the computational burden because each eval- flexibility afforded by the delivery windows. Such a
uation of the recourse function associated with the latitude of action is hardly compatible with the
second stage, for a given set of values of the first period requirements that we have to impose on the recourse
decision variables, still involves solving a deterministic strategy. To untangle the situation, we first force the

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 MIfodels for Allocating Empty Containers / 119

assignment of each customer to a single depot for its to the first period. Indeed, in real-life applications,
uncertain requests of a given type (i.e., supply or one would expect a fair portion of firm customer
demand) over the planning horizon. This is not, in requests to be known some periods in advance. For
general, such a big distortion, because most customers these, it seems natural to come up with firm allocation
will be assigned to a single depot for a given type of decisions at the outset of the planning horizon, and
container and request as a result of planning at the thus to include the decisions concerning the move-
strategic/tactical level. As for delivery windows, we ments used to satisfy them in the first-stage problem.
can adopt, in a stochastic setting, the conservative Second, it was seen in the deterministic model of
policy of shipping containers to customers as late as Section 3 that depot-to-depot movements are largely
possible, i.e., at the end of the forecast delivery win- ruled by exogenous constraints. In such a context, it
dows. Having already forced the assignment of cus- seems preferable to make all such movements deter-
tomers to a specified depot for uncertain demands, ministic, and thus to include them in the first-stage
the period in which a shipment should occur under problem.
this policy is easy to determine. To summarize, the first-stage problem covers all
Alternate hypotheses may be adopted, correspond- fi'rst period decisions, as well as all depot-to-depot
ing to different policies, a long as an uncertain supply flows and movements relative to identified certain
or demand is associated to only one depot and one customer requests in later periods. Given these, the
time period. However, once such an hypothesis is in second-stage model is simply a container inventory
effect, uncertain demands and supplies for each depot model that tracks down how the stocks of available
and time period can be aggregated into a single ran- containers at depots in periods 2-T will fluctuate in
dom variable for each type of movement: reaction to random supplies and demands. We now
specify exactly how this process will take place. (Note
x,: the random demand in the region of depot j in that, in the following, all Greek letters refer to random
periodt; jEDUH,t=2,..., T. variables.)
i': the random supply of containers available in Consider a given depotj at the beginning of a period
period t at depot j; j E D U H, t = 2, .. ., T. t (t > 2). In period t several events may occur that will
impact the level of the stock of containers on-hand.
Deriving the probability distributions of these vari-
ables requires careful analysis, but we defer the dis- 1. There is a net inflow or outflow of empty con-
cussion relative to this issue to Section 7, and simply tainers arising from the deterministic decisions of
assume for the time being that these distributions are the first-stage problem; let iWJ~ denote this inflow
known. (outflow, if negative).
Apart from customer requests, containers are 2. A random supply (if') is received from uncertain
received at ports from outside the system, and the supply customers.
export demands for empty containers are also made 3. There is a random demand (X) from uncertain
at these locations. To simplify the exposition, we demand customers.
assume that, except for the first period in which such
4. If depot j is a port, a random import supply (ij) is
supplies and demands should be known with certainty received, and there is a random export demand
(we retain the notation Xj and Yj for j E H), these (Qj) for empty containers.
are random variables denoted by:
In the deterministic models of the previous sections,
tj: the export demand at port j in period t; it was always assumed that customer demand had to
jEH, t= 2, .. ., T. be met, but that export demand could be left unsatis-
ij: the import supply at port j in period t; fied if it was more economical to do so. Such a latitude
j E H, t = 2, . . ., T. is not easy to accommodate in our restricted recourse
framework. Hence, in the present stochastic model,
Having spelled out the random parameters that
we make the simplifying assumption that all demands
describe uncertain events in our representation of an
have to be satisfied exactly. This allows us to aggregate
empty container distribution system, we now turn to
all random demands and supplies into a single random
the definition of the first-stage problem. First, as men-
variable for each depot and time period:
tioned, all decisions in the first period of the horizon
should be included in this stage. As for later periods, Aj: the net random demand (supply, if negative) at
note that certain customer requests are not restricted depot jinperiod t; jEDUH,t=2, .. ., T;

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
120 / CRAINIC, GENDREAU AND DEJAX

and we have 5.4 The Mathematical Formulation

We now formally write the model that has just been

A/tf_XJ - C, j E D described. In this formulation, we use the definitions
xj IXJ- J + $- - ;jl jE H.
and notation for the variables and parameters of
Note that we were able to perform this aggregation Section 3 and subsection 5.3 (recall, in particular, that
because of the simplifying assumptions made earlier, Latin letters refer to deterministic quantities and
and because we are dealing with a single type of Greek letters to stochastic ones).
container; this would not be possible if we had differ-
ent rules for handling export demand or several con- 5.4.1. The First-Stage Problem

tainer types with possible substitutions. Most of the constraints that link the variables of the
Given these elements, there are only two things that first-stage problem are identical to those of the deter-
one may do at a depot: borrow containers from exter- ministic model of Section 3. We recall them for
nal sources, and hold onto some of the containers for completeness' sake:
later periods (these two actions correspond, respec-
Identified Customer Demand
tively, to the b' and wji decision variables of the
deterministic formulation). To represent these activi-
ties, we introduce the following second-stage variables: i'Si jEJ5

for all i E I', t = 1, 2, .. . T. ( 11)

JW.: the stock of empty containers available at depot
j at the end of period t; j E D U H, t= Identified Customer Supply
2, ...,I T;
E v,;= for all s E S', t = 1, 2, . ,T. (12)
o3: the volume of containers borrowed from external
sources at depot j in period t; j E D U H, t=
2, . .. I T. Stocks at Nonport Depots in Period 1

We can then write the stock balance equation as WIV- W= J keuDllI

+ Ij Tkj=()
+ E Vl' + KJA
(9; = W( -I) + i + oil - A4-
- E wAj! - E vJi for all j E D. (13)
Note that since the values of the second-stage variables keDull iEl'

for any depot j in any period t 2 depend solely on

the inventory levels in previous periods, and thus on Port Supply, Demand and Stock in Period 1
vtj', the deterministic inventory at the end of period 1,
as well as on the net inflows (iiW) and net random
Wjy, =i5 +kEDullITkj=O
bj + E H)l keDU1l
+ KJ - , Wjk
demands (A) for periods t = 2, . .., T, the second-
stage is separable by depot, and the overall recourse
- E VJ'i-Xj' + Yj' + ej' foralljEH. (14)
function of the problem can be written as
JEDuIIQj((w4, ii ;), where Qj(ivt'],, W)j) is the Balancing
recourseFlows
function for depot j, while wi', represents the vector of
IJkI -4 wJ/; lJA for allj E D U H, k E D U H,
iiW' variables for periods t = 2, .. ., T.
Before closing this section and proceeding to the t= 1, 2, ..., T. (15)
formal description of the stochastic model, two
To these constraints, we add a set of new equations
remarks are in order. First, notice that we have lost
which allows for the computation of the auxiliary
the exact location of customer sites by performing the
variables W'J ;
aggregation of uncertain demands and supplies. It is
thus no longer possible to compute the exact cost of
Net Inflow (Outflow) From Deterministic Decisions
movements between these sites and their assigned
at Depots in Periods 2-T
depots. But since these movements are "forced" in the
model, they have no impact on the optimization; they
'r'''= E ( E kJ&i E
can therefore be dropped from the objective function. kCD te ullJITkj=1-1' V j
Second, readers familiar with stochastic programming
will have noticed that the recourse structure we have K+ jk KJl E
lEDUlII iElj
defined is not the "classical" simple recourse structure
with overage and underage variables. It is somewhat
for all j E D U H, t = 1, 2, ..., T. (16)
similar to the null recourse strategy introduced by
Powell (1988). Also, all variables are restricted to be nonnegative,

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 Mlodels]Jbr Allocating Emnptj C7ontainer.s / 121

with the exception of the iij's which are unrestricted where [x]+ denotes the positive part of x (i.e.,
in sign. max (0, x)), and &J(wvJ, -, I ) _i.
The objective function of the first-stage model is It is easy to verify that the depot recourse functions
now to minimize the sum of the costs associated with and, consequently, the global recourse function of the
the deterministic variables, plus the recourse function problem, are convex and piecewise linear for discrete
of the second-stage problem: Aii variables.
To solve the model, several approaches may be
considered. However, keeping in mind that, in most
E E (CjjIUj(i+ E Cy&'k"''jak+ EC/,SVU,s
1=]1 2,..., 7'jeD1)u iG1, keDull As E.S1 j.e,, typical applications, the size of the network is large
(even after performing a substantial aggregation of
+ E (cji,lj+6jbJ)+ E eej + E Qj( ii,, -') (17) demands and supplies), one should focus on tech-
jEDUH jEH jEDUH
niques which exploit the special structure of the prob-

5.4.2 The Second-Stage Problem lem. In our context, stochastic quasigradient methods,
which solve a sequence of subproblems associated with
As pointed out at the end of the previous subsection,
specific realizations of the random parameters (see
the second-stage problem is separable by depot. We
Ermoliev 1988 for further details), seem particularly
thus write a second-stage subproblem for each depot.
attractive because they allow us to exploit the network
For a given depot j, j E D U H, the second-stage
structure of the problem (each subproblem is a mini-
variables are the stocks of empty containers at the end
mum cost flow problem). We intend to further explore
of each period (wj), and the number of containers this avenue in the near future.
borrowed in each period (fj) for periods t = 2 . . . , T.
Aside from nonnegativity restrictions, there is a single
set of constraints linking these variables. 6. COMPARISON WITH OTHER FLEET
MANAGEMENT PROBLEMS
Depot Inventory Balance in Periods 2-T
The empty container allocation problem has received
(0j = (5t-1 )+"J +S for allt =2, . ..,T, ( 18) very little attention in the literature, and, to our
where wj --], by convention. knowledge, there does not exist any model for that
The objective function of the second-stage subprob- specific problem to which the models described in
lem for depot j, given values for vvw' and w-j, is to Sections 3, 4 and 5 could meaningfully be compared.
minimize the expected costs of holding and borrowing On the other hand, operational fleet management
containers: problems arising in other contexts (railroad freight
transportation and truckload trucking) have been
studied more extensively and a number of models
have been developed (see Dejax and Crainic 1987,

= E., minimumL E T + t35)1 Ai1}] (1 9) and Powell 1988 for detailed reviews on this topic).
The purpose of this section is to examine and discuss
the similarities and differences between our models
where i,, wl, and Aj represent the vectors of random
and some of those.
net demands (Ak), inventory levels (wj), and borrowing
levels (d,), respectively, for t = 2, . . ., T. The stochastic empty sailcar- distribution problem is
If we assume a reasonable cost structure for the probably the one which most closely resembles the

problem (for instance, nonincreasing borrowing empty container allocation problem. Apart from the
costs), it is never advantageous to borrow containers fact that it is usually formulated for a single type of
in a period while maintaining a positive inventory at railcar at a time (it is thus a single commodity prob-
the end of the same period. This allows us to determine lem), one has the same general structure of vehicles
the optimal values of the second-stage variables for (railcars, in this instance) becoming free wkhich must
any given realization of the random parameters be distributed to locations (yards) wkhere there is a
through a set of recursive relations: demand for empties. Among the models that wkere
proposed for this pr oblem. we mention those of
sf( 2: i / | I A,) [ f ( 1,tf Wil I A ,, /) + t X ' t - Mendiratta and Turnquist (1982). Jordan and
Turnquist (1983). and Beaujon and Turnquist (1991).
forallt= ..., T. (20)
The d! ynam'nic vehicle allocation (DVA) problem has
been defined and studied bv Povell. Sheffi and Thiriez
( j~ ( 1 ti,) = X| -t i)[ vi , ( u Q f all ) - ]
(1984). Powell (1986. 1987. 1988). Frantzeskakis and
Powell (1990). and Powkell and Frantzeskakis (1993)

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
122 / CRAINIC, GENDREAU AND DEJAX

in the context of truckload trucking. In this problem, Turnquist 1982, Powell 1986, 1987, 1988) or back-
one has to allocate available trucks to potential loads ordered with a unit penalty per period of delay (Jordan
over a time-space network. Its general structure is thus and Turnquist 1983, Beaujon and Turnquist 1991).
similar to that of the container allocation and the Note also that in some of the DVA models it is possible
railcar distribution problems, but several elements to turn down globally unprofitable loads, even if
(most notably, the high level of uncertainty regarding trucks are available to move them. Consequently, the
future demands and the range of options open to the empty container allocation models are stated as cost
decision maker) give the DVA a flavor of its own. minimization problems, while all models which allow
Let us now examine the main aspects in which our for demand to be lost or turned down must be for-
container allocation models differ from the aforemen- mulated as profit maximization problems.
tioned ones. It is also important to remark that in models for
other transportation modes the exact location of
6.1. Multicommodity Versus Single Commodity demand requests is not specified: These occur at rail-
The existence of several container types and complex yards or in cities which are permanent nodes in the
substitution rules is one of the factors which most network structure. Hence, the use of a specific location

distinguishes the container allocation problem from of identified customer requests, and of a variable
dynamic vehicle allocation problems for other trans- network structure within the allocation process itself,
portation modes. As a consequence, the notion of is new.
eqluivalent container had to be introduced into the
multicommodity formulation. 6.4. Particular Characteristics

Under this label, we include the rules that constrain

6.2. Empty Only Versus Combined Empty and
the balancing flows, the external demands and sup-
Loaded Movements
plies of empties, and the relationships with partner
With regard to the allocation process itself, only empty
shipping companies, which obviously are all specific
movements appear as decision variables in the con- features of the container allocation problem, and are
tainer allocation models (however, as indicated in not found in any of the other problems.
Section 1, loaded movements are accounted for in the Some of these features had an important impact on
overall planning system and in the routing model), the formulations proposed. In the case of the stochas-
while in many of the previous models (Powell 1986, tic, single-commodity problem, for instance, the pos-
1987, 1988, Beaujon and Turnquist 199 1) both empty sibility of bringing in containers from outside the
and loaded movements are considered. This stems system (e.g., by borrowing from partner companies)
from the fact that loaded containers are shipped out- allowed us to impose deterministic depot-to-depot
side of the land network (they effectively "disappear" flows in all periods, even though empty container
for all planning purposes), while, in the other cases, inventories are random variables. This is in stark
the loaded vehicles remain inside the boundaries of contrast with the models proposed by Jordan and
the network to become available again for allocation Turnquist (1983) and Powell (1986). In these models,
within a short span of time. The same vehicle may which are also based on the null recourse assumption
thus have to be allocated several times within the of holding in stock vehicles that are not required in a
planning horizon, which is practically never the case given period, the stochastic inventories lead to sto-
in the land container distribution context. An unex- chastic flows between depots. This is because, without
pected consequence of this is that current allocation the possibility of introducing new vehicles in the sys-
decisions have a lesser impact on the future state of tem, one cannot make firm decisions on future move-
the system under consideration. We hope that this ments, since these could turn out to be infeasible. In
may reduce the feedback distortions caused by the such a situation, the best that one can do at the
limited planning horizon, thus allowing us to work beginning of the planning horizon is to allocate frac-
with a shorter horizon. tions of the available inventories to specific move-
ments, which implies stochastic flows.
6.3. Demand Handling
In summary, it can be said that while the empty
In the container allocation problem, all customer container allocation problem presents a general struc-
demand must absolutely be met within specified deliv-ture which is similar to that of other operational fleet
ery windows. In the other models, demands occur at management problems, the models described in this
specific instants in time. Requests which cannot be paper differ substantially from previous models pro-
immediately satisfied are either lost (Mendiratta and posed for these problems.

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 M f odels for Allocating Ems1pty Conitainier-s / 123

7. IMPLEMENTATION ISSUES interdepot balancing arcs, while the regional subnet-

works are suitably aggregated. Particular care has to
Although the main objective of this paper is the pre-
be taken concerning the proper representation of sup-
sentation of formulations of empty container
plies and demands at the nodes of this network, but
allocation models, several issues related to their imple-
the present allocation modeling framework extends
mentation and use in actual practice are relevant and
without much effort to the bilevel case.
are discussed below. We do not aim to be exhaustive.
Other types of organization may be imagined and
We think, however, that presenting and discussing
may present a research interest. For example. one may
these items will enhance the understanding of the
contemplate the case where each regional authority
modeling framework and will put our research effort
operates nearly autonomously, while the central
into its proper perspective.
authority is only concerned with the relationships with
7.1. Modeling Context the external world. A possible modeling avenue would
be to treat the demand/supply balancing problem as
Several extensions may be envisioned to our modeling
an n-person cooperative game. Yet, as we are not
framework. Some of these are rather obvious, even
aware of any practical applications of such a model,
though of significant interest for actual applications.
we do not plan to further explore this alternative in
Minimum inventory levels at depots and contractual
the near future.
obligations to send empty containers from customer
sites back to the original port, are two examples of
7.2. Data Requirements
this case. Some others, such as nonlinearities in the
objective function, are more theoretically interesting, An empty container allocation model, similarly to any
though with the current state of the available data no other operational model of a large and complex orga-
practical application may yet be contemplated. nization, requires, for its proper use in an application
Our interest in this section does not lie, however, environment, that abundant data be readily available.
with elaborating on this type of model extension. This information is varied in nature and may be used
Rather, we are concerned with the organizational either directly (e.g., demand figures) or indirectly (e.g.,
environment where our models may be applied and previous routing decisions and their cost implications)
the variants thereof. It is clear that the present frame- to infer correct approximate figures and formulas, and
work corresponds to a centralized organization, where to update them every time the horizon is advanced.
decisions concerning the allocation of empty con- Thus, the models have been designed as part of a
tainers over the entire territory are taken by a unique decision support system for the short-term land trans-
authority which considers the network as a whole. portation planning of loaded and empty containers
This is indeed a widespread organizational scheme. (see Figure 1). and their appropriate implementation
Alternate types of organization are, however, equally assumes the existence of a reliable information system.
possible. To discuss the definition, origin, evaluation/approxi-
One such alternative corresponds to a bilevel struc- mation and update procedures for all data items
ture, made up of a central authority and several required by an allocation model is clearly beyond the
regional ones. Here, briefly, the territory is divided scope of this paper. Hereafter, we only mention some
into several regions, each of which is managed by its of the most important issues related to specifying
own authority. The latter controls the operations demands, supplies and costs.
inside its region, as well as the movements with origins Accurate and constantly updated forecasts of the
inside the region but with destinations outside itssupply and denmand of empty containers are critical
for a successful implementation of an allocation
borders. The central authority performs the strategic/
tactical planning and coordinates the interregional model. Preliminarv studies (Gendreau et al. 199 1)
exchanges, especially the massive balancing flows and have shown that simple forecasting methods cannot
the relationships with the outside world. capture the complexity of the demand mechanism at
The modeling framework we have presented offers work and similar conclusions can probably be drawn
an adequate representation for this case. The variant with regard to the supplies. A more careful analysis
for a regional authority is a straightforward applica- reveals that observed demands and supplies are indeed
tion, where balancing movements are treated some- the result of the superposition of various processes
what similarly to port imports and exports. The with widely different characteristics. To obtain good
central authority model is principally characterized forecasts,
by it is thus necessary to identify as clearly as
its representation of the network: The entry and exit possible these processes, to monitor them closely
poiInts of the system are emphasized, as well as the through the management information system of the

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
124 / CRAINIC, GENDREAU AND DEJAX

company, to derive good individual forecasts by pro- 7.3. Ending the Planning Horizon
cess, and, finally, to merge all these forecasts into
The models of the previous sections cover only a
specific probability distributions.
planning horizon of T periods, while the operations
We use linear cost functions for all activities consid-
of the company will extend in the future for a much
ered in the allocation models: holding, substitution,
longer span of time. This situation, which is encoun-
borrowing, transport, etc. This is a reasonable hypoth-
tered in almost all dynamic models, requires careful
esis for an operational model, because most cost non-
scrutiny because one would like to minimize (hope-
linearities, such as potential economies of scale offered
fully nullify) the distortions created by the use of a
by certain transportation modes, have been considered
finite horizon model to represent, for all practical
by the strategic/tactical model. Besides, it reflects cur-
purposes, an infinite-horizon reality.
rent industrial practice and data availability. Note also
Obviously, using the models in a rolling horizon
that handling costs, which correspond to the various
mode helps to mitigate the negative effects of this
operations and intermodal transshipment of con-
situation, but it does not by itself eliminate them. It
tainers at ports, depots and customer sites, do not
is well known that, for certain types of models, end-
appear in our formulations because the planning of
of-horizon distortions will tend to feed back on the
these operations is not part of the allocation decision
values of the decision variables, even in the first period
process.
(see, for instance, Hughes and Powell 1988, for a more
Container holding and slubstitiution costs at depots
detailed discussion on this topic). It is thus very impor-
are generally well known. On the other hand, even if
tant to carefully select the length of the planning
transportation costs represent the most significant part
horizon and the end-of-horizon conditions to obtain
of the total system costs, one has a less secure knowl-
valid solutions.
edge of these, because maritime container shipping With regard to the length of the planning horizon,
companies do not usually own land transportation
two key factors have to be taken into account. One
means, and must therefore contract the services of
must first be aware that for any real-life application
carriers of various modes. In the case of trucking, they
of the models, the number of decision variables in any
use the services of a large number of local companies,
period will be fairly large. This implies that the length
and the cost of transportation is often determined on
of the planning horizon must be limited to some
a case-by-case basis, according to many factors, such
reasonable value (say between 10 and 20 periods), if
as the geographical area, the technical association of
the models are to be computationally tractable.
truck size-type and container type, the distance to
The second factor to be considered is the informa-
travel, and the insurance of return freight or paid
tion available on the future supply and demand of
return. For rail shipments, they call for the services of
empty containers. As we go down the time axis, we
national railway companies, which propose an equally
move from an area where a large portion of the supply
complex tariff structure, where the cost of transpor-
and demand requests can be deemed as certain, to
tation depends on the origin and destination zones,
one where everything (or almost) is uncertain. Ideally,
the type, size, weight and number of containers to be
the length of the planning horizon should be such as
shipped together, etc.
to cover certain demands and supplies (identified cus-
We have initiated an investigation aimed at deter-
tomer requests and known import and export
mining formulas for modal unit transportation costs,
movements).
which take into account these particularities, and
If such an ideal planning horizon is possible, one
which are based on historical data. The initial results
may then specify fairly simple end-of-horizon condi-
are extremely encouraging (Vasseur 1989, Crainic,
tions to prevent unrealistic solutions. These conditions
Dejax and Gendreau 1989). However, for any given
should be such as to force reasonable values for the
origin-destination pair, the particular transportation
empty container stocks in each depot at the end of
arrangement may vary from period-to-period, espe-
period T. (These values may be derived from historical
cially when trucking is involved, and this routing
experience and practical operational constraints, or
information may not be available at allocation time.
from an empty container inventory model.) In deter-
Therefore, we plan to integrate a cost update
ministic models, this can be accomplished by impos-
mechanism to the decision support system, to re-
ing lower and, eventually, upper bounds on these
estimate the unit costs prior to the advancement of
stocks:
the planning horizon, based on the actual routing
decisions. /,C - u4j' for allj E- D U H, p E P.

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
 Models for Allocating Empty Containers / 125

One could also include a salage value term in the commodity case, the uncertainty of demand and sup-
holding cost functions for period T to account for the ply data that are characteristic of container allocation
marginal value of the various container types at each and distribution problems.
depot (Hughes and Powell). This last option could To our knowledge, these are the first models specif-
also be implemented in the stochastic formulation. ically aimed at the empty container land allocation
Obviously, in any practical application, some experi- problem. We have attempted, therefore, to explore, as
mentation with possible values for these bounds and comprehensively as possible, the various modeling
salvage terms would be necessary to ascertain that the issues related to this problem, while proposing for-
chosen values are indeed correct. mulations that are computationally tractable. This has
When no feasible ideal horizon exists, the specifi- allowed us to highlight the originality of the problem
cation of the end-of-horizon conditions becomes a and to explain the differences which may be observed
much more difficult task. In such a situation, one first
between our modeling framework and formulations
has to deal with split delivery windows, i.e., delivery for similar allocation problems arising in the context
windows which fall partly within the planning horizon of other transportation modes.
and partly outside of it. Deciding whether or not the We have sketched some of the most promising
corresponding demands should be included in the avenues for the development of efficient solution tech-
planning horizon is not obvious. In most cases, one niques for the proposed models. This, together with a
would probably be well advised to exclude them and continuing exploration of richer stochastic formula-
to slightly adjust the lower bounds on the empty tions, especially for the multicommodity case, consti-
container stocks at the depots from which these tutes the next stage of our research program. We also
demands should be satisfied later on (this is known plan to continue the investigation of the loaded and
from the strategic/tactical plan). Similar remarks empty container routing problem and that of the
apply to known import supplies and export demands container planning system required for an efficient,
occurring after the last period of the horizon. actual implementation of the models.
In summary, we stress that there is no cut-and-dry
answer to the problems created by the necessity to
ACKNOWLEDGMENT
work with a limited planning horizon. The modeling
framework offers a number of means (bounds, salvage The authors wish to thank Professor Warren B.
values) which can be used to tackle these problems, Powell, co-editor of this special issue, the associate
but the specific solution to implement in any practical editor, and the referees for their thorough review of
application of the model is highly dependent upon the the manuscript; their suggestions and comments
situation at-hand. helped us write a better paper. The authors are also
grateful to the Canadian Natural Sciences and
Engineering Research Council, to the Fonds F.C.A.R.
CONCLUSIONS
of the Province of Quebec, and to the Programme de
We have described the problem of allocating empty Cooperation France-Quebec, which provided the
containers in a land distribution and transportation financial support for this research.
system, and have identified its basic structure and
main characteristics. We have introduced two
REFERENCES
dynamic deterministic formulations for the single and
multicommodity cases, respectively, which, although ANTONISSE, K. 1988. A Decision Support System for
based on actual experience with a large container Container Replenishment. TIMS XXVIII/EURO
shipping company, offer a general modeling frame- IX Joint Meeting, Paris, July 6-8.
work for this class of problems. This framework ANTUSH, R. M. 1989. Research Issues at Sealand for
reflects the operational and planning complexity of the 1990's. ORSA/TIMS Joint National Meeting,
New York, October 16-18.
the land component of the international maritime
BEAUJON, G. J., AND M. A. TURNQUIST. 1991. A Model
shipping of containers, particularly the space and time
for Fleet Sizing and Vehicle Allocation. Tr-ans. Sci.
dependency of events, as well as specific operational
25, 19-45.
characteristics, such as substitutions, relationships CRAINIC, T. G., P. J. DEJAX AND L. DELORME. 1989.
with partner companies, imports and exports, and Models for Multimode Multicommodity Location
massive equilibration flows. Finally, we provided a Problems With Interdepot Balancing Requirements.
mathematical formulation for handling, in the single- Ann. Opns. Res. 18, 279-302.

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms
126 / CRAINIC, GENDREAU AND DEJAX

CRAINiC, T. G., P. J. DEJAX AND M. GENDREAU. 1989. HUGHES, R. E., AND W. B. POWELL. 1988. Mitigating
Transport Terreste Multimodal de Conteneurs: Pro- End Effects in the Dynamic Vehicle Allocation
bl&matique et Modeles. In Logistics. Production, Model. Mgm. Sci. 34, 859-879.
Distribution, Transport/Manufiacturing, Distribut-
JARKE, M. 1982. Developing Decision Support Systems:
tion, Tr-anspor-tation. Proceedings of the Joint Con- A Container Management Example. Int. J. Policj
ference on the Practice and Theory of Operations Anal. Inform. Syst. 6, 351-372.
Management and 4es Journ&es Francophones sur la JORDAN, W. C., AND M. A. TURNQUIST. 1983. A Sto-
Logistique et les Transports, AFCET, Paris, chastic, Dynamic Network Model for Railroad Car
105-113. Distribution. Tr-ans. Sci. 17, 123-145.
CRAINiC, T. G., M. GENDREAU AND P. J. DEJAX. 1990. MENDIRATTA, V. B., AND M. A. TURNQUIST. 1982. A
Modelling the Container Fleet Management Prob- Model for the Management of Empty Freight Cars.
lem Using a Stochastic Dynamic Approach. Tr-ans. Res. Rec. 838, 50-55.
Oper-ational Research '90. H. E. Bradley (ed.). POWELL, W. B. 1986. A Stochastic Model of the Dynamic
Pergamon Press, New York, 473-486. Vehicle Allocation Problem. Tr-ans. Sci. 20,
CRAINiC, T. G., M. GENDREAU AND P. J. DEJAX. 1991. 117-129.
A Dynamic Stochastic Model for the Allocation of POWELL, W. B. 1987. An Operational Planning Model
Empty Containers. Publication #713, Centre de for the Dynamic Vehicle Allocation Problem With
Recherche sur les Transports, University of Uncertain Demands. Tr-ans. Res. 21B, 217-232.
Montreal. POWELL, W. B. 1988. A Comparative Review of Alter-
DEJAX, P. J., F. BENAMAR, T. G. CRAINIC AND M. native Algorithms for the Dynamic Vehicle Alloca-
GENDREAU. 1992. Short Term Container Fleet Man- tion Problem. In Vehicle Routing. Methods and
agement: Issues, Models and Tools. In Pr-oceedings Studies, B. L. Golden, and A. A. Assad (eds.). North-
of The 6th WRorld Conference on Tr-anspor-t Resear-ch, Holland, Amsterdam, 249-291.
Lyon, France, June 29-July 3. POWELL, W. B., Y. SHEFFI AND S. THIRIEZ. 1984.
DEJAX, P. J., AND T. G. CRAINIC. 1987. A Review of The Dynamic Vehicle Allocation Problem With
Empty Flows and Fleet Management Models in Uncertain Demands. In 9th Int. S.yrnp. on Tr-ans-
Freight Transportation. Tr-ans. Sci. 21, 227-247. porttation and Ti-affic Theory, J. Volmuller and
DEJAX, P. J., T. G. CRAINIC AND L. DELORME. 1988. R. Hamerslag (eds.). VNU Science Press, Utrecht,
Strategic/Tactical Planning of Container Land The Netherlands.
Transportation Systems. TIMS XXVIII/EURO IX POWELL, W. B., AND L. F. FRANTZESKAKIS. 1992.
Joint Meeting, Paris, July 6-8. Restricted Recourse Strategies for Stochastic,
ERMOLIEV, Y. 1988. Stochastic Quasigradient Methods. Dynamic Networks. Tr-ans. Sci. (to appear).
In Numerical Techniques for Stochastic Optimiza- ROCKAFELLAR, T. R., AND R. J. WETS. 1991. Scenarios
tion. Y. Ermoliev and R. Wets (eds.). Springer and Policy Aggregation in Optimization Under
Verlag, Berlin. Uncertainty. Math. Opns. Res. 16, 119-147.
FRANTZESKAKiS, L. F., AND W. B. POWELL. 1990. A VASSEUR, C. 1989. Les Couts du Transport Ferroviaire
Successive Linear Approximation Procedure for Sto- par Conteneurs en France: Etude de Cas. Publication
chastic, Dynamic Vehicle Allocation Problems. #650, Centre de Recherche sur les Transports,
Tr-ans. Sci. 24, 40-57. University of Montreal.
GENDREAU, M., T. G. CRAINIC, P. DEJAX AND H. WETS, R. 1974. Stochastic Programs With Fixed
STEFFAN. 1991. Forecasting Short-Term Demand for Resources: The Equivalent Deterministic Problem.
Empty Containers: A Case Study. Tr-ans. Res. Rec. SIAM Rev. 16, 309-339.
1328, 98-103.

This content downloaded from

129.97.193.44 on Mon, 13 May 2024 02:48:42 +00:00
All use subject to https://about.jstor.org/terms