You are on page 1of 25

This article was downloaded by: [University of Wisconsin-Milwaukee]

On: 17 October 2014, At: 06:28


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Production Research


Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tprs20

Interconnected logistic networks and protocols:


simulation-based efficiency assessment
a a a b b
Rochdi Sarraj , Eric Ballot , Shenle Pan , Driss Hakimi & Benoit Montreuil
a
Centre de Gestion Scientifique (CGS), Mines ParisTech, Paris, France
b
Cirrelt, Université Laval, Québec, Canada
Published online: 09 Dec 2013.

To cite this article: Rochdi Sarraj, Eric Ballot, Shenle Pan, Driss Hakimi & Benoit Montreuil (2014) Interconnected logistic
networks and protocols: simulation-based efficiency assessment, International Journal of Production Research, 52:11,
3185-3208, DOI: 10.1080/00207543.2013.865853

To link to this article: http://dx.doi.org/10.1080/00207543.2013.865853

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
International Journal of Production Research, 2014
Vol. 52, No. 11, 3185–3208, http://dx.doi.org/10.1080/00207543.2013.865853

Interconnected logistic networks and protocols: simulation-based efficiency assessment


Rochdi Sarraja, Eric Ballota*, Shenle Pana, Driss Hakimib and Benoit Montreuilb
a
Centre de Gestion Scientifique (CGS), Mines ParisTech, Paris, France; bCirrelt, Université Laval, Québec, Canada
(Received 3 May 2013; accepted 5 November 2013)

Logistic networks intensely use means of transportation and storage facilities to deliver goods. However, these logistic
networks are still poorly interconnected and this fragmentation is responsible for a lack of consolidation and thus effi-
ciency. To cope with the seeming contradiction of just-in-time deliveries and challenging emissions targets, a major
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

improvement in supply networks is sought here. This new organisation is based on the universal interconnection of
logistics services, namely a Physical Internet where goods travel in modular containers for the sake of interconnection in
open networks. If from a logical point of view, merging container flows should improve efficiency, no demonstration of
its potential has been carried out prior to the here reported research. To reach this potentiality assessment goal, we model
the asynchronous shipment and creation of containers within an interconnected network of services, find the best path
routing for each container and minimise the use of transportations means. To carry out the demonstration and assess the
associated stakes, we use a set of actual flows from the fast-moving consumer goods sector in France. Various transpor-
tation protocols and scenarios are tested, revealing encouraging results for efficiency indicators such as CO2 emissions,
cost, lead time, delivery travel time, and so forth. As this is a first work in the field of flows transportation, the simula-
tion model and experiment exposes many further research avenues.
Keywords: transportation protocols; open logistic networks; physical Internet; sustainable development; logistics protocols;
shortest path

1. Introduction
Today’s logistics performance is limited in pursuing two antagonist goals within the current organisation of supply
chains. The first goal seeks small high-frequency shipments in a just-in-time philosophy while the second goal seeks
better environmental performance by best use of transportation means, notably heavier yet cleaner means. Mostly
dedicated logistics services and independent supply networks induce the fragmentation of logistics flows and thus the
difficulty to overcome this antagonism. Increasing the collaboration between supply chains or networks is a way to
exploit synergies between them and to then jointly improve their logistics performance, notably of the transportation
activities. However, even if pooling and horizontal collaboration are known concepts, only a few implementations are
operational and there is no sign of widespread generalisation. Hence the interconnection of independent logistic net-
works is becoming a hot issue in logistics towards enabling a performance that is equivalent or better than achieved
through pooling, while dealing with independent logistics organisations. Relative to this issue, in this paper, we design
and evaluate transportation protocols for use in a network of open and interconnected networks, namely the Physical
Internet (PI).
The aim of this paper is to assess efficiency of an open and interconnected PI. However, this efficiency is linked to
the implementation and performance of protocols, the functioning rules guiding operation in such a PI. Thus to reach
this goal, we define a set of transportation protocols for freight transportation in a PI network, then implement them in a
simulation model. The proposed algorithm-based protocols respectively focus on loading modular containers from
orders, finding the best paths for such containers from their origin to their destination within the PI network, and
consolidating containers efficiently to fill up transportation means at each hub towards destination.
The paper is organised as follows. In Section 2, we justify the need for a new logistics framework to cope with
worldwide challenges and then explain the PI approach, comparing it to other approaches. In Section 3, we introduce
the proposed transportation protocols. In Section 4, we present the simulation model in which the current and PI
logistics organisations and protocols are implemented using real-world data. In Section 5, we analyse the results and

*Corresponding author. Email: eric.ballot@mines-paristech.fr

© 2013 Taylor & Francis


3186 R. Sarraj et al.

limitations of this first large-scale experimental assessment of the PI protocols for freight transportation. In Section 6,
we summarise the contribution and provide perspectives for future research.

2. Why interconnecting logistic networks?


Improving efficiency in logistic networks is an old and never-ending quest. Yet the antagonism between the environmen-
tal challenges, particularly CO2 emissions, on the one side and the just-in-time delivery of small orders on the other side
may change the landscape not only because of new objectives like greenhouse gas emission reduction but also because
it may request significant changes in supply chain management and design. We hereafter depict the challenge and
explore main existing solutions before introducing the PI concept and how it could contribute to reaching efficiency
targets, focusing on performance and on CO2 emissions.

2.1 Efficiency issues related to actual logistic networks


Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

Logistic networks and the logistics services based upon those networks are organisationally and spatially fragmented,
being mostly dedicated to specific companies, overlapping each other with minimal synergy, and stretched by the
pressures of high-delivery frequencies induced by inventory reduction policies. For instance, a car manufacturer designs
and controls its own supply network, even if it is operated by a third-party logistics provider (3PL), with multiple deliv-
eries per week from the same supplier. A consumer goods supplier does the same with its distribution network, without
sharing its logistics resources with other suppliers even if it is delivering to the same distribution centres as these other
suppliers.
This fragmentation of logistics, especially in the fast-moving consumer goods sector (FMCG), is highlighted by the
crisscrossing of product flows through disconnected networks of plants, warehouses, distribution centres and stores. It is
wasteful due to the lack of resource sharing (Sarraj et al. forthcoming). As a result, freight transport intensity, measured
in tonnes-kilometres (t.km), reaches unprecedented levels. To emphasise this phenomenon, Figure 1 shows the growth
of flows and CO2 emissions in France, revealing a similar situation to that of many developed countries (Serveau 2011).
It illustrates that, despite all efforts, logistics flows are still growing at a faster pace than the overall progress made
toward reducing greenhouse gas emission. This is clearly unsustainable (Montreuil 2011).
The main assumption made here is the following: the logistics organisations that are at the source of freight flows
can also make major contributions to tackling logistics-induced environmental issues. This hypothesis is supported by
several freight transportation studies such as Kamakaté and Schipper (2009) and Piecyk and Mckinnon (2010). Statistics
from (EuroStat 2007) combined with the results of the survey by Mckinnon (2010) reveal real potential. Indeed truck

Figure 1. Growth of flows and growth of CO2 emissions in France compared to EC targets.
International Journal of Production Research 3187

weight capacity utilisation is less than 60% and around 25% of trucks are travelling empty, with very little improvement
over the years.

2.2 New freight efficiency approaches: horizontal collaboration and pooling


A first approach to address the global logistics environmental and efficiency challenges is to make better use of
available transportation means. A second approach encourages switching from road-based transportation to much less
greenhouse gas emitting transportation means such as trains, ships and barges. However, statistics in past decades in
many countries show little or no progress, despite network optimisation and operations optimisation done by operators
and despite real potential for emission reduction. For example, an empty truck trip releases 70% of the CO2 compared
to a full trip and an electric train releases around 20 times less greenhouse gas than a truck in France, thanks to nuclear
and hydro power stations (Joumard 1999).
Consolidation of freight is a solution avenue known since decades with currently asymptotic results as statistics
show due to adverse factors mentioned before. Therefore, this huge body of literature about consolidation is not
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

discussed here. We rather focus on new approaches.


Horizontal collaboration is a recent approach designed to tackle this problem and that has created significant
enthusiasm in recent years (Chan, Chung, and Wadhwa 2004; Cruijssen, Dullaert, and Fleuren 2007a; Naesens, Gelders,
and Pintelon 2009). An example of this approach is the concept of pooling which aims to merge flows from several
companies to take better advantage of transportation means so as to reduce shipment sizes and stocks. It also has been
demonstrated that it can enable greenhouse gas emission reductions from 14 to 50% (Pan, Ballot, and Fontane 2013).
There are only a few existing implementations of pooling, and they are involving of a very small number of partnering
organisations. Yet they already provide insights on the huge efforts required to make them happen and to keep them
operating. Key issues limiting the scalability of this approach are coalition formation complexity, value-added sharing
and lack of flexibility (Cruijssen et al. 2007b).
In the difficult context of small shipments and lack of multimodal solutions, the shift from truck to heavier transpor-
tation modes is decreasing year after year for manufactured goods, limiting access to the second approach.
The PI offers a new way to look at the consolidation problem, indeed at logistics at large. This approach aims to univer-
sally interconnect logistic networks as the Internet did with computer networks (Montreuil 2009; Sarraj et al. forthcoming).
It can be seen as a generalisation of pooling where coordination complexity is addressed by standardisation and protocols.
The concept involves generalising the use of modular containers in inland logistics to give to shippers a private space in
open networks and routing them like ‘packets’ as done with information in the Digital Internet, yet in a manner adapted to
the needs of logistics that has to harness the distinct characteristics of information and goods flow.
The former works in the literature have already provided a broad introduction to the PI concept and basic properties
(Montreuil 2011; Montreuil, Meller, and Ballot 2013), as well an in-depth analysis of analogy between the PI and the
Digital Internet (Sarraj et al. forthcoming). These works claim that the principles that allowed Digital Internet fast
growth during the past decades can also have a useful application in logistic networks despite the differences between
the information and material worlds. Key differences are taken into account in this paper, especially the physical
constraints relative to the weight and volume of goods, and those relative to transportation means capacity and speed.
The PI does not deal with goods of various shapes, sizes and materials, but rather with smart modular containers of
different sizes. It uses a set of smart and secured containers, namely PI-containers, with modular dimensions and
standardised interfaces for handling and communication. This allows using all kinds of logistics services as long as they
accept the PI-containers (Montreuil 2011).
Instead of dedicated, fragmented and overlapping supply networks, a global meshed network can be developed by
interconnecting different existing networks, as shown in Figure 2. In such a network each supplier, represented by a
square, ships directly its goods in PI-containers to the end destinations, represented by a circle. On the left side of
Figure 2, depicting classical supply networks, the products are sent on pallets directly from a manufacturer’s plant to
one of its warehouses, then to one of distribution centres of clients. On the right side of Figure 2, depicting a PI
network, products are loaded at the plant in a set of PI-containers and each of these PI-container is shipped through
hubs of the network in relay mode toward its destination. The concentration of flows provided by the global network of
logistics services is to also help efficiency thanks to the economy of scale induced by better use of transportation means
and a shift of flows to more efficient transportation means (Ballot et al. 2011a). The impact on detours to reach some
hubs, assumed as a second-order factor, is evaluated in this paper to assess the associated stakes.
One major assumption of the PI is that the transit through several hubs may require more handling (transshipment
and sorting of PI-containers in the hub) as compared to direct shipment, but that this drawback is overcome by handling
composite PI-containers of larger sizes than pallets and other benefits such as stock reduction and better transport
3188 R. Sarraj et al.
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

Figure 2. Overlapping networks vs. interconnected network.

vehicle fill rate (Montreuil, Meller, and Ballot 2010). In this paper, as we deal only with the upstream of supply net-
works, we have shipments sizes ranging from the equivalent of a pallet and to the equivalent of a few trucks. Accord-
ingly, we consider only containers with section of 2.4 by 2.4 m that will not be opened in hubs. A key aim of this
paper is to provide a first demonstration through a simulation experiment that the foreseen efficiency benefits of the
change of logistics schemes can be confirmed when operational constraints and dynamics are taken into account. To
investigate the potentiality, a substitution of the actual logistics organisation by the PI is implemented in a simulation
model. This operationalisation of the PI in the simulation model uncovers a second key aim of this paper that is provid-
ing a primer on defining PI transportation protocols, as tackled in the next section.

3. Transportation protocols in the PI


The process of shipping goods from a supplier to a customer within the PI is significantly different from current
practices. The process implies several steps summarised in Figure 3. First of all, each order is loaded in a ‘best’ fitting

Figure 3. Transportation process from a shipment point of view.


International Journal of Production Research 3189

PI-container or set of PI-containers. When a PI-container is ready to be shipped, the ‘best’ path toward destination is
identified inside the network of all available logistics services. The path made of several segments or logistics services
could start by a truck service to a first hub, then a connection with a train service and so forth until reaching destination
where the goods are offloaded from the PI-container. Between transportation segments, PI-containers are handled in
hubs. For each PI-container getting into a hub, there is a search for best fitting transportation means for the next seg-
ment, in terms of capability, availability and alignment with the economical, environmental and time-related service
objectives of each PI-container. If a transportation service is found, the consolidation of PI-containers is done. Other-
wise, the departure is forced or an alternate path is computed. Each operation is performed by rules and optimisations.
The protocols structuring these decisions and operations are hereafter named transportation protocols.

3.1 Containerisation of goods


The goods containerisation protocol specifies how products ordered for shipment are assigned to be grouped into a
single PI-container or a set of PI-containers, how to decide on the size(s) of PI-container(s) to be used for each shipping
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

group, and how to decide on the loading sequence and pattern of goods within each PI-container. It is applied at each
source of freight to be shipped (e.g. a plant or a warehouse). Goods containerisation is the equivalent of encapsulation
within Internet. The consolidation is clustering orders to be shipped within the same period (e.g. day) and heading for
the same destination, either the same final customer or some common warehouse. Thus, in our case there, is no
unloading and loading of PI-container contents in intermediate hubs.
The definition of the overall set of modular containers for the PI, with accurate sizes, interfaces and functionalities,
is a challenging task beyond the scope of this paper. Even though they do not have the entire set of properties specified
for PI-containers, the current ‘pallet-wide’ (PW) container is an intermodal transport unit used in Europe that provides a
starting basis illustrating the spirit of what is envisioned (Bouley 2010). The PW containers, unlike ISO sea shipping
containers, are adapted to Euro-pallets, thus allowing better cargo optimisation and using more fully the capacity of
container ship.
In this paper, in order to start with a module close to actual logistics practices, we take the Pallet Wide 40-foot
(12-metre) container as a basic reference for dimensioning PI-containers. Indeed, we use the dimension of a Pallet Wide
40’ (12 m) as the maximal length and divide it by 10 for the smallest-length container. We will assume that their
internal volume can be fully used. The used sizes, reported in Table 1, are not to be perceived as prescriptive. They are
rather illustrative of potential sizes in line with actual logistics modules while exploring the advantage of modular
dimensions, in contrast with the rigid 20-foot and 40-foot choices of actual sea shipping containers.
The common section 2.4 m by 2.4 m makes them compatible with pallet wide containers and transportation means
such as ships, trains and trailer trucks, while the different lengths allow flexibility in shipment sizes. No smaller sections
(pallet size or less) are tested here as the destinations are distribution centres and not retail stores in the simulated
experiment.
All reported simulation experiments exploit a specific combination of these modular sizes and measure resulting
performance levels in terms of fill rate, transportation means use and handling intensity. Beyond the current
experimentation, such a simulation model as used here could be part of the evaluation process for sets of PI containers
proposed for standardisation. As the remainder of the paper deals strictly with PI-containers, any further mention of the
term container is to be referring to a PI-container, except if explicitly specified.
To specify the set of containers to encapsulate a set of goods in a PI context, there are several possible strategies.
The first is to use the largest containers possible, minimising the number of containers used. This aims toward reducing
management costs and minimising handling and containerisation costs, as all contents of a container always reach the
same end destination. A second strategy is to use only a small-length container to reduce lead times and stock levels
while exploiting the PI to keep transport costs low and add flexibility. The third strategy is to take advantage of modular

Table 1. Physical Internet container sets for the simulation experiments.

Set Sizes (inner dimensions)

1 2.4 m × 2.4 m × 12 m
2 2.4 m × 2.4 m × 1.2 m
3 2.4 m × 2.4 m × {1.2, 2.4, 3.6, 4.8, 6, 12}
3190 R. Sarraj et al.

dimensions by offering a set of dimensions to fulfil various needs. We explore these strategies with three sets of
containers.
Given a set of allowed container sizes, the selection of specific containers for every order (origin, destination, date)
here aims first to minimise the number of containers and second to maximise space utilisation. In FMCG, weight is very
rarely the constraining factor: volume is the dominant constraining factor. This set of criteria was selected to minimise
the handling efforts and to limit the associated container usage costs. For example, if containers of 12.6 and 3 m are
allowed and available, and if the order occupies the equivalent of 45% of a container of 12 m of length, then the
solution will be to use one container (6 m of length) instead of one 12 m-long or two 3 m-long. Then normally the next
step is to set the load plan of the container. This step is not included in the simulation experiments, assuming loading
feasibility when there is volume and weight feasibility.

3.2 Container routing


Routing is a well-known approach used in several areas such as computer networks, logistics and transportation. In the
latters, routing applications have focused on vehicles or fleets of vehicles to find the shortest path from a source to a
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

destination or to generate an optimal tour of a set of pickup and delivery locations. Here we use routing techniques to
move PI-containers from origins to destinations.
In the Digital Internet, a routing protocol specifies how the nodes communicate the network state and use this infor-
mation to select a route between two nodes. The route selection is done by a routing algorithm according to a ‘cost’ or
a ‘distance’ function (Comer 2006). We use a similar logic to route PI-containers in the PI.
Basically, choices made in the protocol implementation, algorithm design and valuation function are related to the
nature of the network. Hereafter, we discuss briefly the differences between digital datagram routing and physical
container routing, then introduce the container routing protocols.

3.2.1 Differences between data and freight routing


Even if there are a lot of similarities between the routing of containers in the PI and the routing of datagrams in the
Digital Internet, there are also some important differences due to the difference between information and material. For
more on this subject see (Sarraj et al. forthcoming). We focus here only on routing related differences.

3.2.1.1 Computer networks and routing protocol design. Whereas the Digital Internet network infrastructure can be
considered stable during the route selection process, the flows are quite unpredictable and the state of the network can
change faster than the update process. This is why (1) the main routing objective is to minimise load and avoid
congestion points regardless of geographic distance, and (2) planning entire routes from origin to destination is quite
difficult. This has lead the main Digital Internet routing algorithms to have the following properties:
 They are dynamic;
 They maintain a routing table at each node;
 They use a hop-minimising shortest path approach based on distance vectors, such as the Routing Information
Protocol (RIP) and the Border Gateway Protocol (BGP) or state links, such as Open Shortest Path First
(OSPF);
 They focus on a specific metric such as the number of hops or the available bandwidth on arcs to reach
destination;
 They delete data packets in case of delay, which involves a new release of these packets.
For more information about Digital Internet routing protocols (TCP/IP, RIP, BGP and OSPF) and algorithms, see
(Comer 2006; Farooq and Di Caro 2008; Resende and Pardalos 2006).

3.2.1.2 Logistic networks and routing protocol design. Similarly to a data service network in the Digital Internet, a
logistics service network also has irregular traffic patterns. Yet unlike Digital Internet wires with a given capacity, there
is flexibility in the transportation industry (especially trucking) to dynamically adjust the capacity to the demand on each
arc. Traffic is also potentially easier to anticipate because freight is moving more slowly than its corresponding informa-
tion, although this is not always exploited in current practice. In contrast with the Digital Internet, a routing algorithm
in a logistic network seeks, in addition to satisfaction of a certain service level, the saturation of services (e.g. transpor-
tation means), especially scheduled ones. And we rely on postal addressees for site location and assume they are ready
to receive containers as we are dealing with well-known warehouses and distribution centres.
International Journal of Production Research 3191

Therefore, the PI routing algorithms and protocols are expected to have the following properties:
 They are dynamic to take into account demand volume fluctuation and to select best available transportation
service;
 They maintain routing tables at each node to deal with service updates and incoming flows;
 They use a state-link routing approach that needs more storage but establishes a set of best paths with
alternates, allowing a choice between them;
 They focus on a specific metric according to logistics needs, such as cost, time or greenhouse gas emission,
while looking for means saturation;
 They change departure priorities or paths in case of container delay.

3.2.2 Routing objectives within PI


Routing is a key factor for network performance in both the Digital Internet and the PI. Yet the objective functions
differ significantly from one context to the other. This is mostly due to the fact that while the Digital Internet incurs
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

negligible marginal data transportation cost, every freight move in the PI generates significant variable transportation
and handling costs. In the Digital Internet, a dominant objective is to minimise the number of hops by nodes subject to
waiting time constraints. In the PI, optimising the use of means will always be done in compromise with three other
key minimisation objectives as pertinent to the shipper:
 Costs, including transportation, handling, sorting and transiting costs;
 Time, from the instant a container requests a departure until its arrival to destination, including transportation
time and time spent in routing centres (hubs), including waiting time.
 Environmental impact, dominantly measured by transportation and transshipment induced greenhouse gas
emissions and energy consumption.
The routing objectives in PI logistics are complex and differ for each container according to priorities set by the
shipper.
In this paper, we only focus on the question of routing (finding the best path from origin to destination for every
container) and to simplify our modelling system, the question of container fleet management is not considered. Accord-
ing to our ongoing research works, this is a complex and important subject (ownership of container, business models,
equilibrium of flows, reveres logistics, etc.). However, it is too early to be discussed in this paper.
Container traceability is another subject to be studied. Differing from the immateriality of the data in Digital world,
containers in the PI should have exact addressee information for each delivery (transshipment or final destination).
Hence the traceability is a must-have property of PI containers. This information could be achieved by considering a
container as an object of the Internet of Things and emerging ECPGlobal standard to publish the information to the
Digital Internet.

3.2.3 Container routing algorithm


The routing algorithm proposed here assumes asynchronous shipment of each PI-container and the availability of real-
time information about the network state. The algorithm runs in a decentralised manner. In a prospective manner, we
could imagine that first the logistics service provider will use some protocols to check route, price and availability. Then
each hub, managed by a 3PL for example, has a view of the network real state thanks to the open monitoring property
of the PI, exploiting upcoming flows of information through the digital capabilities of the Internet of Things (Montreuil
2011). Then, every hub can apply the transportation protocols for each container to search the best match between the
required service and actual state of the hub and the container’s next steps in the network.
There are numerous shortest (or best) path algorithms in graph theory. In our application, we are not looking for a
dynamic programming approach like that of Bellman-Ford-Moore because of possible convergence issues when
unexpected changes in the network occur. Instead, we propose that each node compute the best path to every destina-
tion. For that purpose, two types of algorithms are possible. The first type embeds the exact algorithms to provide the
optimal solution, such as the algorithms of Kruskal, Prim, Bellman-Ford-Moore or Dijkstra (Bellman 1958; Dijkstra
1959; O’Cearbhaill and O’Mahony 2005). The second type embeds the heuristics that can speed the optimisation pro-
cess to find good solutions (Fu, Sun, and Rilett 2006).
The shortest path problem is well known and finding the optimal solution needs a polynomial computation time that
becomes long in a large-scale graph. This is exactly our case with a France-wide network of over 13,000 arcs (roads
3192 R. Sarraj et al.

and rails) and 500 nodes, and up to over a million containers to route concurrently. For this reason, we use a heuristic
method to rapidly find solutions. After a review of the heuristics related to the problem, we found A*, the most popular
among all heuristic algorithms for this problem (Fu, Sun, and Rilett 2006), to be well suitable to be applied in our
model because it also allows us to have optimal paths.
The A* algorithm is an extension of Dijkstra’s exact algorithm. It uses an evaluation function on each node for
estimating the best path between this node and the destination, thus allowing the orientation of Dijkstra’s algorithm.
This function, which estimates also the remaining cost (or length, time), must be well defined to achieve good solutions.
In our case, we made sure to assign minimum values for this function, ensuring optimum solutions. For instance, for
the time criterion, we implemented Euclidian distance to destination for the faster transportation mean. Readers can refer
to (Dechter and Pearl 1985; Fu, Sun, and Rilett 2006) for more details.
With this algorithm, we are able to find at each node the best path to destination satisfying each container-specific
requirement. In the simulation experiments, without loss of generality, we deal with cost, delivery time and CO2
emissions. In the simulations, we currently do not use real-time state of the network to seek for the best services
(transshipment load factor in hubs, road congestion, etc.).
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

3.3 Container consolidation on transportation means


The container consolidation protocol aims to fully load the selected transportation services.

3.3.1 Container consolidation protocol


At each node of the PI, the PI-containers asynchronously arrive with information about their next arc (next destination).
The protocol optimises the consolidation of containers by next destination in transportation means while looking to have
urgent containers shipped first, as shown in Figure 4. The optimisation of loading corresponds to the well-known
bin-packing problem that will be discussed in the next subsection.
The container consolidation protocol aims to respect the target delivery times, assigning a higher value to urgent
containers so that they shall be immediately transported. However, if the load of one of these means (provided by
solving the bin-packing problem) is not deemed sufficient, a new path is identified for the PI-container (with a truck
instead of a train for example). Yet if there is no alternative solution, the departure will be done, regardless of the load
status. The consolidation and routing protocols are interdependent. Furthermore, we look not only for consolidation to
the next common hub but to the subsequent common destination whenever possible in order to avoid as much as
possible PI-container unloading, sorting and loading operations.
In summary, the protocol cheques container consolidation per common destination (potentially not only the next
one) and eventually picks a near optimal arc (transportation service) according to flow level, if none is available the
departure is forced and in all cases loading is optimised.

Figure 4. Routing principle in a multimodal hub.


International Journal of Production Research 3193

3.3.2 The loading problem


After the containers are grouped by next destination, they are associated to transportation means. This creates a need for
optimising the fill rate of each means, in order to minimise the number of used means. This can be modelled as a bin-
packing problem: given a set of objects (here PI-containers), each with their own priority, size and weight, as well as a
set of bins (transportation means) with their individual capacities, find the minimum number of bins so that each
container is assigned to a transportation mean. The bin-packing problem has been addressed since the 50s by numerous
researchers in operations research and combinatorial optimisation, such as the classical (Gilmore and Gomory 1961,
1963).
In our empirical simulation based research context where (1) the modular sizes (allowed in Table 1) for the PI-con-
tainers have all the same height and width, differing only in volume by their length, and (2) the means have identical
capacity for each mode (railcars, trucks), the loading problem minimising the number of loaded means can be
formulated mathematically as the following two-dimensional bin-packing problem:
X
k
min yi (1)
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

i¼1

yj utilisation of transportation means (bin) j (1 if bin j is used, esle 0)


xji loading container i in means of transportation j (1 if container i is assigned to bin j, esle 0)
ci container i volume, i 2 f1; . . .; kg
wi container i weight, i 2 f1; . . .; kg
C means of transportation volume capacity
W means of transportation weight capacity
X
k
xji : ci  C : yj ; 8j 2 f1; . . .; kg (2)
i¼1

X
k
xji : wi  W : yj ; 8j 2 f1; . . .; kg (3)
i¼1

X
k
xji ¼ 1; 8i 2 f1; . . .; kg (4)
j¼1

xji 2 f0; 1g; 8i 2 f1; . . .; kg; 8j 2 f1; . . .; kg

yi 2 f0; 1g; 8j 2 f1; . . .; kg


The objective function is to minimise the number of means used (Equation (1)). Three constraints exist, respectively to
ensure to not exceed the mean’s volume (length) and weight capacity (Equation (2), Equation (3)), and that each
container is assigned to one and only one means (Equation (4)).
This combinatorial problem has been long known to be NP-hard. The difficulty of finding the optimal solu-
tion grows exponentially with the number of PI-containers. A Branch-and-Bound algorithm based on a linear
programming relaxation could provide optimal solution. Yet the performance of such a methodology limits its
use to small-scale problems when fast solution time is critical. For large-scale problems, heuristic approaches
are used with the aim to generate near optimal solutions in a short time. Among the most used heuristics
(Johnson et al. 1974; Kellerer, Pferschy, and Pisinger 2004), the following ones are particularly appropriate in
our context:
 First Fit (FF): Initially a single bin is considered, and objects are loaded into the bin according to their order of
arrival. When there is not enough space in the first bin for a given object, it is loaded into the second bin
without closing the first one. Then, the following object is tried to get into the first bin before the second,
unless the first one is full and thus closed, and so on.
 First Fit Decreasing (FFD): Knowing that the sorting criterion according crucial for the quality of the solution
(Johnson et al. 1974), the FFD method consists in sorting out objects in decreasing order of their size and then
loading them into the bins by order, using the First Fit methodology.
3194 R. Sarraj et al.

 Best Fit (BF): As in the FF heuristic, the objects are sorted by arrival. The difference is that in BF an object is
tried first into the fullest bin amongst all the open bins, and if there is not enough place it is tried in the second
fullest, and so on.
 Best Fit Decreasing (BFD): The BFD heuristic differs from the BF heuristic by sorting first the objects in
decreasing size order.
In general, the FFD and BFD yield better solutions than the others (Johnson et al. 1974; Simchi-Levi 1994). We
have experimented with both of them in our model and we found that their results are very close whereas the FFD
needs less computation time since it does not require sorting bins according to their fill rate. Based on those reasons,
we chose to implement the FFD algorithm for our simulation modelling needs. Yet in other contexts, more sophisticated
heuristics could be used (Coffman, Garey, and Johnson 1996).
This completes our primer contribution toward a set of PI protocols, focused on enabling our experimental simula-
tion-based investigation of decentralised PI-container transportation.
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

4. Simulation model and experiment


In this section, we describe how we implemented these transportation protocols in a simulation model using real data
from fast moving customer goods (FMCG) supply chains in France. We also define scenarios evaluating different
aspects of interconnected PI networks. Finally, we present an analysis of the results.

4.1 Simulation model


In this research, a simulation model has been developed as the instrument used for experimentally evaluating the impact
of, generally, the interconnected PI networks as compared to the current logistic networks and of, specifically, the
designed transportation protocols.
The simulator is built exploiting the AnyLogic multi-agent-based system (Van der Zee 2006; Lee and Kim 2008;
Ruiz et al. 2011) using a discrete-event simulation modelling approach. The complexity of the model requires using the
native Java environment of the AnyLogic platform to implement the needed modelling extensions in Java code and to
integrate external data sources. At this stage, the simulation uses only real data as demand input without additional ran-
dom perturbation. The agent approach was chosen to implement a decentralised framework and to enable negotiations
between agents, to be investigated in further work.

Figure 5. The simulator architecture.


International Journal of Production Research 3195

Basically, the simulation model uploads the logistic networks data (current and PI), the maps, the order data, and
generates agent instances in each node. Figure 5 illustrates the global architecture of the simulation. We have in parallel:
one part dedicated to the current supply network scenario and another part to the PI scenarios.
To establish the reference scenario and validate the simulation model, we use input data corresponding to real orders
from the FMCG sector in France and the current logistic networks (the real location of logistics sites as well as roads
and railroads). We implemented in the model the current practices in terms of transportation (creating pallets, loading,
transshipment, unloading, etc.). The simulator generates output data stored in a database from which are compiled the
Key Performance Indicators (KPIs).
In parallel, we use the same orders to simulate the operations of the PI protocols and obtain the data output
then used for compiling the PI KPIs. This methodology allows us to compare the performance of the current
logistic networks and the PI networks, to validate our model by comparison of the simulation results of current
logistic practices using real-life data, and to analyse the potential of PI with different hypotheses and protocols
settings.
There are three main agent classes corresponding to the transportation protocols presented in Section 3 (Con-
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

tainerisation, Routing, Consolidation). Figure 6 represents the agent classes implemented in every sourcing point
(a plant or a warehouse) of the simulation model, their operating process and the communications between them.
Figure 7 explains the operation at the level of a hub. When a container arrives at destination (a warehouse or
a distribution centre) it is simply offloaded and information is recorded. Furthermore, Figure 8 provides a global
view of the Consolidation Agent process in a simple case not considering re-routing and intermodal use of
train.
The main levers for the simulation model are route selection, transportation mean selection and loading as well as
consolidation at each hub.
The simulation model is also based on a large set of parameters. They were set according current practices. Table 2
indicates the main ones.

Figure 6. The simulator agents (for one node) in a sourcing point.


3196 R. Sarraj et al.
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

Figure 7. The simulator agents in a hub.

Figure 8. The Consolidation Agent process scenarios (A simple case: not considering re-routing and intermodal train usage).
International Journal of Production Research 3197

Table 2. Default settings of the model.

No Coefficient Value

1 Truck Capacity in useful volume 76,032 m3


2 Truck Capacity in weight 25 tonnes
3 Train Capacity in useful volume (6 railcars) 456,192 m3
4 Unit cost of road transport (full truck) 1€/km + 0.95€/l of diesela
5 Unit cost of rail transport (full train) 40€/km8
6 Cost of entry/exit-stock 10€/container3 1€/ palette
7 Handling unit cost per loading, transshipment For a container: 5€/ movement + 3.3 × Length €b. For a pallet: 2€
or unloading movement
8 Cost of container rental / dayc 1.8 × 0.35 × Length €/day
9 Energy consumed by electric train 0022 kWh/t-kmd
10 Emission factor by train 0,0035 × x kg CO2/kme. x: train weight (tonnes)
11 Emission factor per truck (0772 + 0013 × x) kg CO2/kmf. x: truck weight (tonnes)
12 Minimum sojourn time in a hub (Loading/Unloading time, 60 min
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

sorting operations, movements, etc.) for a container


a
Pan (2010). Contribution à la définition et à l’évaluation de la mutualisation de chaînes logistiques pour réduire les émissions de co2
du transport: Application au cas de la grande distribution. Ph.D. Mines ParisTech.
b
Calculated from the values of see shipping containers (20’ and 40’). These are approximate values but increased to ensure the poten-
tial of the PI, and to give credibility to results.
c
http://conteneurs.fr/Resotainer-Louer-8–2.html. With this formula we obtain a Handling cost of 44€60 for a 40-foot long shipping
container.
d
Report Carbon accounting, Ademe (2007). Guide des facteurs d’émissions. Bilan carbone. Paris: Agence de l’Environnement et de
la Maîtrise de l’Energie, Mission Interministérielle de l’Effet de Serre.
e
= 0022 kWh/t-km × 0.045 kg CO2/kWh (French rate) = 0001 kg CO2 / t-km, but for a train we have to include the net weight of
locomotive and waggons, then 0.0035.
f
Pan (2010). Contribution à la définition et à l’évaluation de la mutualisation de chaînes logistiques pour réduire les émissions de co2
du transport: Application au cas de la grande distribution. Ph.D. Mines ParisTech. Example of full truck with 25 tonnes:
0772 + 0013 × 25 = 1097 kg CO2/km.

4.2 Input data


The data-set used for the demonstration came from two major retail chains (Carrefour and Casino) in France and their
106 biggest common suppliers. The data encompass the flows from plants in France and in Europe (taken into account
from France border) to warehouses and/or to distribution centres according to the design of actual logistic networks.
Déméter Club, an association dedicated to sustainable logistics for Fast Moving Consumer Goods, provided real
logistics order data over 12 weeks.

4.2.1 Context data


The geographic position of the facilities is specified through real longitude and latitude coordinates. These facilities are
linked to hubs by roads, highways and railroads registered in the ROUTE 120® IGN geographic information system
that includes the top 120,000 km of roads in France. All logistics services (arcs) are selected according to lowest energy
consumption criterion (very close to shortest time) between all nodes with the infrastructure database.
The driving and rest times for drivers are taken from French regulations. Economic data such as pallet handling
costs in a warehouse, trucking prices, etc. have been provided by industry.

4.2.2 Supply networks and flows


As mentioned before, the flows used for our demonstration are actual flows from two retailers in France. Two steps have
constructed the database: first we identified the biggest common suppliers of the two retail chains and second we
extracted these suppliers’ history flows within the first 12 weeks of 2006. Thus, the database concerning a supply
network of 106 suppliers and 2 retailers was constructed.
The supply networks include 303 plants, 57 warehouses (operated by a manufacturer or sub-contracted to a 3PL)
and 58 distribution centres (operated by a retailer or sub-contracted to a 3PL) all across France. They supply three
product families: liquids, grocery as well as personal and home care. Figure 9 represents the locations of facilities and
3198 R. Sarraj et al.

the 4451 flows of 2,582,692 (pallets) operated independently. This corresponds to around 20% of the French FMCG
market share for the considered products families. We explicitly model three types of flows: from plant to distribution
centre (DC), from plant to DC, and from warehouse (WH) to DC. Table 3 gives more details on flow data. This data
stem from a mix of full-truckload and less-than-truckload shipments.
The database containing real FMCG flows is used as input to the simulation. This origin ensures the resemblance of
the simulation’s results with shipments that take place in current practice. This usage of current flows is a purposeful
and voluntary limitation: the study of contexts where the existence of the PI induces changes in shipment sizes and
demand patterns is left for further research. In this regard, the results reported in this paper are in many ways conserva-
tive in regard to the PI potential. They are deemed necessary for both the scientific and professional communities to
assess the potentiality, from a perspective of transitioning from the current state in a phased approach towards a mature
PI. This paper addresses the transition phase.
The networks displayed in Figure 9, overlapping each other into a dense spaghetti tangle, reflect the existing
structure and are implemented in the reference scenario.
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

Figure 9. Actual supply networks – flow view.


International Journal of Production Research 3199

Table 3. Current flow data description.

Categories Liquids Grocery Care

Number of product 130 355 217


Number of supplier 29 55 24
Flow characteristic by link Plant to DC Plant to WH WH to DC

Links 2738 251 1461


Mean dist. (km) 511 444 441
(St. Dev. of dist.) (276) (302) (262)
Orders (in number) 56,710 22,730 37,604
Mean order size (Pallet) 19,77 32,30 19,20
(St. Dev. of order size) (10, 11) (2, 69) (10, 28)
tonne 991,323 443,540 435,149
t.km 474,459,658 165,930,963 163,419,052
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

4.2.3 PI networks
In practice, the PI network construction would be a long and evolutionary multi-actor process. In order to simulate
transportation through a realistic PI-network, we created one through optimisation using the input data. This optimisa-
tion problem was solved with a heuristic approach that mimics an evolutionary process. The PI-network is optimised
according to a cost function and with operation constraints such as maximum length of truck trips to avoid round trip in
more than one driver shift. For details on the optimisation process, hub location and arcs determination, refer to Ballot
et al. (2011b).
The resulting network of logistics services between origins, hubs and delivery points is an input to the simulation
model. Figure 10 illustrates a PI network design.

4.3 Scenarios
To investigate the potential of the PI, several scenarios have been set. These scenarios represent progressive levels of its
deployment. Three families of PI scenarios are studied in addition to the reference scenario. Each family has

Figure 10. Interconnected networks, scenario 2 – traffic and flow views.


3200 R. Sarraj et al.

sub-scenarios, as shown in Table 5. For simplification purposes, the term scenario 1 refers in the text to scenario family
1, and so on. Table 4 summarises the four families of scenarios.

4.3.1 Reference scenario 0: simulating the current logistic network


Scenario 0 is the reference or status quo scenario approximating the historic events based on the provided data. It
simulates the current logistics practices applied to the flows from a plant (PL) to a warehouse (WH) of a supplier, and
to a distribution centre (DC) of a retailer, as well as from a PL directly to a DC. The flows are transported in a point-
to-point mode by trucks only, as shown in Figure 11. This scenario enables the assessment of the performance of the
transportation in the current logistic networks as a basis of comparison with the PI scenarios.

4.3.2 Scenario family 1: road-based PI transportation


The first family of PI scenarios keeps the structure of the current supply chain as in Scenario 0, transportation is still by
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

road yet it is carried out through the PI network, as illustrated in Figure 12. In this scenario the goods travel exclusively
in PI-containers.
The objective of this scenario family is to investigate the potential of interconnection without changing the current
supply chain structure and transportation means. This scenario was tested with the three sets of containers defined in
Table 1, with time and cost routing criteria.

4.3.3 Scenario family 2: multimodal PI transportation


The second family of scenarios studies the performance of the PI network presented in Scenario 1, now exploiting
multimodal transportation, as depicted in Figure 13. The aim of this scenario is to establish the potential of PI-enabled
multimodality exploiting heavy transportation means that are more efficient, especially in terms of CO2 emissions
reduction, yet largely inaccessible to the FMCG companies in the setting of the current logistics system. The PI network
permits the consolidation of the PI-containers of different shippers, generating enough volume and thus allowing the
shift to railway service. Furthermore, the containerisation of goods enhances the efficiency of multimodal transportation
as it facilitates intermodal transitions.
Like scenarios in family 1, the scenarios in family 2 contrast different routing criteria such as cost and time, as well
as greenhouse gas emissions.

4.3.4 Scenario family 3: PI network without manufacturers’ warehouses


The consolidation of small shipments of different actors in the PI attenuates the need for each actor to ship large loads
on its own. This allows getting away from them having to send full truckloads from each plant to a centralised
warehouse where the products are stored and prepared in pallets for further distribution. To investigate this potentiality,
we simulate the direct shipment from plants to the retailers’ distribution centres by pushing the distribution centres’
orders directly to plants, to be transported from origin to final destination in inter-hub relay mode, as depicted in
Figure 14. It is a generalisation of current practice for the bestsellers as they have enough volume to ship directly to
each distribution centre. Like Scenario 2, this scenario supports multimodal transportation.

5. Results
5.1 Model validation
For validation and reference purposes, we built and simulated a base scenario 0 representing the current situation.
Validation was carried out in two steps. The first one checked the consistency of the model. We compared the

Table 4. Scenarios presentation.

Scenario Family 0 1 2 3

Orders Actual Actual Actual Without WHa


Network Actual Physical Internet Physical Internet Physical Internet
Train No No Yes Yes
a
WH refers to warehouse.
International Journal of Production Research 3201

Table 5. KPIs of the simulation results.

Scenario Scenario Container Route Mean weight Mean hubs per Mean lead Logistic cost
Family Id. set criterion loading (%) route time (h) CO2 (t) (k€)

0 NA NA NA 59 0 5.86 52,742 81,976


1 1.1.C 1 Cost 62 2.24 5.4 53,107 77,679
1 1.2.C 2 Cost 76 2.20 10.02 45,545 104,268
1 1.3.C 3 Cost 76 2.25 8.89 45,599 68,429
1 1.3.T 3 Time 75 1.93 8.69 46,682 68,861
2 2.3.C 3 Cost 72 2.28 8.73 29,878 62,275
2 2.3.T 3 Time 70 1.66 8.45 33,035 64,591
2 2.3.E 3 CO2 72 2.94 9.88 23,600 65,720
3 3.3.E 3 CO2 67 2.95 11.84 22,120 57,724
3 3.3.C 3 Cost 65 2.30 10.69 27,584 56,057
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

Figure 11. Reference scenario: Current Supply Chain Model.

Figure 12. Scenario 1: Physical Internet with road transportation only.

Figure 13. Scenario 2: Physical Internet with multimodal road–rail transportation.

Figure 14. Scenario 3: Direct flows from plant to distribution centre within Physical Internet.

simulation output of Scenario 0 and the real field results from the database in terms of number of pallets in the system,
total weight of transported products, number of fulfilled orders, their size, etc., to make sure the model correctly
simulated the playback of the input field data. The second step was a comparison between the input flow data and the
existent studies of transportation in FMCG chains in the literature. Even though many indicators are measured in each
simulation run, the key performance indicator used for validation purposes was the truck-trailer weight fill rate as it is
the only one available in official statistics for the FMCG sector. The simulated fill rate of 59% is in line with the results
computed from the database provided by the Demeter Club, with judgement of consulted field experts, as well as with
openly available statistics: 53% in food supply chain survey in the UK (Mckinnon et al. 2003) and 56% in France on
3202 R. Sarraj et al.

average (EuroStat 2007). The difference, around 3–6% can be explained by our sample of products with a lot of
beverages, and our massive flows that corresponds to the 106 biggest suppliers. Therefore, we can conclude that the
model and our data are representative.

5.2. Simulation results and analysis


The simulation experiment was constructed based on the three scenario families previously described that shift toward a
more intense use of the PI for transportation and transshipment purposes. For each scenario family, distinct scenarios
have been investigated as depicted in Table 5. Within the specifications of each scenario family, scenarios vary in terms
of the container set allowed and the routing criterion. The results provided in Table 5 and Figures 15–20 are in terms of
key operational, economic and environmental performance indicators.
Many variations have been tested for the targeted scenarios, notably according to the minimum load allowed for
immediate departure; maximum allowed waiting time and the handling cost structure. These variations have enabled us
to fine tune the specific parameters to be used for each scenario. For instance, we found that a maximum waiting time
for a container of 12 h compared to 3 h does not really improve the fill rate but actually impairs the lead time. Illustra-
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

tively, the scenario results here presented are thus based on an allowance for a maximum three-hour waiting time at
each node and an immediate departure allowance whenever the truck/trailer becomes at least 90% full. The latter allows
the departure of a European trailer truck with 12 m long of container(s) in a 13.2 trailer.
For each indicator, the variations of the three scenarios are hereafter presented simultaneously and compared to the
status quo scenario.
The Table 5 does not describe all combinations done in the research but rather focuses on some experiments to high-
light the impact of key factors. The complete research report with all results and KPI is available (Ballot et al. 2012).

5.2.1 Container size usage


Figure 15 contrasts the use of PI-container sizes within scenarios and between scenarios. The distribution of container
size usage in the simulation experiments demonstrates that actual orders target the size of the trailer, in line with the

Figure 15. Container trips per size among scenarios (out of scale: 117,585 for size 12 m in Sc 1.1 and 677,551 for size 1.2 in Sc
1.2).
International Journal of Production Research 3203
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

Figure 16. Transportation means filling ratio for each scenario.

Figure 17. Average number of passages through a hub for each scenario.

Figure 18. Average delivery travel time of PI-containers in hours by scenario.

actual approach focusing on full-load dedicated transport, and contributing a priori toward minimising transportation
costs. Indeed, except for scenario 1.2 that imposes small 1.2 m long PI-containers, in all cases there is a domination of
the usage of 12-m PI-containers. However, as is later reinforced with other indicators, the limitation to the 12 m
container reduces transportation performances, like the fill rate. Note that scenario 1.2, with its small PI-containers, is a
3204 R. Sarraj et al.
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

Figure 19. Scenarios’ costs breakdown.

Figure 20. CO2 emissions per scenario (rail emissions are in blue).

variant encouraging the reduction of shipment sizes, thus inducing increase in transport filling rates, shipment frequency
and reduction in stock requirements.
It is a limitation of our study: as we did not alter the current orders, we penalise the PI alternatives by reducing their
degrees of freedom and forcing them to play with a hand optimised for the current constraints. For example, suppliers
try to saturate means, particularly in the upstream between factories and warehouses where full-truck ordering/shipping
is the rule, thus impacting supply lead times and frequencies. It is therefore not surprising that scenario 3 results in the
most balanced set of container sizes, since the orders from the distribution centres to the plants are not packaged a priori
to fit full trucks.

5.2.2 Transportation means fill rate


The fill rate results provided in 5 and depicted in Figure 16 show that a shift towards PI implementation in the studied
setting enables a highly significant gain from a current 59% utilisation of the load capacity of transportation means up
International Journal of Production Research 3205

to between 65 and 76% in weight, and exhibit a similar tendency in volume. The results in volume are less impressive
due to the relative heavy density of products used here, mostly liquids (drinks and home care).
Scenario 1 produces an increase of the weight fill rate by 17% comparing to the current scenario. This is due to the
consolidation of flows in the PI Network. The best gains are with scenario 1.2 using only the small 1,2 m-long PI-con-
tainers, providing the maximum degree of freedom in loading vehicles.
The gain is lower in scenario 2 because we divide the flows in two parts: a first part using trucks through road trans-
portation and a second part using railcars through train transportation, with the railcars being more difficult to fill due to
their larger size. This division of flows reduces the possibilities of consolidation. Again, the fact that we model only
flows with two out of the top five FMCG retailers in France and did not include all smaller retailers and suppliers
reduces the potential of open consolidation in a PI setting. So these results are to be considered conservative.
Scenario 3 has a lower fill rate than scenarios 1 and 2 but still is better than scenario 0 by about 6%. This is due to
the absence of warehouses allowing consolidation of flows through the exploitation of large inventory. In scenario 3,
truck fill rates diminish because full truckloads from plants to distribution centres nearly do not exist. This a compro-
mise between on one side the reduction of transport fill rates and on the other side total system sojourn time reduction,
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

stock reduction and facilities operating cost reduction due to the elimination of stays in warehouses and the direct con-
tact between plants and distribution centres. Again, the focus on two retailers only limits the consolidation potential out
of the factories that in reality feed the five largest retailers as well as the many smaller retailers.
Globally, the obtained results show a significant fill rate progress (up to 17%) that represents an important gain, and
this despite several unfavourable factors: order frequency not adjusted to the size of containers, the use of heavier means
more difficult to fill in scenarios 2 and 3 due to partial FMCG industry modelling, and flow more fragmented when we
do not use warehouses as in scenario 3.

5.2.3 Mean Number of Visited hubs per container route


A perceived possible drawback of the PI stems from the shift from point-to-point transport to distributed transport
involving having the trucks and containers being transshipped through a number of hubs, inducing potentially significant
costs and delays. Results depicted in Figure 17 show an average of 2.43 hubs and warehouses included in the route
taken by a container from source to final destination. In scenarios 0, 1 and 2, there is an average of 0.44 warehouse
transitions per pallet or PI-container. This transition is not needed anymore in scenarios three. Scenarios 2.3.E and 3.3.E
justified an extra hub to reach trains, so as to optimise emissions.
The minimum number of transshipments is reached in scenarios focusing on moving PI-containers as quickly as pos-
sible to their destinations and limiting the number of loading and unloading activities in the intermediate hubs (scenario
2.1.T).

5.2.4 Delivery travel time


Regarding delivery travel time, we observe (Figure 18) that in the PI scenarios, deliveries require an average extra
2-to-4-h, except in scenario 3. This difference of up to 4 h of transportation time in one-or-two-day lead time is not
significant. In general, shipments are almost systematically delivered within 24 h in the PI simulations.
Scenario 3 should also be considered differently since it should at least be compared to two transportations
(upstream and downstream of the warehouse), without counting the time of storage in the warehouse. In absolute terms,
scenario 3 offers much lower delivery times as it avoids the transit through a warehouse.

5.2.5 Costs
Throughout the entire experiment, we have taken into consideration the actual cost parameters for truck transportation,
rail transportation, transshipment, loading and unloading of PI-containers, as well as PI-container rental. We have been
conservative in not using lower costs in the PI context, such as for handling, even though they are expected to be lower.
We also conservatively assumed that loading and unloading of pallets in truck is about the same price as loading and
unloading the PI-containers itself.
Figure 19 shows that the PI scenarios result in significant lower costs. Even scenario 1.2.C which uses only small
containers, thus increasing the costs associated to handling and containerisation, is cheaper than scenario 0. In addition,
the benefits on stocks will accentuate the cost difference between these two scenarios.
Scenario 2 is more efficient in terms of costs than scenario 1, mainly due to the fact that trains have cheaper energy
consumption than gas-oil fuelled trucks. Scenario 3 is more cost efficient because we avoid detours to warehouses, and
3206 R. Sarraj et al.

the difference with other scenarios will be larger if we include storage cost reduction from the elimination of
warehouses.
In addition, we used our model to change the costs to check the robustness of the PI paradigm. For instance, we are
able to change gas-oil price, container handling price, etc. The dominance of the PI was proved with important changes
of prices, for instance, with the cost of 150€ to transship a 12 m container in an inland hub. Furthermore, increases of
gasoil price are in favour of the PI.

5.2.6 Greenhouse gas emissions


The results relative to greenhouse gas emission, here assessed through CO2 emission, relate only to those due to trans-
portation. Two means of transportation are considered in our simulator: truck (25-tonne semi-trailer) and electrical train.
To evaluate their CO2 emissions, we use the emissions evaluation functions (Table 2) developed in (Pan, Ballot, and
Fontane 2013) which are based on the MEET report (Jorgensen and Sorenson 1998; Hickman et al. 1999).
In France, electrical trains emit much less than trucks because of the low-carbon electrical power used in the coun-
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

try. The reduction of CO2 emissions was the most significant among all the achievements of the PI scenarios (Figure 20)
in this experiment. In fact, thanks to the relative increase in fill rate of up to 17% and despite additional kilometres, we
note a significant reduction in emissions even when using only road transport.
When the electric railroad means are used, it allows a drastic reduction of the order for up to 60% of CO2 emissions.
This reduction is better than the one given by pooling approach (between 14 and 50% according to Pan, Ballot, and
Fontane (2013)) as it expands significantly the number of actors involved and it builds on a network of open hubs,
creating more possibilities to consolidate flows and reduce overall travel.
Even more than collaborative pooling approaches, the PI paradigm has the ability with its open pooling to concur-
rently reduce CO2 emissions and costs without concessions on the shipment size and with little extra transshipments. It
is important to notice that this 60%-reduction result corresponds to the target of the European Commission by 2050.
Scenarios 2 and 3 allow rail transportation to account between 37 and 57% in terms of travelled t.km, as compared
to an actual negligible share. These results relative to the environmental contribution of the PI, though quite important,
still have margins for further progress related to the volume and the reconfiguration of supply chains. Notably, the
increase of volumes by adding numerous other actors would allow the increase of the share of multimodal
transportation.

6. Conclusion and further work


In this work we have demonstrated, using a simulation experiment based on real-world orders from a FMCG-sector
database, that the PI concept could significantly improve transportation efficiency and sustainability. The combination of
an open hub network and a set of modular containers makes the difference as compared to the current organisation even
if we impose a limited choice of container sizes. The fill rate of transportation means is increased by almost 17%, the
share of rail transportation becomes significant and leads to a 60% reduction of CO2 emissions (thanks to low carbon
production of electricity), without compromising lead times or jeopardising the operational costs. Additionally, all sce-
narios come with significantly lower costs as compared to the status quo scenario, which means there is no need to fully
implement all the elements of the PI concept to start collecting the benefits.
In this paper, we provided evidence that the PI is very efficient within the FMCG supply networks of two large
retailers and we expect that the impact will increase significantly as the network increases with more retailers and sup-
pliers. Also very potent is the extension from stopping at the distribution centre level, as was done in this paper, to
incorporating deliveries to retail outlets. There is also wide open important research opportunities in testing PI potential
in other sectors, such as the automotive industry, that have very different organisations and specifications.
From a research point of view, it is interesting to note that the results from our experiment were obtained without
extra coordination between logistics service providers, except relative to the information publication required for routing.
Also, all shipments were sent independently and thus asynchronously. Our work opens numerous research avenues. The
first one is the improvement of the routing algorithms to make them more dynamic and efficient by taking into account
resource-congestion feedback effects on lead times. The second one is to allow several levels of containerisation. Small
PI-containers could be filled in the plant, then dynamically grouped and ungrouped to deliver the shops. The third one
is to investigate novel supply and inventory deployment methods to take advantage of the PI capability to spread
PI-containers all over interconnected networks. Investigating how this capability can be implemented and instrumented,
as well as its potential impact on logistic and supply chain efficiency, requires to expand from the transportation and
handling levels to incorporate the inventory and purchase levels of supply chains.
International Journal of Production Research 3207

The research reported in this paper has exploited simulation modelling so as to provide an assessment of the PI
potentiality. Modulushca, a recently launched university-industry, three-year, ten-country, sixteen-partner collaborative
European project is undertaking a next phase of investigation of the potential of PI enabled interconnected logistics in
the FMCG sector. It notably aims to realise the first field-based pilot test of the PI. The first results from this large-scale
investigation and experimentation are expected by the end of 2014 and will open further research and innovation ave-
nues.

Funding
This work has been funded by the French Ministry of Transportation under the National Research Program PREDIT grant
10-MT-PREDITGO4-4-CVS-124-20.

References
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

Ademe. 2007. Guide des facteurs d’émissions. Bilan carbone [Emission Factor Guide: Carbon Footprint]. Paris: Agence de l’Envi-
ronnement et de la Maîtrise de l’Energie, Mission Interministérielle de l’Effet de Serre.
Ballot, E., R. Glardon, and B. Montreuil. 2011a. “Openfret: Contribution à la conceptualisation et à la réalisation d’un hub rail-route de
l’internet physique [OpenFreight: Contribution to the Conceptualization and Implementation of a Physical Internet Road-rail
Hub].” In Recherche et innovation dans les transports terrestres: carrefour à mi-parcours du Predit 4, edited by Predit, 252–255.
Paris: Predit.
Ballot, E., O. Gobet, and B. Montreuil. 2011b. “Physical Internet Enabled Open Hub Network Design for Distributed Networked
Operations.” In Service Orientation in Holonic and Multi-agent Manufacturing Control, Studies in Computational Intelligence,
Series vol: 402, edited by T. Borangiu, A. Thomas and D. Trentesaux, 279–292. Berlin: Springer.
Ballot, E., R. Sarraj, S. Pan, F. Fontane, B. Montreuil, D. Hakimi, R. Glardon, M. Thémans, and O. Gobet. 2012. Simulation de
l’internet physique: Contribution à la simulation des enjeux et à sa définition [Physical Internet Simulation: Contribution to its
Definition and Stakes Assessment], 199. Paris: Predit/Ministère De L’écologie, Du Développement Durable Et De La Mer En
Charge Des Technologies Vertes Et Des Négociations Sur Le Climat.
Bellman, R. 1958. “On a Routing Problem.” Quarterly of Applied Mathematics 16 (1): 87–90.
Bouley, C. 2010. Manifeste pour le 45’ palletwide: Un conteneur vert pour l’europe. Paris: Geodis Global Solutions.
Chan, F. T., S. Chung, and S. Wadhwa. 2004. “A Heuristic Methodology for Order Distribution in a Demand Driven Collaborative
Supply Chain.” International Journal of Production Research 42 (1): 1–19.
Coffman Jr, E. G., M. R. Garey, and D. S. Johnson. 1996. “Approximation Algorithms for Bin Packing: A Survey.” In Approximation
Algorithms for NP-hard Problems, 46–93. Boston, MA:PWS.
Comer, D. 2006. Internetworking with TCP/IP. 5th ed. Pearson Prentice Hall: Upper Saddle River, NJ.
Cruijssen, F., W. Dullaert, and H. Fleuren. 2007a. “Horizontal Cooperation in Transport and Logistics: A Literature Review.” Trans-
portation Journal 46 (3): 22–39.
Cruijssen, F., M. Cools, and W. Dullaert. 2007b. “Horizontal Cooperation in Logistics: Opportunities and Impediments.” Transporta-
tion Research Part E: Logistics and Transportation Review 43 (2): 129–142.
Dechter, R., and J. Pearl. 1985. “Generalized Best-first Search Strategies and the Optimality af A*.” Journal of the ACM 32 (3):
505–536.
Dijkstra, E. W. 1959. “A Note on Two Problems in Connexion with Graphs.” Numerische Mathematik 1 (1): 269–271.
Eurostat. 2007. Average Loads, Distances and Empty Running in Road Freight Transport – 2005 [No. 117/2007]. Luxembourg:
European Communities.
Farooq, M., and G. A. Di Caro. 2008. “Routing Protocols for Next-generation Networks Inspired by Collective Behaviors of Insect
Societies: An Overview.” In Swarm Intelligence, edited by C. Blum and D. Merkle, 101–160. Heidelberg: Springer Berlin.
Fu, L., D. Sun, and L. R. Rilett. 2006. “Heuristic Shortest Path Algorithms for Transportation Applications: State of the Art.” Com-
puters & Operations Research 33 (11): 3324–3343.
Gilmore, P. C., and R. E. Gomory. 1961. “A Linear Programming Approach to the Cutting-stock Problem.” Operations Research
9 (6):849–859.
Gilmore, P. C., and R. E. Gomory. 1963. “A Linear Programming Approach to the Cutting Stock Problem-part II.” Operations
Research 11 (6):863–888.
Hickman, J., D. Hassel, R. Joumard, Z. Samaras, and S. Sorenson. 1999. Methodology for Calculating Transport Emissions and
Energy Consumption (Report for Project Meet). Crowthorne: Transport Research Laboratory.
Johnson, D. S., A. Demers, J. D. Ullman, M. R. Garey, and R. L. Graham. 1974. “Worst-case Performance Bounds for Simple
One-dimensional Packing Algorithms.” SIAM Journal on Computing 3 (4): 299–325.
Jorgensen, M. W., and S. C. Sorenson. 1998. Estimating Emissions from Railway Traffic (Report for the Project Meet) Lyngby:
Technical University of Denmark.
3208 R. Sarraj et al.

Joumard, R. 1999. Methods of Estimation of Atmospheric Emissions from Transport: European Scientist Network and Scientific State-
of-the Art Cost 319 Final Report. Bron: Inrets, LTE 9901.
Kamakaté, F., and L. Schipper. 2009. “Trends in Truck Freight Energy Use and Carbon Emissions in Selected OECD Countries from
1973 to 2005.” Energy Policy 37 (10): 3743–3751.
Kellerer, H., U. Pferschy, and D. Pisinger. 2004. Knapsack Problems, 546. Berlin: Springer.
Lee, J.-H., and C.-O. Kim. 2008. “Multi-agent Systems Applications in Manufacturing Systems and Supply Chain Management: A
Review Paper.” International Journal of Production Research 46 (1): 233–265.
Mckinnon, A. 2010. “European Freight Transport Statistics: Limitations, Misinterpretations and Aspirations.” In 5th ACEA a Scientific
Advisory Group Meeting, edited by Acea, 23. Bruxelles: ACEA.
Mckinnon, A., Y. Ge, and D. Leuchars. 2003. Analysis of Transport Efficiency in the UK Food Supply Chain. Edinburgh: University,
H.-W.
Montreuil, B. 2009. “Physical internet manifesto: Globally transforming the way physical objects are handled, moved, stored, realized,
supplied and used.” [online]. Accessed 2011. www.physicalInternetInitiative.org
Montreuil, B. 2011. “Towards a Physical Internet: Meeting the Global Logistics Sustainability Grand Challenge.” Logistics Research
3 (2–3): 71–87.
Montreuil, B., R. D. Meller, and E. Ballot. 2010. “Towards a Physical Internet : The Impact on Logistics Facilities and Material Han-
Downloaded by [University of Wisconsin-Milwaukee] at 06:28 17 October 2014

dling Systems Design and Innovation.” In Progress in Material Handling Research, edited by K. Gue, 305–327, Charlotte,
NC: Material Handling Industry of America.
Montreuil, B., R. D. Meller, and E. Ballot. 2013. “Physical Internet Foundations.” In Service Orientation in Holonic and Multi Agent
Manufacturing and Robotics, edited by T. Borangiu, A. Thomas, and D. Trentesaux, 151–166. Springer.
Naesens, K., L. Gelders, and L. Pintelon. 2009. “A Swift Response Framework for Measuring the Strategic Fit for a Horizontal Col-
laborative Initiative.” International Journal of Production Economics 121 (2): 550–561.
O’Cearbhaill, E. A., and M. O’Mahony. 2005. “Parallel Implementation of a Transportation Network Model.” Journal of Parallel and
Distributed Computing 65 (1): 1–14.
Pan, S. 2010. “Contribution à la définition et à l’évaluation de la mutualisation de chaînes logistiques pour réduire les émissions de
co2 du transport: Application au cas de la grande distribution.” PhD, Mines ParisTech.
Pan, S., E. Ballot, and F. Fontane. 2013. “The Reduction of Greenhouse Gas Emissions from Freight Transport by Pooling Supply
Chains.” International Journal of Production Economics 143 (1): 86–94.
Piecyk, M. I., and A. C. Mckinnon. 2010. “Forecasting the Carbon Footprint of Road Freight Transport in 2020.” International Jour-
nal of Production Economics 128 (1): 31–42.
Resende, M. G., and P. M. Pardalos. 2006. Handbook of Optimization in Telecommunications. New York: Springer-Verlag.
Ruiz, N., A. Giret, V. Botti, and V. Feria. 2011. “Agent-supported Simulation Environment for Intelligent Manufacturing and Ware-
house Management Systems.” International Journal of Production Research 49 (5): 1469–1482.
Sarraj, R., E. Ballot, S. Pan, and D. Hakimi. Forthcoming. “Analogies between Internet Network and Logistics Service Networks:
Challenges Involved in the Interconnection.” Journal of Intelligent Manufacturing. doi:10.1007/s10845-012-0697-7.
Serveau, L. T. 2011. Inventaire des émissions de polluants dans l’atmosphère en France [Inventory of Poluants Emissions into the
Atmosphere in France]. Paris: CITEPA/format SECTEN.
Simchi-Levi, D. 1994. “New worst-case Results for the Bin-packing Problem.” Naval Research Logistics 41 (4): 579.
Van Der Zee, D. 2006. “Modeling Decision Making and Control in Manufacturing Simulation.” International Journal of Production
Economics 100 (1): 155–167.

You might also like