This art icle was downloaded by: [ Nat ional I nst it ut e of Technology ]

On: 20 August 2011, At : 22: 01

Publisher: Taylor & Francis
I nforma Lt d Regist ered in England and Wales Regist ered Number: 1072954 Regist ered office: Mort imer House,
37- 41 Mort imer St reet , London W1T 3JH, UK
Int ernat i onal Journal of Syst ems Sci ence
Publ i cat i on det ai l s, i ncl udi ng i nst r uct i ons f or aut hor s and subscr i pt i on i nf or mat i on:
ht t p: / / www. t andf onl i ne. com/ l oi / t sys20
Di st ri but ed opt i mi sat i on of a l ogi st i c syst em and i t s
suppl i ers usi ng ant col oni es
C. A. Si l va
a
, J. M. C. Sousa
a
, T. A. Runkl er
b
& J. M. G. S da Cost a
a
a
Depar t ment of Mechani cal Engi neer i ng/ GCAR - IDMEC Av. , Rovi sco Pai s, Inst i t ut o Super i or
Tcni co, 1049- 001 Li sbon, Por t ugal
b
Inf or mat i on and Communi cat i ons CT IC4, Si emens AG Cor por at e Technol ogy, 81730
Muni ch, Ger many
Avai l abl e onl i ne: 02 Sep 2006
To ci t e t hi s art i cl e: C. A. Si l va, J. M. C. Sousa, T. A. Runkl er & J. M. G. S da Cost a ( 2006) : Di st r i but ed opt i mi sat i on of a
l ogi st i c syst em and i t s suppl i er s usi ng ant col oni es, Int er nat i onal Jour nal of Syst ems Sci ence, 37: 8, 503- 512
To l i nk t o t hi s art i cl e: ht t p: / / dx. doi . or g/ 10. 1080/ 00207720600784452
PLEASE SCROLL DOWN FOR ARTI CLE
Full t erms and condit ions of use: ht t p: / / www. t andfonline. com/ page/ t erms- and- condit ions
This art icle may be used for research, t eaching and privat e st udy purposes. Any subst ant ial or syst emat ic
reproduct ion, re- dist ribut ion, re- selling, loan, sub- licensing, syst emat ic supply or dist ribut ion in any form t o
anyone is expressly forbidden.
The publisher does not give any warrant y express or implied or make any represent at ion t hat t he cont ent s
will be complet e or accurat e or up t o dat e. The accuracy of any inst ruct ions, formulae and drug doses should
be independent ly verified wit h primary sources. The publisher shall not be liable for any loss, act ions, claims,
proceedings, demand or cost s or damages what soever or howsoever caused arising direct ly or indirect ly in
connect ion wit h or arising out of t he use of t his mat erial.
International Journal of Systems Science
Vol. 37, No. 8, 20 June 2006, 503512
Distributed optimisation of a logistic system
and its suppliers using ant colonies
C. A. SILVA*y, J. M. C. SOUSAy, T. A. RUNKLERz and
J. M. G. SA

DA COSTAy
yDepartment of Mechanical Engineering/GCAR - IDMEC Av.
Rovisco Pais, Instituto Superior Te cnico, 1049-001 Lisbon, Portugal
zInformation and Communications CT IC4, Siemens AG,
Corporate Technology, 81730 Munich, Germany
(Received 1 October 2004; accepted 15 April 2005)
This paper introduces a new multi-agent approach for collaborative management of logistic
and supply systems based on the ant colony optimisation (ACO) meta-heuristic. The logistic
system and its suppliers can be modelled as partners of a supply chain. The management
methodology is defined as a set of distributed scheduling problems that exchange information
during the optimisation process. Each problem is solved by an ant colony agent that uses
the pheromone matrix as the communication platform. A simulation example shows that
the proposed coordination mechanism improves the supply-chain performance compared
to a traditional management approach, where both problems are considered separately.
Keywords: Ant colonies; Supply-chain system; Distributed optimisation
1. Introduction
An increasing number of companies are adopting the
supply-chain organisation methodology in order to
improve competitiveness and profitability (Beamon
1998, Mentzer 2004). These companies are organised
as a world-wide network of external partners (suppliers,
warehouses and distribution centres) through which
raw materials are acquired, transformed into products
and delivered to costumers (Barbuceanu and Fox 1996).
The companys job is no longer to produce the goods,
but manage all the different partners in a coordinated
manner so that in the end, the customer receives a
quality product at a certain desired date.
The different partners in a supply chain operate under
different sets of objectives and constraints. However, the
systems are highly interdependent and the optimisation
of objectives such as on-time deliveries or costs of one
partner will influence the performance of the remaining
partners (Ridallis et al. 2000). The supply chain is
purely a distributed system with several independent
but coupled optimisation problems that have to be
solved in parallel. The coherence between the different
decision-making centres can be accomplished by a multi-
agent framework (Shen and Norrie 1999). These systems
are based on explicit communication between specialised
agents assigned to structural elements of the chain,
e.g. supplying or logistic agents. They exchange
information about their tasks, like scheduling or
routing plans, using an interaction protocol with
specific message structure, conversation rules, action
and reaction behaviours (Barbuceanu and Fox 1996,
Swaminathan et al. 1998).
The research in this field has tackled mainly the
interaction between the agents and the optimisation
issues are usually solved through some simple heuristic
(Barbuceanu and Fox 1996). However, heuristics are
usually not sufficient to deal with the complexity of the
real-world problems (Pinedo 2002) and the agents need
to use more powerful optimisation techniques.
Moreover, to take full advantage of the supply-chain
framework, the communication protocols should *Corresponding author. Email: csilva@dem.ist.utl.pt
International Journal of Systems Science
ISSN 00207721 print/ISSN 14645319 online 2006 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/00207720600784452
D
o
w
n
l
o
a
d
e
d

b
y

[
N
a
t
i
o
n
a
l

I
n
s
t
i
t
u
t
e

o
f

T
e
c
h
n
o
l
o
g
y

]

a
t

2
2
:
0
1

2
0

A
u
g
u
s
t

2
0
1
1

support the possibility of exchanging information
during the optimisation processes to allow agents to
react to failures or other type of dynamic disturbances.
This paper introduces a novel multi-agent supply-
chain management based on the description of the
supply chain as a set of different distributed optimisa-
tion problems and using the ant colony optimisation
(ACO) meta-heuristic as agents (Dorigo et al. 1996,
Dorigo and Stu tzle 2004). The ACO algorithm builds a
pheromone matrix, which is an indirect record of the
optimisation steps, that is manipulated at all times
during the optimisation process. If each system in a
supply chain is optimised by its own ant colony agent,
the pheromone matrices can be used to exchange the
information between the different agents as a multi-
agent system, introducing in this way a coordination
mechanism between the different systems in the supply
chain.
This new concept is presented for a supply chain-
system with a logistic system and its suppliers, which is
described in section 2. Section 3 models the management
problem as a distributed optimisation problem and
section 4 shows how the ACO can be used to solve this
problems as a multi-agent system. The simulation results
are presented in section 5 and section 6 concludes the
paper and presents the possible future work.
2. Description of a supply chain
A common supply-chain configuration includes at least
a logistic system and its suppliers. The logistic system is
the core of the company and, for the clients, it represents
the whole company: it collects the orders from the
customers, purchases the components from external
suppliers and schedules the components delivered by the
suppliers as orders. The supply system is a network of
external suppliers that manufactures the components
and delivers them to the logistic system.
2.1 Logistic system
At each day, the logistic system has an order list O of n
orders waiting to be delivered. An order o
j
2 O with
j 1, . . . , n is a set of m different types of items, called
the components c
i
with i 1, . . . , m, in certain quantities
q
ij
. Therefore, an order can be defined as an m-tuple
o
j
(q
1j
, . . . , q
mj
). When a new order o
j
arrives, it receives
two labels: the arrival data or release date r
j
and the
desired delivery data or due date d
j
, which is the date
when the client wishes to receive the order. The order is
delivered at the completion date C
j
.
Assuming that the system does not deliver orders if
they are ready before the due date, the difference
between the completion date and the due date is called
the tardiness T
j
C
j
d
j
. The objective is to match both
dates, i.e. to have for all orders T
j
0. However, two
disturbances may influence the system: the fact that the
suppliers service may not be respected and the fact
that some clients ask for desired delivery dates not
compatible with the supplier services.
In order to define the cost function that best describes
the objective of the logistic system, consider the
following definitions. Given the set O of orders in
the system waiting to be delivered, the subset of orders
that are going to be delivered is defined as O
D
O.
The complementary subset of orders that are not
delivered and remain in the system is defined as
O
ND
O, such that O
D
[ O
ND
O. Consider further
that the subset of orders that are delivered at the correct
date is defined as O
0
D
O
D
and the subset of orders that
are not delivered and are already delayed is defined as
O
d
ND
O
ND
. The optimisation objective is to minimise
the cost function given by
f
L

P
j2O
T
j
jO
d
ND
j
jO
0
D
j c
1
where
P
j2O
T
j
accounts for the minimisation of the
tardiness of the total set of orders in the system O; jO
d
ND
j
is the cardinality of the subset O
d
ND
and refers to the
minimisation of the number of orders that are not
delivered and are already delayed; and finally jO
0
D
j c is
the cardinality of the subset O
0
D
and accounts for the
maximisation of the number of orders delivered at
the correct date. The c is a small constant that avoids the
infinity value when no orders are delivered at the correct
date. This decision step is done once per day, but
different solutions for the same daily problem originate
different next-day scheduling problems. The supply-
chain management is a dynamic succession of daily
optimisation problems that are treated independently,
even though they are not. Therefore, to evaluate the
performance of the supply chain, larger periods of time,
such as weeks or months, should be considered.
Figure 1 presents an example of a supply chain that
sells all possible combinations of m5 different types of
components in any quantity. Consider that at day 1,
the logistic system receives: order o
1
(1, 0, 3, 0, 0),
that consists of one item of component c
1
and three
items of component c
3
, with due date d
1
1; and
o
2
(0, 1, 0, 2, 1), that consists of one item of
component c
2
, two items of component c
4
and
one item of component c
5
, with due date d
2
2.
These components enter a purchase list that is sent to
the suppliers.
At the same day, based on the purchase list, supplier
M
1
delivers one item of components c
1
, c
2
and c
3
and
supplier M
2
delivers two items of component c
4
.
504 C. A. Silva et al.
D
o
w
n
l
o
a
d
e
d

b
y

[
N
a
t
i
o
n
a
l

I
n
s
t
i
t
u
t
e

o
f

T
e
c
h
n
o
l
o
g
y

]

a
t

2
2
:
0
1

2
0

A
u
g
u
s
t

2
0
1
1

The logistic system is not able to deliver any order.
At day 2, two new orders enter the system: o
3
(0, 2, 0,
0, 0), that consists of two items of component c
2
; and
o
4
(1, 0, 0, 1, 0), that consists of one item of
component c
1
and one item of component c
4
. At that
day, supplier M
1
delivers two items of component c
3
and
one item of component c
1
and supplier M
2
delivers one
item of component c
5
. Then it is possible for the logistic
system to deliver orders o
1
and o
2
at the completion date
C
1
C
2
2, with T
1
C
1
d
1
1 and T
2
C
2
d
2
0.
Notice that the supply-chain performance is mainly
determined by the logistic system performance.
However, this performance highly depends on the
performance of the supplying system, as well as on the
scheduling method.
2.2 Supply system
The supplying sub-system is a network of s different
suppliers or manufacturers M
i
with i 1, . . . , s, each one
producing its own set of jobs J
M
i
, where each job
refers to a type of component c
i
requested by the logistic
sub-system. Each supplier is independent and therefore
it optimises its own problem called the local supplier
problem. However, from the point of view of the logistic
system, the supplier can be virtually considered as one
single entity, and the optimisation problem is called the
global supplier problem.
2.2.1 Local supplier problem. The optimisation pro-
blem of each M
i
supplier can be modelled as a single
machine scheduling problem (Pinedo 2002). There is one
machine that produces all the n
i
jobs on the waiting
list J
M
i
of the supplier. The objective function to be
minimised is the total tardiness of all the jobs
f
M
i

X
j2J
M
i
T
j
2
where tardiness is defined as in the logistic system.
Figure 2(a) represents the local problems for each of the
suppliers M
1
and M
2
at day 1: supplier M
1
produces the
jobs c
1
, c
2
, c
3
of the list J
M
1
fc
1
, c
2
, c
3
, c
3
, c
3
g; supplier
M
2
produces the two jobs of type c
4
of the list
J
M
2
fc
4
, c
4
, c
5
g.
2.2.2 Global supplier problem. The global supplier
problem describes the supply system from the logistics
point of view. The problem can be modelled as an open
job-shop problem (Pinedo 2002) with an s number of
machines, where each machine corresponds to a
different supplier. In this problem, the set of the
machines produces all jobs J J
M
1
[ . . . J
M
s
, but each
machine i can only produce the jobs J
M
i
. The objective is
to minimise the logistics system objective function f
L
as
defined in equation (1).
Figure 2(b) represents the global problem at day 1.
From the logistic point of view, there are two machines
producing all type of jobs. However, machine M
1
can
only produce jobs J
M
1
fc
1
, c
2
, c
3
, c
3
, c
3
g and machine
M
2
can only produce jobs J
M
2
fc
4
, c
4
, c
5
g. The
advantage of analysing the supply system as a whole is
explained in the following. Imagine that for the supplier
M
1
, producing components c
1
, c
2
, c
3
at day 1 is as
optimal as producing components c
1
, c
3
, c
3
or c
2
, c
3
, c
3
.
From the logistic point of view, only the second case
leads to an optimal solution in terms of logistic process:
if supplier M
1
produces components c
1
, c
3
, c
3
, the
logistic system is able to deliver order o
1
with T
1
0
Figure 1. Supply-chain behaviour: (a) Supply chain at day 1; (b) Supply chain at day 2.
Distributed optimisation of logistic system 505
D
o
w
n
l
o
a
d
e
d

b
y

[
N
a
t
i
o
n
a
l

I
n
s
t
i
t
u
t
e

o
f

T
e
c
h
n
o
l
o
g
y

]

a
t

2
2
:
0
1

2
0

A
u
g
u
s
t

2
0
1
1

at day 1 and deliver order o
2
with T
2
0 at day 2.
Otherwise, as represented in figures 1 and 2(a), the
solution is non-optimal and order o
1
is delivered with
delay, i.e. T
1
1. In this case, the role of the supply-
chain management system is to lead supplier M
1
to
choose, among the several optimal solutions, the only
solution that is optimal for the logistic system.
3. Distributed optimisation in supply-chain
management
In general, the suppliers are independent manufacturers
that provide the same type of components for different
companies organised as supply chains. In this way, their
manufacturing plans are determined by some scheduling
tool, working on a local supplier problem as previously
described.
The logistic system can, however, influence the
suppliers local optimisation, through the algorithm
presented in algorithm 1. Consider that the supply chain
previously described is a distributed system
S fS
L
, S
M
1
, . . . , S
M
s
g 3
where S
L
is the logistic system and S
M
i
with i 1, . . . , s
are the suppliers. Consider further that f
L
and f
M
i
are
the respective cost functions, as defined in (1) and (2).
If the logistic system is able to exchange information
with the suppliers, it can provide them the solution of
the global supplier problem, i.e. what would be the best
solution from the logistic point of view. The suppliers
read this information and try to adapt their local
solution according to this: if the performance is not
affected, they accept the logistic system suggestion;
otherwise, they reject the suggestion and re-optimise the
process according to the local specifications. This
process can be repeated Z times. After this process,
the suppliers optimise the local problem independently
and deliver the manufactured components to the logistic
system. At the end, the logistic system optimises the
logistic problem based on the available stock.
Algorithm 1 Distributed optimisation
for t 1: Z do
Each system S
M
i
optimises f
M
i
with i 1, . . . , s;
All S
M
i
systems deliver the optimisation results of f
M
i
to system S
L
;
System S
L
optimises f
L
, using the optimisation results
of f
M
i
;
System S
L
delivers the optimisation result of f
L
to all
systems S
M
i
;
end for
Each system S
M
i
optimises f
M
i
with i 1, . . . , s;
System S
L
optimises the cost function f
L
.
This is the concept of distributed optimisation
(Tsitsiklis et al. 1986). If a problem can be described
as a set different but coupled distributed problems, the
optimisation process can allow each problem to use
intermediate results of the other problems, such that at
the end, the global problem is optimised according to
the solutions of all the local problems. The next section
shows how this procedure can be easily implemented
using the ACO algorithm.
4. Ant colony optimisation
The ACO is an optimisation method suited to find
minimum cost paths in problems described by graphs
(Dorigo and Stu tzle 2004). These techniques have proved
to be very powerful optimisation methods in several
different scheduling problems, such as routing problems
(Gambardella et al. 1999), job-shop problems (Blum and
Sampels 2004) or logistic problems (Silva et al. 2002).
Consider a problem with n nodes and a colony of g
ants. Initially, the g ants are randomly placed in g n
different nodes. The probability that an ant k in node i
chooses node j as the next node to visit is given by
p
k
ij
t
t
a
ij
j
[
ij
P
n
r, 2
t
o
ir
j
[
ir
if j 2
0 otherwise
8
>
>
>
<
>
>
>
:
4
Figure 2. Supply system: (a) Local supplier problem; (b) Global supply problem.
506 C. A. Silva et al.
D
o
w
n
l
o
a
d
e
d

b
y

[
N
a
t
i
o
n
a
l

I
n
s
t
i
t
u
t
e

o
f

T
e
c
h
n
o
l
o
g
y

]

a
t

2
2
:
0
1

2
0

A
u
g
u
s
t

2
0
1
1

where t
ij
and j
ij
are the entries of the pheromone
concentration matrix t and of the heuristic function
matrix j, respectively, for the path (i, j). The pheromone
matrix values are limited to ]t
min
, t
max
], with t
min
0
and t
max
1. is the tabu list, which acts as the memory
of the ants and contains all the trials that the ants have
already passed and cannot be chosen again. The
parameters o and [ measure the relative importance of
trial pheromone and heuristic knowledge, respectively.
After a complete tour t, when all the g ants have
visited all the n nodes, the pheromone concentration
in the trails is updated by
t
ij
t 1 t
ij
t 1 , t
q
ij
5
where , 2 0, 1 expresses the pheromone evaporation
phenomenon and t
q
ij
are pheromones deposited on the
trials (i, j) followed by ant q that found the best solution
f
q
for this tour:
t
q
ij

1
f
q
if arc i, j is used by the ant q
0 otherwise
8
<
:
6
The algorithm stops after a pre-defined number of N
iterations.
4.1 ACO for the logistic problem
In the scheduling problem of the logistic system, the n
orders waiting to be delivered are the nodes of the graph,
and the role of the ants is to find the minimum cost
path connecting the orders that should be delivered.
We consider that each ant is carrying a bag with the
available stock and distributes it between the orders that
it is visiting. It only visits orders whose components
are able to deliver, building in this way only feasible
solutions. When the stocks bag is empty or the
remaining components are not sufficient to deliver any
missing order, the search performed by an ant is
complete.
The heuristic function j is the orders tardiness: if an
order has already a positive tardiness, the ant will feel
a stronger attraction to visit it, because the order is
already delayed. We define this function as an expo-
nential function in the interval [0, 1] where the value 0
corresponds to the order that has the minimum
tardiness min(T) and 1 to be most delayed order
max(T) (Silva et al. 2003). The objective is that the
orders already delayed attract ants much more than
the orders not yet delayed:
j
j

e
fT
j
minT,maxTminTg
1
e 1
7
Notice that in this case the heuristic information is only
order-dependent, therefore j
j
j
ij
. Since the pheromone
trails t
ij
are also restricted to the interval [0, 1], o <[
will indicate a higher relative weight of the pheromones
trail. The tabu list is the list of orders already delivered
by the ant and also the orders which are not possible to
visit, due to lack of stocks. The objective function to be
minimised by each ant k is f
L
as defined in equation (1).
4.2 ACO for the local supplier problem
In this problem, there are n
i
jobs J
M
i
f1, . . . , n
i
g to be
done in one machine. The nodes of the graph are the
jobs to be done plus a dummy source node representing
the jobs queue. The objective function f
M
i
to be
minimised is the one defined in equation (2). It is
necessary to define a production time vector P, with size
(n
i
), describing the processing time of the n
i
jobs at the
machine.
The heuristic matrix j is defined in this case as the
Earliest Due Date (Pinedo 2002), where orders are
scheduled in increasing order of due dates d
j
and then
normalised to the interval [0, 1]:
j
j

mind d
j
mind maxd
8
The tabu list gives the jobs already assigned to the
machine. The algorithm starts with all the ants placed in
the source node. Then, from the list of jobs to do, each
ant will choose the next node to visit, i.e. the next job to
place in the machines. The tabu list is updated and the
algorithm proceeds until all the jobs have been assigned.
4.3 ACO for the global supplier problem
In this problem, there are n
P
s
i1
n
i
jobs to be
performed in the s machines. In this case it is necessary
to define the technological matrix Y with size n s
describing the sequence of operations of the n jobs in
each machine. Since each job is only produced at one
machine, the technological matrix Y is a sparse matrix
with entries equal to 1 when job j has to be done in
machine i. The production P is now an n s matrix
describing the production times of job j 2 J
J
M
1
[ [ J
M
s
at machine i. The number of nodes is
(n s 1) representing all the possible assignments of
jobs to machines plus the source node. Notice however
that some of these nodes do not correspond to real
assignments. The heuristic function is defined as in
equation (8) and the objective function to be minimised
is the one defined in equation (1). The algorithm follows
the same sequence as for the local supplier problem,
and at the end the scheduling result describes the
Distributed optimisation of logistic system 507
D
o
w
n
l
o
a
d
e
d

b
y

[
N
a
t
i
o
n
a
l

I
n
s
t
i
t
u
t
e

o
f

T
e
c
h
n
o
l
o
g
y

]

a
t

2
2
:
0
1

2
0

A
u
g
u
s
t

2
0
1
1

scheduling plans of each of the machines, i.e. the
suppliers.
4.4 ACO for the distributed optimisation problem
The ACO is by definition a multi-agent system: several
autonomous agents solving its own problem, commu-
nicating among themselves over the pheromone matrix
t
ij
and solving the global system through cooperation.
In this paper, we propose another level of abstraction
which is the cooperation between colonies of ants: the
colonies associated with each system use a common
pheromone matrix to inform and to be informed about
the different optimisation problems. In this way, each
colony is solving its problem autonomously, considering
relevant information of the remaining colonies.
Consider the supply chain example described in
section 2, represented in figure 1, and the distributed
optimisation algorithm proposed in algorithm 1. At day
1, the logistic system purchases components {c
1
, c
2
, c
3
,
c
3
, c
3
} to supplier M
1
and components {c
4
, c
4
, c
5
} to
supplier M
2
. Suppliers M
1
and M
2
start in parallel
independent optimisation processes with the ACO,
during which pheromone matrices t
M
1
and t
M
2
are
created. At a certain moment in time, the logistic system
starts an optimisation processes to solve the global
supplier problem, using the matrix t
L
, which is a
composition of matrices t
M
1
and t
M
2
, given by:
t
L

t
M
1
0
0 t
M
2

. 9
At the end of this optimisation process, suppliers M
1
and M
2
receive the information from the logistic system
and utilise it to re-optimise their own processes. The
pheromone matrices used by each of the suppliers to
solve the local problem are the updated sub-matrices t
M
1
and t
M
2
of the pheromone matrix t
L
. This exchange of
information occurs Z times, until suppliers M
1
and M
2
define their last scheduling solution. Notice that the last
optimisation procedure is always done by the suppliers
themselves, guaranteeing in this way that the final
solution satisfies their own local objectives. At the end,
the suppliers deliver the components to the logistic
system, that starts the optimisation process of the
function f
L
in (1). Figure 3 shows how the information
is exchanged between the different systems. The
continuous arrows indicated the information exchanged
during the distributed optimisation process, while the
dotted arrows indicate the exchange of other type of
information, such as components lists. For the sake of
simplicity, this work considers that the systems exchange
information only once, i.e. Z1.
5. Example
To demonstrate the potential of the proposed manage-
ment technique, we consider simulation examples of
a supply chain as described in section 2, with a logistic
system and its suppliers. The distributed optimisation
technique is compared to a traditional management
approach, the independent optimisation technique.
This methodology considers that each system is opti-
mised independently: first, the suppliers generate their
planning according to their best interests; afterwards, the
logistic system schedules the delivered orders based on
the components produced and delivered by the suppliers.
The methodologies are tested on three instances that
describe different supply-chain working conditions,
based on four different parameters:
. A average number of orders that enter the system
every day;
. B number of different types of components
produced by the suppliers;
. C average number of different types of components
existing in each order;
. D maximum quantity of each type of component
per order.
The designation of each instance results from the
combination of the four parameters into A-B-C-D.
In this way, we have defined the following instances:
55310, which represents a small instance, with a large
overlap of components between the orders and therefore
with a high optimisation potential; 1010230, which
represents a large instance, with small overlap of
components between the orders and therefore, with
Purchase list M
1
Purchase list M
2
Supplier M
1

M
1

M
2

M
1

M
2
Z iterations
Logistics Logistics Logistics
Components
Components
Supplier M
1
Supplier M
2
Supplier M
2
Supplier M
1
Figure 3. Information flow between the logistic system and the supplying system with two suppliers.
508 C. A. Silva et al.
D
o
w
n
l
o
a
d
e
d

b
y

[
N
a
t
i
o
n
a
l

I
n
s
t
i
t
u
t
e

o
f

T
e
c
h
n
o
l
o
g
y

]

a
t

2
2
:
0
1

2
0

A
u
g
u
s
t

2
0
1
1

low optimisation potential; 2010520, that represents
a large instance, with a large overlap of components
between orders and that mimics a real-world instance
(Silva et al. 2003). The instances describe data from
60 working days, considering empty orders, purchase
and stock lists at day 1. The results presented here
describe the supply chain between the fictive days 31 and
60, in order to test the management techniques in chains
working with stationary conditions, with full lists of
orders and stocks.
The presentation of results is done in two different
steps. First, the optimisation are presented results at day
31, for instance 55310, considering the distributed
and the independent supply-chain management
approaches. This example clearly shows how the
distributed optimisation concept allows the improve-
ment of the global supply-chain performance. We also
show the optimisation results for day 45, where
the distributed approach does not outperform the
independent method.
Later, we present the management results for all the
instances, considering a whole months problem.
As previously said, the optimisation of the supply
chain is a daily problem, but the solution determines
the initial conditions of the next-day problem. Thus,
the performance of the system must be analysed
considering longer time units, such as weeks, months
or years.
5.1 Daily optimisation problem
In this section, the performance of the supply-chain
is analysed considering two different scenarios for the
55310 instance: in the first one, the stock of
components is produced and delivered at the logistic
system by the supplier considering only the local
suppliers optimisation (independent management
approach); the second scenario considers that the stock
is produced following the distributed optimisation
approach.
Figure 4 represents the scheduling result in a Gant
chart for both suppliers M
1
and M
2
at day 31. At this
day, supplier M
1
has two jobs 1 and 2 (defined in the M
1
space), that are expected at the logistic system at day
d
1
32 and d
2
33 respectively. Supplier M
2
has also
two jobs 1 and 2 (defined in the M
2
space), that are
expected to be delivered at the logistic center at day
d
1
d
2
33. The dots represent the desired delivery
dates of the components at the logistic centre.
For each supplier, there are two optimal solutions:
(1) supplier M
1
can produce job 1 followed by job 2 or
job 2 followed by job 1, resulting in each case
f
M
1

P
T
j
0; (2) supplier M
2
can produce job 1
followed by job 2 or job 2 followed by job 1, resulting
also in this case f
M
2

P
T
j
0. At the end, depending
on the combination of the suppliers solution, there are
four possible different sets of components that are
delivered at the logistic centre. Consider further that the
solution is the one represented in figure 4, which might
occur with a probability of 25%.
In terms of the logistic problem, the solution for this
specific day is the one presented in table 1, in the
independent management row. This table shows the cost
function value of the logistic system f
L
, the number of
orders delivered at the correct data jO
0
D
j, the number
of delivered orders already with delay jO
d
D
j and the
maximum tardiness of the delivered orders max(T
j
). The
results show that the logistic system is able to deliver two
orders at the correct date and two orders that were
already delayed one day (max(T
j
) 1).
When the distributed approach is used, the logistic
system induces the suppliers to choose a different
solution. Although any of the possible solutions is
optimal for the suppliers, for the logistic system they
represent four different stock conditions. From those,
the one that leads the system to a better performance is
the case where supplier M
1
produces job 2 followed by
job 1 and supplier M
2
produces job 2 followed by job 1.
This is the solution represented in figure 5. Since this
solution is still optimal for the suppliers, the logistic
system can influence them to adopt this solution as their
Figure 4. Suppliers scheduling solution following an independent management approach at day 31: (a) Supplier M
1
solution;
(b) Supplier M
2
solution.
Distributed optimisation of logistic system 509
D
o
w
n
l
o
a
d
e
d

b
y

[
N
a
t
i
o
n
a
l

I
n
s
t
i
t
u
t
e

o
f

T
e
c
h
n
o
l
o
g
y

]

a
t

2
2
:
0
1

2
0

A
u
g
u
s
t

2
0
1
1

own solution, as represented in the Gant charts on
the right side of figure 5. Observe that those solutions,
for both suppliers M
1
and M
2
are different from
the ones obtained with the independent approach,
represented in the left Gant charts of figure 5. These
solutions drive the logistic system to a much better
performance, as can be seen in table 1, in the
distributed management row. With the distributed
methodology, the logistic system is able to deliver
eight delayed orders, some of which were already two
days delayed (max(T
j
) 2). The logistic system perfor-
mance is improved (lower f
L
value), without compro-
mising the local suppliers performance, which is still
optimal.
Note however, as explained before, that the suppliers
are influenced by the exchange of information only if
their local solution is still optimal. In cases where the
global supplier problem solution decreases the perfor-
mance of the local supplier problem, the suppliers do not
accept the suggestion of the logistic system. Figure 6
shows an example for day d 45. At this day,
both suppliers propose solutions that are locally optimal:
supplier M
1
produces job 3, then job 1 and finally job 2;
supplier M
2
produces job 1 and then job 2. However, the
global supplier problem proposes a different solution for
supplier M
1
: producing job 1 first, then job 3 and finally
job 2. This solution obliges supplier M
1
to produce job 2
with delay, which is not optimal. This can be observed by
the fact that job 2 enters production after the expected
completion date, represented by the grey dot. In this case,
the supplier M
1
does not accept the solution and
produces job 3, then job 1 and only job 2 at the end,
which is the same solution as the one found by the
independent management solution.
Figure 5. Suppliers scheduling solution following a distributed management approach at day 31.
Table 1. Logistic system scheduling results at day
d 31 Instance 55310.
Management ( f
L
) jO
0
D
j jO
d
D
j max(T
j
)
Independent 39.04 2 2 1
Distributed 36.19 2 8 2
510 C. A. Silva et al.
D
o
w
n
l
o
a
d
e
d

b
y

[
N
a
t
i
o
n
a
l

I
n
s
t
i
t
u
t
e

o
f

T
e
c
h
n
o
l
o
g
y

]

a
t

2
2
:
0
1

2
0

A
u
g
u
s
t

2
0
1
1

Remember that the system is dynamic and that on the
next day, new orders will arrive and the problem data
will change. Although the improvement of the solution
for a specific day does not seem outstanding, for the
logistic system and therefore the supply-chain perfor-
mance, the daily improvements can represent significant
improvements in the long term. This analysis is done on
the following section.
5.2 Monthly optimisation problem
The comparison between the independent and distrib-
uted approach for one month is presented in table 2.
This table represents the cost function values for each
system: f
M
1
and f
M
2
are the suppliers costs; and f
L
is the
cost function of the logistic system. Due to the stochastic
nature of the ACO algorithm, the analysis considers
30 runs of each method and the results are presented in
terms of mean and standard deviation o values.
It is possible to observe that the mean value
of the supplier cost function is very similar for both
management methods, for all instances. This means that
the suppliers performance does not change, irrespective
of the independent or the distributed approach used.
The standard deviation values are slightly higher for the
distributed case, which is explained by the fact that
the pheromone matrix values are changed during
the suppliers optimisation process, which introduces
some disturbance to the algorithms convergence.
Nevertheless, we can conclude that the suppliers
performance does not decrease when the distributed
management approach is used.
The logistic system performance significantly
increases when the distributed optimisation method is
used, for all instances. In those cases, the mean value
of the cost function is clearly lower. Further, as the
standard deviations show, the worst solutions obtained
with the distributed method are clearly better than the
best solutions obtained using the independent approach.
In this way, we can conclude that the supply-chain
management can be optimised through distributed
optimisation using the ant colony meta-heuristic.
Figure 6. Suppliers scheduling solution following a distributed management approach at day 45.
Distributed optimisation of logistic system 511
D
o
w
n
l
o
a
d
e
d

b
y

[
N
a
t
i
o
n
a
l

I
n
s
t
i
t
u
t
e

o
f

T
e
c
h
n
o
l
o
g
y

]

a
t

2
2
:
0
1

2
0

A
u
g
u
s
t

2
0
1
1

6. Conclusions
This paper proposes a new multi-agent methodology for
supply-chain management. The methodology is based
on a description of the management problem as a
distributed optimisation problem solved by several ant
colonies. The ant colonies exchange information about
the distributed scheduling problems through the pher-
omone matrix generated during the optimisation proce-
dure. In this way, a multi-agent framework that
coordinates the different systems of the supply chain is
created. A simulation example for a two-partners supply
chain showed how this new management concept can
improve the global systems performance. The next
research steps aim the extension of the methodology
to other supply-chain partners, such as the distribution
system.
Acknowledgements
This work is supported by the German Ministry of
Education and Research (BMBF) under Contract no.
13N7906 (project Nivelli) and by the Portuguese
Foundation for Science and Technology (FCT) under
Grant no. SFRH/BD/6366/2001 and Project POCTI/
EME/59191/2004.
References
M. Barbuceanu and M. Fox, Coordinating multiple agents in the
supply chain, in Proceedings of the Fifth Workshops on Enabling
Technology for Collaborative Enterprises, WET ICE96, IEEE
Computer Society Press, 1996, pp. 134141.
B.M. Beamon, Supply chain design and analysis: models and
methods, International Journal of Production Economics, 55, pp.
281294, 1998.
C. Blum and M. Sampels, An ant colony optimization algorithm for
shop scheduling problems, Journal of Mathematical Modelling and
Algorithms, 3(3), pp. 285308, 2004.
M. Dorigo and T. Stu tzle, Ant Colony Optimization, Cambridge, MA:
MIT Press/Bradford Books, 2004.
M. Dorigo, V. Maniezzo and A. Colorni, The Ant System:
Optimization by a colony of cooperating agents, IEEE
Transactions on Systems, Man, and Cybernetics Part B:
Cybernetics, 26(1), pp. 2941, 1996.
L.M. Gambardella, E. Taillard and G. Agazzi, MACS-VRPTW:
a multiple ant colony system for vehicle routing
problems with time windows, in New Ideas in Optimization (Eds)
D. Corne, M. Dorigo, and F. Glover, McGraw-Hill, 1999,
pp. 6376.
J.T. Mentzer, Fundamentals of Supply Chain Management Twelve
Drivers of Competitive Advantage, London, New Delhi: Sage
Publications, 2004.
M. Pinedo, Scheduling: Theory, Algorithms, and Systems, 2nd ed.,
Prentice Hall, 2002.
C.E. Ridallis, S. Bennett and N.S. Tipi, Modelling the dynamics
of supply chains, International Journal of Systems Science, 31(8),
pp. 969976, 2000.
W. Shen and D.H. Norrie, Agent-based systems for intelligent
manufacturing: a state-of-the-art survey, Knowledge and
Information Systems, an International Journal, 1(2), pp. 129156,
1999.
C.A. Silva, T.A. Runkler, J.M. Sousa and R. Palm, Ant colonies as
logistic process optimizers, in Ant Algorithms, International
Workshop ANTS 2002, Brussels, Belgium (Eds) M. Dorigo,
G.D. Caro, and M. Samples, Springer: Heidelberg, 2002,
pp. 7687. Lecture Notes in Computer Science LNCS 2463.
C.A. Silva, T.A. Runkler, J.M. Sousa and J.M.S. da Costa,
Optimization of logistic processes in supply-chains using meta-
heuristics, in LNAI 2902 (Ed.) F.M. Pires, Berlin Heidelberg:
Springer Verlag, 2003, pp. 923.
J.M. Swaminathan, S.F. Smith and N.M. Sadeh, Modeling supply
chain dynamics: A multiagent approach, Decision Sciences Journal,
29(3), pp. 607632, 1998.
J.N. Tsitsiklis, D.P. Bertsekas and M. Athans, Distributed
asynchronous deterministic and stochastic gradient optimization
algorithms, IEEE Transactions on Automatic Control, 9(31),
pp. 803812, 1986.
Table 2. Supply-chain performance between d 31 and d 60.
Instance Management f
M1
f
M2
f
L
55310 Independent (, o) (0.20, 0.01) (1.37, 0.21) (61.40, 15.33)
Distributed (, o) (0.21, 0.03) (1.35, 0.34) (49.15, 1.41)
1010230 Independent (, o) (2.20, 0.09) (9.90, 0.32) (47.07, 1.64)
Distributed (, o) (2.50, 0.12) (9.50, 0.73) (40.02, 0.67)
2010520 Independent (, o) (0.90, 0.31) (0.00, 0.00) (58.48, 2.14)
Distributed (, o) (1.01, 0.20) (0.02, 0.00) (56.36, 2.69)
512 C. A. Silva et al.
D
o
w
n
l
o
a
d
e
d

b
y

[
N
a
t
i
o
n
a
l

I
n
s
t
i
t
u
t
e

o
f

T
e
c
h
n
o
l
o
g
y

]

a
t

2
2
:
0
1

2
0

A
u
g
u
s
t

2
0
1
1