DEVELOPMENT OF A MULTI-AGENT PLATFORM
FOR SUPPLY CHAIN-WIDE ORDER FULFILMENT
Roberto Domínguez Cañizares and Jose M. Framiñán
Industrial Management, School of Engineering, University of Seville,
Camino de los Descubrimientos, s/n, Isla de la Cartuja, 41092, Sevilla, Spain
Keywords: Supply Chain Management, Order Fulfilment Process, Multi-Agent System, Simulation.
Abstract: In this paper we describe an agent-based framework for modelling and simulating different processes taking
place in supply networks, resulting in a supply chain simulator called SCOPE (Sistemas COoperativos para
la Programación y Ejecución de pedidos). The framework is composed of reusable elements (agents and
objects) allowing easy modelling of real-scale supply chains, with different companies and products. Each
company in the model can use different policies and parameters for the different business functions. The
framework is implemented using Swarm. Furthermore, its generic and modular structure allows to easily
adding new and more complex functions for the agents. The final aim of SCOPE is to serve as a testbed to
implement and analyse the effects of different management decisions related to order fulfilment over real-
scale supply chains. SCOPE has been validated using different supply chains described in the literature.
1 INTRODUCTION
In today's dynamic and complex manufacturing
environment, an important enabler in gaining
competitiveness is the ability of a company to
respond quickly and effectively to satisfy customers
(Framinan, 2009) via an effective Order Fulfilment
Process (OFP), which starts with receiving orders
from the customers and ends with having the
finished goods delivered (Lin et al., 1998).
OFP is a complex process because it is usually
composed of several activities, executed by different
functional entities heavily interdependent among the
tasks, resources, and agents involved in the process
(Lin and Shaw, 1998). Furthermore, manufacturing
practice is shifting towards the outsourcing
paradigm, so OFP activities may take place across
different companies, thus hindering the centralized
planning and scheduling (Lin and Lin, 2006).
Therefore, OFP is likely to be executed over a
Supply Chain Network (SCN).
Available-To-Promise (ATP) systems refer to a
variety of methods and tools to enhance the
responsiveness of order promising and the reliability
of order fulfillment (Framinan, 2009). It is difficult
to predict the effects of new management policies
and strategies to improve OFP on a real SCN, but at
the same time this may be of great benefit for
companies. One of the most popular approaches is
using analytic models, like linear programming,
integer/mixed integer programming to analyze the
performance of a SCN. However, a SCN involves
stochasticity and uncertainty features not suited for
these analytical models. In addition, analytical
models may not be solved due to their complexity
and nonlinearity. Simulation, especially the multi-
agent-based distributed simulation, turns out to be
one of the most effective tools to model and analyze
SCNs (Long et al., 2011).
Because of its ease for modeling and
comprehensive description of complex systems
(particularly its capability of handling their
dynamics and stochastic behaviour), simulation has
been widely used in SCN management. There is a
great interest in modeling SCNs as Agent-Based
Systems (ABS) because there is a natural
correspondence between SCN participants and
agents in a simulation model. In addition, SCNs tend
to be decentralized systems with the participants
acting independently, according to their own
interests and policies (Long et al., 2011). Thus, the
use of an agent-based approach is suitable to model
and simulate SCNs (Chatfield et al., 2007).
The rest of this paper is summarized as follows:
Section 2 describes the framework, Section 3
describes the implementation of the framework,
203
Domínguez Cañizares R. and M. Framiñán J..
DEVELOPMENT OF A MULTI-AGENT PLATFORM FOR SUPPLY CHAIN-WIDE ORDER FULFILMENT.
DOI: 10.5220/0003747802030208
In Proceedings of the 4th International Conference on Agents and Artificial Intelligence (ICAART-2012), pages 203-208
ISBN: 978-989-8425-96-6
Copyright
c
 2012 SCITEPRESS (Science and Technology Publications, Lda.)

Section 4 is validation of the simulation platform
obtained (SCOPE) and Section 5 is the conclusion.
2 FRAMEWORK
Real SCNs have multiple layers of abstraction (Lin
et al., 2002), as they can be studied in different
levels of details. Thus, we model two different
layers: an Enterprise Layer containing all enterprises
in the SCN, and a Functional Layer, including the
main functions/departments of the enterprises. This
structure will allow studying inter-enterprises
relationships and intra-enterprises relationships.
The Enterprise Layer is modelled by one generic
and reusable agent (Enterprise Agent) composed of
several functional agents modelling physical and
planning tasks. By doing this, every department in
the enterprise is encapsulated in one agent, with its
characteristics of independency and autonomy, and
being able to take its own decisions.
A central feature of ABS is the bottom-up
methodology to construct a model. In this
methodology the user assumes that he/she cannot
understand the whole phenomenon of interest but
can observe, at a micro level, specific activities and
processes, and tries to understand their behaviour
and their objectives. These agents interact and
communicate with other agents and they join to form
a coherent whole on a macro level (Nilsson and
Darley, 2006). This whole is the emergent
behaviour, which cannot be predicted in advance.
In accordance with the bottom-up methodology,
SCN and the Enterprise Agent are not explicitly
modelled. Instead, the Enterprise Agent behaviour
emerges from its components’ behaviours (i.e.,
functional agents), which are easier to understand
and model. Similarly, the global SCN behaviour
emerges from that of its components enterprises.
2.1 Enterprise Agent
The Enterprise Agent is able to model any kind of
company in the SCN. The composition of their
functional agents determines its behaviour. To
simplify the Enterprise Agent configuration we
identify four roles of the companies in the SCN.
Companies with similar functions (in terms of
functional agents) belong to the same category.
These four role categories and their main
characteristics are summarized in the Table 1.
To model a new company, it is enough to select
one of the roles from Table 1 for the Enterprise
Agent and automatically the required functional
agents are assigned to it. The basic agents always
belong to the Enterprise Agent for the selected role,
while the optional agents are selected depending on
the enterprise characteristics.
2.2 Functional Agents
A good design of the functional agents is crucial
because they must capture the internal dynamic of a
real enterprise and should model the key aspects of
enterprise management. We base our choice in the
level 1 of the SCOR model (SCC, 2006) and the
literature revised.
Level 1 of the SCOR model summarises business
operations in five main activities: Plan, Source,
Make, Deliver and Return. Activities are divided in
two groups: Physical activities (Source, Make,
Deliver and Return) to manage the physical
resources of the enterprise, and planning activities
(Plan) to make decisions. Each one of the physical
activities (with the exception of Return) is modelled
by one agent, so there are three physical agents:
Table 1: Roles of the Enterprise Agent.
Roles in
Framework
Description Examples types Basic Agents Optional Agents
Manufacturer
The enterprise is in the SCN and
has manufacturing or assembly
capacity
Manufacturer,
Assembler
Demand Fulfillment, MRP,
Scheduling, Source, Make,
Deliver
Demand Forecast,
Master Planning,
Production Planning
Intermediate
The enterprise is in the SCN and
has not manufacturing or
assembly capacity
Distributor,
Wholesaler,
Retailer
Demand Fulfillment, MRP,
Deliver
Demand Forecast
External
Provider
Any enterprise who provide
something to the SCN, but it’s
out of the SCN
Provider
Demand Fulfillment,
Deliver
No
External
Customer
Any enterprise who demand
something to the SCN, but it’s
out of the SCN
Customer MRP, Source No
ICAART 2012 - International Conference on Agents and Artificial Intelligence
204

Figure 1: Multi-Agent Framework: general configuration of the Enterprise Agent.
Source Agent, Make Agent and Deliver Agent. The
Return activity is implicitly implemented by
allowing the Deliver Agent to return products or
receive returned products. According to the Supply
Chain Planning Matrix in Stadtler (2005), the Plan
activity has been divided into six planning functions,
being each of these functions carried out by a
different agent: Demand Fulfilment Agent, Demand
Forecast Agent, Master Planning Agent, Production
Planning Agent, MRP (Material Resource Planning)
Agent and Scheduling Agent.
The enterprise is then modelled as a mix of
planning agents and physical agents. Planning agents
store management policies and take the main
decisions. Physical agents control the physical
resources of the enterprise and share information
with planning agents. The overall configuration of
the Enterprise Agent, with all its functional agents is
shown in Figure 1. A brief description of the agents
is presented below:
 Source Agent, to handle the arrival and
storage of raw materials, and its delivery to
the manufacturing process when needed.
 Make Agent, to monitor the manufacturing
process by controlling machines and the flow
of jobs in the shop floor.
 Deliver Agent, to handle the arrival and
storage of finished products, and the
delivering of orders to customers.
 Demand Fulfillment Agent, in charge of
demand management and inventory control. It
checks incoming RFQs (Request for
Quotations) from customers and quotes the
due dates. If the order is accepted, it tries to
fulfill them from inventory, if available. If
inventory is not enough, it sends a production
order (a Job) to the Scheduling Agent if the
enterprise is a manufacturer. Otherwise it
sends a purchase order to the MRP Agent.
 Demand Forecast Agent: This agent requests
the enterprise demand historical data to the
Demand Fulfillment Agent in every forecast
period. Then it forecasts the demand for each
product in the next periods using a forecasting
rule (like Simple Moving Average, etc.).
 Master Planning Agent: In every planning
period, this agent uses forecast information
from the Demand Forecast Agent and
generates an aggregate Master Plan for the
products concepts defined by the company by
solving a linear programming model (that
includes capacity and inventory restrictions, as
well as production and inventory holding
costs), obtaining the production needs per
product concept for each period.
 Production Planning Agent: It receives the
Master Plan and takes only the production
needs for the first planning period. Then it
generates a detailed Production Plan solving
one model for each product concept and
obtaining the production needs for the final
products to accomplish the Master Plan.
 MRP Agent: It creates a detailed material plan
to fulfil the Master Plan. If there is no Master
Plan, it takes control of the raw material
inventory levels by using some inventory
policy. This agent is the one in charge of the
purchasing functionality in the enterprise.
 Scheduling Agent: This agent schedules
production orders (Jobs) coming from the
Production Planning/Demand Fulfillment
DEVELOPMENT OF A MULTI-AGENT PLATFORM FOR SUPPLY CHAIN-WIDE ORDER FULFILMENT
205

Agent by using first some priority rule to
create an initial solution, and then a heuristic
to improve the initial solution according to
certain objective. It also calculates starting and
ending times for each job (so it can help the
Demand Fulfillment Agent for due date
calculation), generating a detailed schedule.
According to this schedule, at the starting time
of each job it sends the job information to the
Make Agent to start its production.
3 IMPLEMENTATION
SCOPE has been implemented in Swarm (Java
version) and NetBeans IDE 6.7 as implementation
framework. Swarm is a multi-agent software
platform for the simulation of complex adaptive
systems (Minar et al., 1996). It provides object
oriented libraries of reusable components for
building models and analyzing, displaying, and
controlling experiments on those models. Swarm is
suitable for modeling SCN (Lin et al., 1998). In Lin
et al., (2002) we can see a comparison between
SCN’s features and Swarm. Furthermore, models are
coded in Java, so low-level functions can be added.
We have considered other platforms for the
implementation of our framework, such as NetLogo
or Repast. NetLogo stands out for its ease of use, but
it is basically designed for mobile agents acting
concurrently on a grid space with behaviour
dominated by local interactions over short times, and
that are not extremely complex (Railsback et al.,
2006). Repast was initially conceived for
implementing Swarm in Java. However, it does not
implement swarms, which are very helpful for
organizing models. Furthermore the schedule
executes top-level actions in randomized order
(which is not desirable), while Swarm allows a
precise control of the sequence of actions.
In our model we have four different types of
objects, nine agents and three swarms. Objects are
information containers created by the agents, who
use them and send them to other agents. A typical
object is represented by a java class, where all
variables and methods are defined. The objects
available are: RFQ, Job, Product and Machine.
Agents are the basic elements in the simulation
model. They represent the main functions in the
enterprise (functional agents in the framework).
Their behavior is modeled by writing methods.
Information is passed by arguments, which makes
easy to add new capabilities to the agents, by simply
adding new methods, or overwriting existing ones. A
typical agent in our model is represented by a java
class that extends the Agent class in Swarm. A
pseudo code example is shown below:
Public class AgentName extends
SwarmObjectImpl {
//Internal Variables
Private Type name1;
Private Type name2;
...
//Constructor
Public AgentName (Type name1, Type
name2,...) {
this.name1=name1;
this.name2=name2;
...
}
/*Methods: define behaviours and
abilities of the agent*/
Public Type behaviour1 (Type name3) {...}
Public void behaviour2 () {...}
...
}
Swarms are agent’s containers with schedules of
actions. The three swarms in the model are the
following:
 Enterprise: this is the main swarm, and it
serves to model an entire enterprise. It is
formed by a combination of the nine types of
agents described before and contains the
schedules for all these agents, controlling their
actions. The behaviour of the enterprise
swarm is defined by the emergent phenomena
of the agents inside themselves.
 Model: it models the SCN environment, and it
contains all the enterprises.
 Observer: it contains the model swarm and
special methods to present all relevant
information from the agents in the model
swarm, for a post-simulation analysis.
For solving the planning models included in the
Master Planning and Production Planning agents,
SCOPE can be connected with Gurobi solver
through a special library for Java. Gurobi is a
commercial software package for solving large-scale
mixed-integer linear optimization problems.
4 VALIDATION
In order to validate SCOPE we have looked into the
literature for SCNs modeled and simulated by other
authors and have compared their results with those
provided by our platform. Particularly, in Chatfield
et al. (2004) their Simulator for Integrated Supply
Chain Operations (SISCO) is validated by
comparing their results with those by Chen et al.
(2000) and Dejonckheere et al. (2003). Therefore,
ICAART 2012 - International Conference on Agents and Artificial Intelligence
206

we validate SCOPE using the same two scenarios as
in Chatfield et al. (2004), and simulate a third
scenario reproducing some of the experiments
conducted in Chatfield et al. (2004) after the
validation of SISCO.
The SCN’s structure is the same for the first two
scenarios: Only one product, and SCN structured by
a serial formation of customer, retailer, wholesaler,
distributor, and factory levels. The lower node
places orders with the next upper node and this node
fills these orders. The customer does not fill orders
and the factory places orders with an outside
supplier. There are other features that remain
identical for both scenarios, like the use of a fixed
lead time, or the forecast rule and inventory policy.
For more details see Chatfield et al. (2004).
4.1 Scenario 1: Chen et al. (2000)
Chen et al. calculated a statistical lower bound for
the variance amplification in the SCN described
above, and obtained the result shown in equation (1).
(

)
()
≥1+
2


+
2







,∀
(1)
We have obtained very similar results to those
offered by SISCO, although we found a greater
amplification at upper SCN stages than in Chen et
al. (2000). Chatfield et al. justify these results
arguing that the bounds provided by Chen et al. do
not account for interactions and interdependencies
present in a multi-stage system. To test this, they
perform a “sequential pairs execution” simulation, in
which they broke the supply chain into four two-
node sub-chains (customer-retailer, retailer-
wholesaler, wholesaler - distributor, distributor -
factory). Then, they simulate each sub-chain using
the ordering mean and standard deviation obtained
from the simulation of the previous sub-chain. We
reproduced this new scenario, obtaining the results
in Table 2, which are extremely close to those
predicted by Chen et al., and SISCO.
Table 2: Amplification Ratio for Chen et al. (2000),
“sequential pairs execution” experiment. Parameters:
demand rate =N (50,20
2
); protection time = L+R = 4+1
= 5; MA(15) forecasting; simulation time = 5200 time
periods (200 for warm-up).
Enterprise Chen et al. SISCO SCOPE
Retailer
1.89
1.90
1.90
Wholesaler
3.57 3.59 3.53
Distributor 6.74
6.70 6.66
Factory 12.73
12.84 12.58
4.2 Scenario 2: Dejonckheere et al.
(2003)
These authors used a Control Engineering
methodology to study the variance amplification in
the SCN described before, and obtained the result
shown in equation (2). The results for this scenario
are summarized in Table 3, where it is possible to
notice that SCOPE performs very similar to SISCO
and to Dejonckheere et al. (2003).


=
−2−+2

+

+





(2)
Table 3: Amplification Ratio for Dejonckheere et al.
(2003). Parameters: demand rate =
N (100,10
2
);
protection time = L+R = 3+2 = 5; MA(19) forecasting;
simulation time = 5200 time periods (200 for warm-up).
Enterprise Dejonckheere
et al.
SISCO SCOPE
Retailer
1.67
1.67
1.71
Wholesaler
2.99 2.99 3.10
Distributor 5.72
5.72 5.96
Factory 11.43
11.43 11.93
4.3 Scenario 3: Chatfield et al. (2004)
After the validation of SISCO these authors studied
the impact of information sharing and different
degrees of information quality on the Bullwhip
Effect when the lead time between companies is
stochastic. The SCN used is similar to the one used
in the previous scenarios, with only some
differences like the mentioned stochastic lead time
or the inventory policy used (see Chatfield et al.
2004 for more details). We have used SCOPE to
reproduce two of these experiments. The first
experiment analyzes the influence of different
information quality levels on the Bullwhip Effect for
a lead time c.v. = 0.50. The results obtained by
SCOPE (Figure 2) are identical to the results
obtained by SISCO in Chatfield et al. (2004). A
second experiment analyzes the impact of the
variance of the lead time on the standard deviation
of orders for a given information quality level.
Again the results obtained by SCOPE were identical
to those by SISCO. Therefore we can conclude that
the validation of SCOPE is successful.
DEVELOPMENT OF A MULTI-AGENT PLATFORM FOR SUPPLY CHAIN-WIDE ORDER FULFILMENT
207

Figure 2: SCOPE results for Scenario 3.
5 CONCLUSIONS
We have developed SCOPE, a framework with a
modular design for SCN simulation and analysis.
This tool may help SCN managers and researchers
to better understand how a given SCN configuration
performs in the presence of external and/or internal
disturbances. SCOPE is completely open for
improvement, which can be done in several ways:
 New functions, i.e. more policies, planning
models, heuristics methods, priority rules,
forecast methods, etc.
 New behaviours in order to have agents more
proactive and with negotiation abilities.
 Including a User Interface.
Future work includes the following lines:
 Studying the Bullwhip Effect and optimizing
inventory policies when there are multiple
Providers, each one providing different raw
materials with different stochastic lead times.
 Testing up-to-date heuristic rules for
scheduling. Simulation of full-scale SCNs and
analysis of the impact that these have on lead
times and customer satisfaction when they are
implemented at different levels of the SCN.
 Studying different policies for purchase
selection, giving to the agents the ability of
selecting the best offer in each purchase and
analyzing their individual and global benefits.
REFERENCES
Chatfield, D., Kim, J., Harrison, T., Hayya, J., 2004. The
Bullwhip Effect—Impact of Stochastic Lead Time, In-
formation Quality, and Information Sharing: A Simu-
lation Study. Production and Operations Management,
Vol. 13(4), pp. 340-353.
Chatfield, D., Hayya, J., Harrison, T., 2007. A multi-for-
malism architecture for agent-based, order-centric
supply chain simulation. Simulation Modelling Pra-
ctice and Theory, Vol. 15(2), pp. 153-174.
Chen, F., Drezner, Z., Ryan, J., Simchi-Levi, D., 2000.
Quantifying the bullwhip effect in a simple supply
chain: the impact of forecasting, lead times, and infor-
mation. Management Science, Vol. 46(3), pp. 436-
443.
Dejonckheere, J., Disney, S., Lambrecht, M., Towill, D.,
2003. The impact of information enrichment on the
Bullwhip effect in supply chains: A control enginee-
ring perspective. European Journal of Operational Re-
search, Vol. 153, pp. 727-750.
Framinan, J. M, 2009. Managing resources for order pro-
mising in Available-To-Promise (ATP) systems: A
simulation study. International Conference on Indus-
trial Engineering and Systems Management.
Lin, F.-R., Tan, G. W., Shaw, M. J., 1998. Modeling
Supply-Chain Networks by a Multi-Agent System. Pro-
ceedings of the Hawaii International Conference on
System Sciences, Vol. 5, pp. 105-114.
Lin, F.-R., Shaw, M., 1998. Reengineering the Order Ful-
fillment Process in Supply Chain Networks. Interna-
tional Journal of Flexible Manufacturing Systems,
Vol. 10 (3), pp. 197-229.
Lin, F.-R., Huang, S.-H., Lin, S.-C, 2002. Effects of Infor-
mation Sharing on Supply Chain Performance in Ele-
ctronic Commerce. IEEE Transactions on Enginee-
ring Management, Vol. 49 (3), pp. 258-268.
Lin, F.-r., Lin, Y., 2006. Integrating multi-agent nego-
tiation to resolve constraints in fulfilling supply chain
orders. Electronic Commerce Research and Appli-
cations, Vol. 5(4), pp. 313-322.
Long, Q., Lin, J., Sun, Z., 2011. Modeling and distributed
simulation of supply chain with a multi-agent plat-
form. International Journal of Advanced Manufactu-
ring Technology, pp. 1-12.
Minar, N., Burkhart, R., Langton, C., Askenazi, M., 1996.
The Swarm simulation system: A toolkit for building
multi-agent simulations. Working Paper 96-06-042,
Santa Fe Institute, Santa Fe.
Nilsson, F., Darley, V., 2006. On complex adaptive sys-
tems and agent-based modelling for improving deci-
sion-making in manufacturing and logistics settings:
Experiences from a packaging company. International
Journal of Operations and Production Management,
Vol. 26 (12), pp. 1351-1373.
Railsback, S., Lytinen, S., Jackson, S., 2006. Agent-based
Simulation Platforms: Review and Development Reco-
mmendations. Simulation, Vol. 82(9), pp. 609-623.
SCC, 2006. SCOR v8.0. Supply Chain Council, Inc,
Washington.
Stadtler, H., 2005. Supply chain management and advan-
ced planning - Basics, overview and challenges. Euro-
pean Journal of Operational Research, Vol. 163 (3),
pp. 575-588.
ICAART 2012 - International Conference on Agents and Artificial Intelligence
208