Computers and Chemical Engineering 26 (2002) 1755 /1769

www.elsevier.com/locate/compchemeng

Agent-based supply chain management* 1: framework /

Nirupam Julka 1, Rajagopalan Srinivasan *, I. Karimi

Department of Chemical and Environmental Engineering, National University of Singapore, 4 Engineering Drive, Singapore 117576, Singapore

Received 26 July 2001; received in revised form 12 July 2002; accepted 12 July 2002

Abstract

In the face of highly competitive markets and constant pressure to reduce lead times, enterprises today consider supply chain
management to be the key area where improvements can significantly impact the bottom line. More enterprises now consider the
entire supply chain structure while taking business decisions. They try to identify and manage all critical relationships both upstream
and downstream in their supply chains. Some impediments to this are that the necessary information usually resides across a
multitude of resources, is ever changing, and is present in multiple formats. Most supply chain decision support systems (DSSs) are
specific to an enterprise and its supply chain, and cannot be easily modified to assist other similar enterprises and industries. In this
two-part paper, we propose a unified framework for modeling, monitoring and management of supply chains. The first part of the
paper describes the framework while the second part illustrates its application to a refinery supply chain. The framework integrates
the various elements of the supply chain such as enterprises, their production processes, the associated business data and knowledge
and represents them in a unified, intelligent and object-oriented fashion. Supply chain elements are classified as entities, flows and
relationships. Software agents are used to emulate the entities i.e. various enterprises and their internal departments. Flows */
material and information */are modeled as objects. The framework helps to analyze the business policies with respect to different
situations arising in the supply chain. We illustrate the framework by means of two case studies. A DSS for petrochemical cluster
management is described together with a prototype DSS for crude procurement in a refinery.
# 2002 Elsevier Science Ltd. All rights reserved.

Keywords: Supply chain modeling; Software agents; Grafcets

Nomenclature
3PL third party logistics
ADE agent development environment
DSS decision support system
EC extra clustorial
iCOP intelligent cluster optimizer
KPI key performance index
PRISMS petroleum refinery supply chain modeler and simulator
RFQ request-for-quote
RRFQ reply-for-request-for-quote
SCM supply chain management
SCN supply chain network
SFC sequential function charts

* Corresponding author. Tel.: /65-687-48041; fax: /65-6779-91936

E-mail address: [email protected] (R. Srinivasan).
1
Present address: Singapore Institute of Manufacturing Technology, Singapore.

0098-1354/02/$ - see front matter # 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 9 8 - 1 3 5 4 ( 0 2 ) 0 0 1 5 0 - 3
1756 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769

1. Introduction consider the interactions with their suppliers and

customers and incorporate them into their decision-
The management of relationships across a supply making process. They also need to reformulate their
chain to capture the synergy of intra- and inter-company business policies to enable them to incorporate the
business processes is referred to as supply chain manage- information regarding their supply chain into their
ment (SCM) (Lambert & Cooper, 2000). Historically, decisions. Most companies are now paying close atten-
companies have focused only on their resources, con- tion to SCM. This is largely due to the success stories of
straints, and policies to make decisions and reduce costs. major corporations around the world saving billions of
With intense competition and reducing profit margins, dollars in inventory and logistics costs by efficient
this approach is no longer sufficient. They need to management of their supply chains. SCM has risen to

Fig. 1. (a) A chemical supply chain and (b) the internal departments of a typical manufacturer.
 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769 1757

Fig. 2. Grafcet of an enterprise agent.

be a key strategic area that has direct impact on the This paper addresses this critical need by presenting
success of any enterprise in today’s highly competitive an agent-based framework for supply chain analysis.
business environment. Software agents exhibit collaboration, intelligence, and
There are several challenges in effective supply chain mobility, and thus are ideal for modeling and analysis of
decision-making. The first challenge is that the informa- supply chains. We propose an agent-based framework
tion across all the departments and enterprises is that can emulate a supply chain at different levels of
distributed, dynamic, and disparate in nature. Secondly, granularity */cluster, inter-enterprise, and inter-depart-
in a present-day enterprise, decision centers reside in ment. We begin with a discussion of software agents and
different departments. For instance, consider the refin- their application to different chemical engineering
ery supply chain shown in Fig. 1. The enterprises and problems, as well as SCM related problems in Section
their internal departments take a number of decisions 2. Previous work related to SCM is reviewed in Section
regarding information and material flows across the 3. Then, we present our agent-based framework in
supply chain. These decisions are usually based on the Section 4 and discuss its features in Section 5. In Section
department’s own constraints, and the upper manage- 6, we illustrate a decision support system (DSS) for the
ment handles the conflicts. Most often, the decisions are management of a petrochemical cluster. In the second
taken independent of the interdependency of various part of this paper, we describe in detail another
factors contributing to the overall business process of application for managing the crude oil procurement
the enterprise. The decisions are optimized locally within process in a refinery.
the departments but do not assure a global optimum for
the enterprise. Of course, decision support tools exist for
local decision-making, e.g. planning and scheduling
systems, inventory management systems, market trading 2. Software agents
optimization systems, etc. A mere electronic integration
of these tools would not solve the problem and there is a Software agents can be defined in different ways
need for a unified approach for modeling and analysis of depending on the way they are implemented and the
supply chains, which explicitly captures the interactions tasks they perform. Wooldridge and Jennings (1995)
among enterprises and within the departments of an suggest that any computer system (software or hard-
enterprise. Such an approach would enable integrated ware) should have the following properties to be termed
supply chain decision-making. as an agent:
1758 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769

(a) Autonomy. It should have some control over its is used to define the activity. Alternatively, a program
actions and should work without human interven- can also define the activity. AdeGrafcet is based on
tion. Grafcet or Sequential Function Charts (SFC), a graphi-
(b) Social ability. It should be able to communicate cal language that has been accepted as an industrial
with other agents and/or with human operators. standard (IEC 848 and IEC 1131-3) for PLC-level
(c) Reactivity. It should be able to react to changes sequential logic control (David & Alla, 1992). The
in its environment. Grafcet chart is made of nodes and links (see Fig. 2).
(d) Pro-activeness. It should also be able to take Nodes represent states of a system. The main nodes are
initiative based on pre-specified goals. transitions and steps and the main links are branches
and joins. An AdeGrafcet step may contain small
The above-mentioned properties are generic for an programs, unlike standard Grafcets, where the opera-
agent. An agent may exhibit more of one property than tions in the steps are Boolean in nature. A transition
another based on its architecture and embedded intelli- marks the change of the system from one state to
gence. another. It represents an event or condition, which is
Toolkits are commonly used to develop multi-agent necessary to change the state of the system. Links
systems. In our work, we used Agent Development connect two nodes in a Grafcet. Examples of links
Environment (ADE), which is built on G2, Gensym’s include First-True Branch and Asynchronous Join.
expert system shell. ADE has the following main
components: agent, message, activity, host and environ- 2.4. Host
ment. These components are explained in detail below.
There is a host for every multi-agent application.
2.1. Agent Each agent, when created, is assigned to a particular
host. The most important task of the host is to provide
An ADE agent is an autonomous, multi-threaded message delivery service. All messages exchanged be-
object, which communicates with other agents through tween agents are routed through the host. In this, the
messages. Every agent in ADE has a unique name, host serves the function of the post office.
which serves as its address. This enables an agent to
communicate with other agents irrespective of their 2.5. Environment
location across a computer network. ADE provides a
facility similar to ‘‘Yellow pages’’ to find the addresses An environment is a group of agents. The agents may
of agents with some particular pre-defined property. be registered to different hosts and can change hosts but
Agents can send messages to other agents based on these they cannot change their environment. Agents within an
properties. The task of an agent is described in the form environment can be disallowed from communicating
of one or more activity classes (explained below), which with agents external to their environment.
is embedded inside the agent. At any particular time, an The above architecture permits agents to perform
agent may perform multiple tasks, each of which is parallel tasks just as a human would do. Addition of
represented as an instance of a specific activity. more activities or threads to a particular activity will
also add new functionalities to an agent. Further, the
2.2. Message same agent class may be used inside different applica-
tions. Tasks performed at each step may be as simple as
The base level message class provided by ADE is querying a database or as complex as collaborating with
called AdeMessage. All agents communicate with each other agents to jointly solve an optimization problem.
other by sending objects of AdeMessage class or its Agents have been used to accomplish a variety of
subclasses. Information is embedded inside these objects tasks. They have been used to assist process and
in the form of attributes. equipment design, perform enterprise integration stu-
dies, and address the decision-making processes in a
2.3. Activity business. But a question often asked is what makes a
task suitable for the use of multi-agent systems? In other
An activity defines a specific behavior of an agent. An words, what should be the characteristics of a problem
agent may perform multiple activities of the same or to be addressed using multi-agent systems? Aylett,
different types with multiple threads active in an Brazier, Jennings, Luck, Nwana, and Priest (1997)
activity. Once an activity is initiated, messages can be justify the application of multi-agent systems to a
sent specifically to it. An agent handler defines the problem that is inherently distributive in nature and
destination activity for each message. Thus, activities are requires the use of AI. However, they also mention that
initiated by the receipt of a particular message class, multi-agent systems are not required merely to produce
defined in the message handler of the agent. AdeGrafcet modularity, extra speed, reliability, flexibility or re-
 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769 1759

usability. We discuss multi-agent applications in chemi- comprises five smaller modules. These modules repre-
cal engineering as well as those in the field of SCM. sent and manage communication, service availability,
service execution, external agent service information,
2.6. Chemical engineering applications and situation assessment. Garcia-Flores, Wang, and
Goltz (2000) present an agent-based framework to
Most chemical engineering problems addressed by represent information flow in process industry supply
multi-agent systems are ones in which coordination of chains. The agents are based on the ADEPT agent
multiple entities is required. One such activity is the architecture. The entities in the supply chain perform the
design process. Han, Douglas, and Stephanopoulos tasks of either clients (placing orders) or servers (ful-
(1995) use an agent-based approach to develop a filling orders). Gjerdrum, Shah, and Papageorgiou
computer-aided design support system. The entire de- (2001) present a combination of numerical optimization
sign process is decomposed into tasks and an agent is with expert system techniques to address supply chain
assigned to each task. Central agents as well as a human modeling and performance assessment. They present a
operator do the coordination between these agents to multi-tier supply chain with customers, inbound and
assist in the design task. Maguire, Scott, Paterson, and outbound logistics, and production facilities. Their work
Struthers (1995) propose an agent-based modeling focuses on ordering and replenishment and does not
environment for process design. They introduce ‘‘auton- consider cross-department decision support within an
omous cognitive entities’’ which follow four simple steps organization.
to take any decision. These steps are assess, decide, act,
and review (ADAR approach). They use this modeling
environment to model a flash process in Maguire, 3. Previous work in supply chain analysis
Struthers, Scott, and Paterson, (1998). Yang and Yuan
(1999) present an agent-based framework for process Beamon (1998) provides a focused review of the
plant operations where agents model tasks in the plant literature on modeling multi-stage supply chains. Mod-
operations, e.g. monitoring, fault diagnosis, etc. Eo, eling approaches are classified as deterministic analyti-
Chang, Shin, and Yoon (2000) propose an agent-based cal (Cohen & Moon, 1990; Williams, 1983), stochastic
framework for the diagnosis of chemical processes. analytical (Cohen & Lee, 1998; Lee & Billington, 1993;
Agents take care of a set of process units and commu- Svoronos & Zipkin, 1991) and simulation (Towill, 1991;
nicate observations with each other. They use a knowl- Towill, Naim, & Wikner, 1992; Wikner, Towill, &
edge base pertaining to the unit to take decisions. Naim, 1991). Most of the works focus on specific
McGreavy, Wang, Lu, Zhang, and Yang (1996) problem areas of SCM, e.g. inventory management,
propose an agent architecture for concurrent engineer- distribution systems, demand forecasting, etc. These
ing. They define three main agents as process, control, solutions are thus inadequate to consider the complete
and equipment design agents. Each of these agents has supply chain structure and all the related knowledge and
subagents, e.g. the process design agent has reaction, information. Also, they do not utilize the Internet,
separation, energy integration, and utility subagents. which today is a very important channel for doing
The main agents exchange design data to perform business and sharing business related information in a
iterations in the design and negotiate for improvement seamless manner. To address these shortcomings, in the
in the complete design process. The agents inherently last few years, multi-agent approaches to supply chain
use application programs for the design and simulation problems have emerged. In this section, we summarize
of various sub-problems. The results from these pro- the literature in supply chain optimization and manage-
grams are then compiled and sent to the other agents for ment in process industry. We also discuss existing multi-
improvements. agent systems for SCM.
Batres, Lu, and Naka (1997), Batres, Asprey, Fu-
chino, and Naka (1999) address a concurrent process- 3.1. Chemical supply chain optimization and management
engineering problem where programs which perform
different design tasks are distributed over a local Chemical industry supply chains are typically long
network and reside in different computers. The agents and plant configurations are rigid. Also, chemical plants
deployed at different locations and using applications are normally their own customers. Thus, supply chain
written in different languages communicate with each solutions from other industries are not directly applic-
other using Knowledge Query and Manipulation Lan- able in the chemical industry.
guage (KQML). Struthers (1997) presents an applica- Perea-Lopez, Grossman, Ydstie, and Tahmassebi
tion of multi-agent systems to the task of organizing and (2001) present a framework for dynamically modeling
managing a Pressure Relief and Blow down study. The decentralized supply chains to assist in evaluating
agents are designed based on the Advanced Decision different heuristic control laws in their ability to dampen
Environment for Process Tasks (ADEPT). Each agent demand amplification in supply chains. They use a
1760 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769

polymer manufacturer’s supply chain as their case study. firm is similar to those between different firms. This
They consider decisions as the control variables of the helps in easy integration of the supply chain elements.
dynamic system and associate the optimization of the Agents also represent the resources of the firm. Each
decision-making process to designing a control law for a agent has a committed capacity profile, which shows its
dynamic system. The approach though novel in nature, time-varying load. In case a new task has to be
is specific to the demand amplification problem. An- scheduled, each agent bids for that task with a bid
other approach in managing chemical manufacturing that is inversely proportional to the committed capacity
supply chains under uncertainty and risk is presented by of the agent. Thus, resources are allocated to tasks
Applequist, Pekny, and Reklaitis (1999, 2000). without resorting to rescheduling the entire supply
Papageorgiou, Rotstein, and Shah (2001) address the chain.
issue of strategic supply chain design for pharmaceutical Swaminathan, Smith, and Sadeh (1998) present a
companies. They develop a mixed-integer linear pro- modeling and simulation framework for developing
gramming (MILP) model to select a product develop- customized decision support tools for supply chain
ment and introduction strategy and a capacity planning reengineering. Agents represent supply chain entities,
and investment strategy. Similarly, Bok, Grossman, and e.g. customers, manufacturers, and transportation.
Park (2000) have addressed a multi-site continuous These agents use different interaction protocols and
flexible process network problem. The objective is to help in simulation of material, information, and cash
maximize the profit of a SCN. The MILP model flows. These interaction protocols are in the form of
provides the sales, production and inventory profiles, messages of various classes. Message handlers are
and the production shortfalls for a scenario. Mathema- associated with each message class and consider the
tical models are capable of providing accurate results, agent receiving the message, to decide upon the message-
but cannot handle the computational complexity of the processing semantics. The agents use various control
entire SCM problem. These are again examples of policies to manage inventory, procure components, and
decision-support tools for specific problems. determine optimal transportation routes. The frame-
Voudouris (1996) addresses the fine chemical industry work helps in the configuration of a simulation model
supply chain problem. He emphasizes improving the by selection, instantiation, and composition of sets of
efficiency of the supply chain by improving the logistics components without the need for extensive program-
of the downstream formulation and packaging plants. ming expertise. These simulation models are then used
An MILP model is developed by introducing a number to analyze different reengineering options for a com-
of binary variables to represent discrete decisions pany.
pertaining to the supply chain. A crude oil supply In the next section, we describe our framework for
problem is addressed by Shah (1996). The scope of the supply chain analysis. We discuss how our framework is
problem is limited to the allocation of crudes to the particularly suited for the chemical industry SCNs and
jetties, tanks and crude distillation units (CDUs). A addresses some of the drawbacks of the existing
similar problem of distribution of crudes to distillation approaches.
columns is addressed by Kim, Choi, Kim, and Lee
(1999).
4. Framework for supply chain analysis
3.2. Supply chain management using multi-agent systems
A supply chain is defined as a network of suppliers,
SCM problems are both distributive in nature, and factories, warehouses, distribution centers, and retailers
require extensive intelligent decision-making. Thus, in through which raw materials are acquired, transformed,
the last few years, multi-agent systems have been a and delivered to customers (Fox, Barbuceanu, & Teigen,
preferred tool for solving supply chain problems. Good- 2000). In order to make supply chain decisions, it is
win, Keskinocak, Murthy, Wu, and Akkiraju (1999) necessary to identify the entities and flows in a supply
present a framework for providing decision support for chain. This calls for first modeling them to understand
an online exchange. They use a multi-agent system to the supply chain and classifying them as critical and
find matches of demand and supply on the exchange and non-critical entities (Lambert & Cooper, 2000). The next
provide the user with the best set of transactions. The step involves monitoring the critical elements. Finally,
user then chooses the best match based on his/her the identified and monitored elements need to be
discretion. Sauter, Parunak, and Goic (1999) present managed to improve the overall working of the supply
an architecture called Agent Network for Task Schedul- chain. This would eventually optimize the supply chain.
ing (ANTS), inspired by insect colonies and humans. We identify the elements in a supply chain, their
Agents represent elements in the supply chain and features, and the challenges associated with SCM. We
within a factory. Each firm in turn is viewed as a small classify the elements in a supply chain as entities and
supply chain, thus the interface between agents within a flows. Entities include all manufacturers, logistics pro-
 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769 1761

Fig. 3. Agent modeling of the supply chain entities.

viders, electronic exchanges and all their internal leave based on their own interest. This changes the
departments that participate in the business process. structure and flows in the supply chain. Information
These entities are essentially the operators in the supply in the supply chain e.g. prices, demands, technolo-
chain. Flows are of three types */material, information gies, etc. is also changing continuously.
and finance, and these are the operands in the supply 2) Distributed: The elements are distributed across
chain. These entities have three common features: various geographical locations. The planning and
operating systems used by an entity may also be
1) Dynamic: The supply chains are more flexible now.
geographically distributed e.g. there may be a
In today’s business environment, there are no
dedicated inventory database residing at each ware-
obligations for companies to be part of a supply
house of a manufacturer. The SCM related infor-
chain for a certain time period and they may join or
mation might even reside as rules-of-thumb with the
people responsible for performing the various tasks
2
We have currently not considered cash flows in our framework. in the business process.
1762 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769

Fig. 4. Snapshot of i COP displaying the agent hierarchy and attributes of agents and objects.

3) Disparate: The entities in a supply chain use processes can be emulated. Fig. 3 illustrates the model-
different systems built on different platforms for ing of the supply chain based on the framework.
planning and management of their business. Infor- Materials and information are modeled as objects.
mation pertaining to the various elements is also Each supply chain entity such as manufacturers and
disparate in form. For example, tracking of chemi- third party logistics (3PLs) are modeled as agents. The
cal shipment could be through emails, faxes, tele- internal departments of these entities are modeled as
phone calls or online reports. sub-agents. The material and information flows are
emulated by the exchange of the objects between these
Any attempt to model these elements and flows as a agents.2
mathematical formulation to yield optimized results The framework has two important elements. The first
would be cumbersome. Furthermore, this task would is the object modeling of supply chain flows such as
be specific to an enterprise and modification to suit the material and information. This modeling involves two
needs of other enterprises or authorities taking supply classes of objects, ‘commodity’ and ‘message’. These
chain related decisions would not be straightforward. objects are exchanged between agents to simulate the
Given this reality, there is a clear need for a new material and information flow in the SCN. The details
paradigm for DSSs, which is capable of gathering the of the material and information are stored as attributes
dispersed information required for the normal function of the respective object. Fig. 4 shows the attributes of an
of an enterprise and providing a structured way to make instance of a fulfilled order, a commodity object. The
decisions. details of the commodity object include name, amount
The foundation of our framework is a unified agent- of material, sender, receiver, time when it was dis-
based modeling of the entire supply chain network patched, time when it was received, and the present
(SCN) in an object-oriented fashion. This supply chain location. Fig. 4 also shows an instance of a reply-to-
model provides an environment where all business request-for-quote (RRFQ). This message object has the
 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769 1763

details of name of the material, amount of material, tively. The sales agent processes these bids and decides
status of order, source agent, destination agent and time on the outcome of the RFQ e.g. to be supplied from
of dispatch and arrival of message. stock, to be produced or cannot be fulfilled. The sales
The second important element of the framework is the agent replies to its enterprise agent, who then forwards
agent modeling. We use three classes of agents: emula- the message to the customer.
tion agent, query agent and project agent. The three
classes are explained in detail below. 4.1.3. Warehouse agent
This agent models the material-handling department
4.1. Emulation agents of an enterprise. It issues the required commodity to
other agents and replenishes raw materials stock by
Emulation agents model the supply chain entities. procuring from other enterprises. Thus the warehouse
These include the manufacturers, 3PLs, internal depart- agent also acts as the ‘‘sticky-end’’ attaching the
ments of an enterprise, etc. Each agent may have one or enterprise to a network of suppliers. Our framework
more sub-agents. For example, in Fig. 4, the PCS enables the incorporation of different inventory policies
enterprise agent that models the manufacturer has sub- such as periodic review, stock level maintenance, etc.
agents modeling the internal sales, warehouse, produc- into the warehouse agent.
tion, and line departments of the manufacturer. These
internal departments have generic functions in all 4.1.4. Production agent
enterprises, which can be customized to the practices This agent models the operations or production
of an enterprise. Some of the agents and their roles are department of an enterprise. It has sub-agents, which
explained below. model the various production lines or facilities of an
enterprise. This agent awards received orders to the
4.1.1. Enterprise agent production lines. On receiving a RFQ, it calls for bids
In our framework, an enterprise is defined as a plant from the line agents representing the production facil-
site producing a set of products from a set of raw ities. Based on these bids it generates a RRFQ. The
materials by utilizing a set of technologies. Enterprise production agent also procures material from the ware-
agent can have sub-agents that perform tasks internal to house and allocates them to the lines based on their
an enterprise. This agent handles all communication needs.
with the enterprise. On receiving messages from other
enterprises, this agent decides the further course of 4.1.5. Line agent
action. For example, if a request-for-quote (RFQ) is This agent models the physical setup converting raw
received, it forwards the message to its sales agent. materials into finished products. Each agent has a
Similarly, if a message requesting inventory information schedule of tasks to be performed. Whenever the
is received, it creates an internal request for this production agent calls for bids, the line agents capable
information and sends it to the warehouse agent. The of making that product estimate the cost to produce the
enterprise agent also holds information on its products requested amount of product. This cost takes into
and the technologies available for use by its line agents. account the availability of the line as well as the
technology used. Each line may use a different technol-
4.1.2. Sales agent ogy to convert raw materials to finished products. The
This agent models the sales department of an technologies available are defined under the enterprise
enterprise. It is responsible for processing of RFQs agent’s technology library. Different scheduling algo-
and subsequent orders received by the enterprise. rithms can be incorporated into these agents.
Customer information resides only with the sales agent. The enterprise agent with its sub-agents is one
It keeps track of market demand and prepares demand example of an emulation agent. There is another
plans. For this, it takes into account long-term alliances category of emulation agents that can be used to model
with customers, earlier demand patterns, and demands non-manufacturing enterprises such as the Internet
for related products on the exchanges. It keeps a record exchanges of the present e-commerce era. An agent
of all orders received, fulfilled, or queued. In case of belonging to this category is the exchange agent. This
made-to-stock goods, the sales agent generates the agent models a real life online exchange. Each posting
order. Consider a situation where the enterprise agent on the exchange is modeled as an object and is an
receives a RFQ from a customer. It forwards the RFQ attribute of the exchange agent. The deletion of the
to its sales agent. The sales agent queries the warehouse object representing the posting reflects the removal of a
and the production agents regarding the order and its posting from the online exchange. Multiple instances of
specifications and if they can fulfill it. The warehouse the exchange agent represent different electronic ex-
and production agents bid for that order based on changes. Apart from including information on the
present stock and production line availability, respec- postings, the information associated with the online
1764 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769

exchange is also encapsulated in the exchange agent. solved using a project agent is the identification of new
Other atypical businesses can also be similarly modeled plants for a petrochemical cluster. Both of these are
as emulation agents. But our goal is not just modeling of discussed later in Section 6.
the supply chain. We also wish to analyze the supply Having presented the design and elements of the
chain and perform studies to get an insight into the framework, we now discuss the implementation of the
supply chain behavior to assist in design, management, framework and how it facilitates supply chain DSSs for
and operation of the supply chain. In our framework, the new e-commerce scenario.
the query agent assists in analysis of the supply chain
models.
5. Framework features
4.2. Query Agents
We first identify features needed in any supply chain
These agents do not emulate any supply chain entity. DSS. We then discuss how the proposed framework
They handle queries from the user and assist in supply enables DSSs to address problems at different levels.
chain analysis. Each enterprise has one or more query Finally, we present the different modes in which a DSS
agent. There are also query agents for clusters of based on the proposed framework could work. These
enterprises. The query agents are active only in the would jointly bring out the advantages and features of
analysis mode (described in Section 5). For example, a the framework.
user may wish to know the status of backorders across
an industrial cluster comprising of a number of manu-
facturing enterprises. The cluster level query agent 5.1. Needs of supply chain decision support systems
receives the request, processes it, and identifies the
information needed to reply to this request. It then Fox, Barbuceanu, and Teigen (2000), Bui and Lee
contacts the query agents of the enterprises from which (1999) have identified the following issues for develop-
it needs the data. The query agents of the enterprise ment of DSS for SCM:
access the data of the agents under that enterprise and
. Distribution of supply chain activities among agents
reply to the cluster level query agent. In the above
. Coordination among DSS components
example, the enterprise level query agent contacts the
. Responsiveness to the modeling environment
sales agent and production agent of the enterprise and
. Interface with the present modules e.g. MRP systems
collects the data on backorders from them. This data is
. Maximum re-use of DSS components
processed, a reply generated and sent to the cluster level
query agent. The cluster level query agent compiles the With the new e-commerce scenario in mind (Julka,
data from all enterprise level query agents and displays Srinivasan, Karimi, Viswanadham, and Behl, 2000) we
it to the user. Sometimes studies are to be performed or propose the following features for the modern day
certain problems are to be solved using the supply chain supply chain DSSs:
model, which require more diverse steps than that
provided by the query agent. Such studies or problem 1) Knowledge encapsulation: Knowledge is an invalu-
solving methodologies can be modeled as project agents. able resource to an organization and the DSS
should be able to use all this knowledge. The
4.3. Project agents knowledge includes information on the structure
of various entities, their working, and their relations
The project agent models any new project related to with other entities. It also includes the material and
the SCN. A project is defined as a study or problem to information flow details. The DSS should have the
be solved in the SCN. A project agent performs the tasks knowledge organized in a manner which aids
needed to perform the study or solve the problem. The addition, deletion, modification and easy access. It
solution methodology includes querying various other should facilitate handling very specific queries.We
agents present in the framework either directly or use an object-oriented approach to encapsulate the
through the query agent, accessing databases to retrieve data and knowledge in a SCN. The working of the
information, creating or deleting material objects to different entities is captured using Grafcets as
simulate production and running optimization modules explained in Section 2.
for local decision making. The concept of a project 2) Intelligent inference: Intelligent inference is an
agent provides a way of adding analysis algorithms to important feature for efficient query handling and
the system in an easy manner without requiring ex- problem solving. Many intelligent inference engines
tensive modification to the system. An example of a are available for the development of DSSs. In our
study modeled as a project agent is the cumulative work, we have used G2, Gensym’s intelligent expert
inventory report for a petrochemical cluster. A problem system shell.
 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769 1765

3) Connectivity: The knowledge bases used by the enterprise, and intra-enterprise. The cluster level con-
system have to be constantly updated with the latest stitutes a number of enterprises in a business environ-
information. An example of this could be the ment forming a SCN e.g. a petrochemical cluster. The
current offers for various chemicals at the chemical inter-enterprise level deals with a particular company
exchanges on the World Wide Web (WWW). The and its supply chain e.g. a computer manufacturer
system should have the capability to interface with buying various components from its suppliers, assem-
the WWW, Wide Area Networks (WANs), and bling them and selling to its dealers. The intra-enterprise
available enterprise databases. Furthermore, it level views the supply chain activities within an enter-
should be able to intelligently mine for information prise e.g. the crude oil procurement process in a refinery.
from these sources. Our framework facilitates con- Each of the decision levels is explained in detail below.
nection to external sources of information by
deploying agents for the task. These agents pick 5.2.1. Cluster-level
up information from the source, convert it into the Carrie (2000) defines a cluster to be ‘‘a network of
required format, and feed it to the system for companies, their customers and suppliers of all the
incorporation into the decision-making process. relevant factors, including materials and components,
4) Flexibility: The system should be able to provide equipment, training, finance, and so on.’’ Decisions at
support in a variety of manners. It should be able to this level are usually taken by agencies responsible for
address queries regarding policy decisions, suppli- industrial and economic development of geographical
ers, information on inventory levels, pending orders, regions. For instance, an agency is responsible for
etc. It should help the user visualize the entire managing an industrial cluster. Investors planning
network from the view of a single enterprise as well projects in the cluster approach the agency with
as from the view of a planning authority managing a proposals. The agency would evaluate each proposal
large industrial cluster. It should not only allow based on the value-added to the cluster and the project’s
queries of ‘what is’, ‘how much’ and ‘when’ but also impact on space availability, synergies with existing
regarding ‘what if’.Apart from intelligent query plants, environmental and other constraints. For mak-
handling, our framework allows the user to add ing the best decision, the agency uses information on the
problem solving functionalities to the system using present product spectrum of the cluster, the demand and
project agents. It also permits performing simula- supply of each product and the imports and exports
tion of the business processes based on user-defined associated with the cluster. It then studies which projects
parameters and analysis based on user-defined fit well in the cluster and increase the synergy within the
performance metrics. cluster. With our framework, the user can include
5) Collaboration and scalability: The DSS should be enterprises on the fly and perform ‘‘what-if’’ studies
able to collaborate with legacy systems. This is an regarding total value addition of materials in related
essential feature as a large amount of knowledge is supply chains, effect on the logistics and other support
available in existing systems and rewriting them for industries, changes in the environment, and associated
a new DSS would require enormous amounts of issues. Another example for cluster-level analysis is an
time, effort and money, if not altogether infeasible. agency handling the logistics of a large industrial cluster.
The system should be scalable and modular in This agency needs information on various materials and
nature. This is to avoid a large number of changes modes of logistics used for their transportation, the total
in case two different systems need to be merged as bulk volume handled by each mode of logistics, etc. A
may occur in a situation where the same company DSS based on our framework can provide the agency
has multiple sites operating at different geographi- answers to these questions and thus help in taking better
cal locations but has a central planning facility. decisions.
Modularity should also assist incorporation of DSS
components developed elsewhere into the present 5.2.2. Inter-enterprise level
system. The decisions at this level are related to a particular
company and its supply chain. Choosing the right
Since our framework models each entity in a supply suppliers, estimating demands, setting distribution
chain as an agent, merging of multiple DSS into one is points, and developing new products, etc., are some
possible by just placing all the agents in the same issues that can be addressed at this level. Often the
environment. critical links in a company’s supply chain may not be the
immediately adjacent firms; instead the bottleneck may
5.2. Levels of decision support lie in the second or third tier of suppliers or customers.
An example, from the semiconductor industry illustrates
The framework helps in decision support of supply such a situation (Lambert & Cooper, 2000) where all
chain problems at three levels, namely */cluster, inter- tier-one suppliers purchase from the same tier-two
1766 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769

supplier. Thus, in case of shortages, the bottlenecks are technologies involved and entities to be considered. The
the tier-two suppliers. A revelation of such dependencies mode allows the user to configure very specific queries.
may help the manufacturer decide on collaborating with The assistant agents, that are active only in this mode,
the tier-two suppliers in decisions related to production handle these queries. The assistant agents reply to these
planning. With our framework, the user can get over- queries using different presentation methods. For ex-
views of various consumer /supplier relationships and ample, a query regarding the production path to get
discover such dependencies. from material A to material B at minimum cost shall be
displayed in the form of a serial supply chain of existing
5.2.3. Intra-enterprise level facilities. Each facility detail can then be accessed. A
The decisions at this level are mostly related to query regarding the inventory status of a particular
internal departments of an enterprise and their func- material over the entire supply chain would be displayed
tions. The decision domain may lie completely within in the form of a spreadsheet.
the enterprise or at the interfaces between enterprises.
The decisions on inventory policies, scheduling metho-
dology and prioritizing orders belong to this level. The
6. A decision support system based on the framework
agent architecture allows the modeling of a company
and its business practices and policies. The model can
We have used our framework to address two pro-
then be used to study a specific problem using the
blems related to chemical SCNs. The first pertains to
project agent. A simulation of material and information
decision support to manage an industrial cluster. A
flow in this model is performed. The user can view the
system, called intelligent Cluster Optimizer (i COP), has
results with the help of pre-defined metrics or by using
been customized to manage a petrochemical cluster. The
project agents to perform studies and ‘‘what-if’’ analysis.
second system, called Petroleum Refinery Integrated
Supply chain Modeler and Simulator (PRISMS), has
5.3. Modes of decision support
been developed to provide decision support of the crude
oil procurement process in a refinery. In this paper, we
A DSS based on the framework can work in two
describe i COP while PRISMS will be presented in Part
different modes */simulation and analysis. Both these
2.
modes complement each other. The former is a platform
Consider a petrochemical cluster similar to that in
to simulate supply chain operations while the latter
Rotterdam (The Netherlands), Houston (USA), Jubail
provides an environment to address queries and perform
(Saudi Arabia), or Jurong Island (Singapore). An
‘‘what-if’’ analysis. The modes are usually used one after
agency responsible for the overall strategic management
the other.
of the cluster would be required to make decisions
related to:
5.3.1. Simulation
In this mode, based on the parameters entered by the . Selecting projects from a group of proposals
user, the DSS recreates the scenario of how a supply . What are the effects of the new projects on the
chain system with those parameters would perform. The demand and supply patterns of materials in the
user is able to view different key performance indices cluster?
(KPIs) like inventory profiles, order fulfillment, etc., but . Will the new projects increase the overall efficiency
is not permitted to dynamically change any parameter of the cluster?
values. Only emulation agents are active in this mode. . Net value generation by the cluster and measures to
They simulate the processes of the supply chain under increase it
study. The simulation mode can be run based on a time . What are the total imports and exports for the
period or on the basis of occurrence of an event. An cluster?
example of an event may be a backorder of 50% of the . What is the net value addition?
capacity of a firm. Once the simulation time is over or . Enterprises with what product spectra will be able
the event has occurred the system can be used in the to contribute positively to this figure?
analysis mode to perform various studies. . Effect on support organizations and shared services
of the cluster.
5.3.2. Analysis
In this mode, the DSS addresses user queries about The agent-based framework described in Section 4
the simulation performed and allows him to configure can be used to build a DSS for such an agency. iCOP is
the system for another simulation based on the results of an example of such a DSS. The various chemical
the previous run. It helps the user evaluate the efficacy manufacturers in the cluster are modeled as enterprise
of policies and strategies. The user configures the agents. Companies in the cluster may import from or
scenario by specifying parameters like inventory levels, export to enterprises outside the cluster. These external
 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769 1767

Fig. 5. Mapping of the SCN and the material chain network in a cluster using i COP.

enterprises are modeled as ‘Extra Clustorial’ (EC) cluster. The steps followed by the project agent are as
agents. A scenario is configured by choosing the agents follows:
and customizing them. These scenarios are then ana-
lyzed using project agents. An example of a particular 1) Request the sub-agents of the cluster for all
study performed by the project agent is the compilation customer contracts with EC agents.
of the information of the total inventory in the cluster. 2) Prepare a list of all materials that are exported.
3) Find out available commercial technologies which
The various steps followed by the project agent in this
use the materials in the list as raw materials.
case are as follows:
4) Evaluate if technology specific constraints are
1) Contact the query agent of the cluster and request satisfied.
the complete list of all enterprises in the cluster. 5) Display the potential projects.
2) Request the query agent of the each enterprise for
The data needed for iCOP include product spectra of
information of their respective inventories.
the companies and their capacities, present business
3) Prepare a list of all materials and their quantities
links and supply chains, logistics information, and
(stored and in process).
4) Take input from the user regarding the format of export /import information. All this information is
display (decreasing amount, total value, according incorporated into the DSS based on the guidelines given
to the enterprise). in the framework. The methodology of solving some of
5) Display the information of the inventory of materi- the above problems is provided by the agency. These
als over the cluster. methodologies are converted into decision algorithms
and included as Grafcets. A prototype of the system has
Another example of a problem solved by the project been put to use to address a few of the problems defined
agent is the identification of potential projects for the above. The user specifies the information related to a
1768 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769

proposed project. iCOP is able to provide the user with References

the changes in the export and import amounts of
products over the whole cluster. It provides information Applequist, G. E., Pekny, J. F., & Reklaitis, G. V. (1999). Economic
risk management for design and planning of chemical manufactur-
regarding excess capacities at different enterprises of the
ing supply chains. In Proceedings of the fifth international con-
network. It is able to provide the user with the changes ference on foundations of computer-aided process design . c01
in the supply pattern of the product produced by the Applequist, G. E., Pekny, J. F. & Reklaitis, G. V. (2000). Risk and
project and corresponding demand present in the uncertainty in managing chemical manufacturing supply chains.
system. The above study can be repeated for different Computers and Chemical Engineering 24 (9), 2211.
Aylett, R., Brazier, F., Jennings, N., Luck, M., Nwana, H., & Priest,
products and the effects can be compared. The DSS can
C. (1997). Agent systems and applications. In Panel discussion
provide a map of the transformation path followed by a report. Second UK workshop on foundations of multi-agent applica-
material both with respect to the manufacturers and the tions (FoMAS) .
material conversion. Fig. 5 shows a mapping of the SCN Batres, R., Asprey, S. P., Fuchino, T. & Naka, Y. (1999). A KQML
and the crude oil material conversion in a cluster as multi-agent environment for concurrent process engineering.
Computers and Chemical Engineering 23 (Supplement), S653.
generated by i COP. By including specific decision- Batres, R., Lu, M. L. & Naka, Y. (1997). An agent-based environment
support algorithms and providing the required data, for operational design. Computers and Chemical Engineering 21
the system can be customized for any cluster manage- (Supplement), S71.
ment agency. Beamon, B. M. (1998). Supply chain design and analysis: Models and
methods. International Journal of Production Economics 55 (3),
281.
Bok, J. K., Grossman, I. E. & Park, S. (2000). Supply chain
optimization in continuous flexible process networks. Industrial
7. Conclusions and Engineering Chemistry Research 39 (5), 1279.
Bui, T. & Lee, J. (1999). An agent-based framework for building
decision support systems. Decision Support Systems 25 (3), 225.
SCM has become the key strategic area that has direct Carrie, A. S. (2000). From integrated enterprises to regional clusters:
impact over the success of any enterprise in today’s the changing basis of competition. Computers in Industry 42 (2 /3),
highly competitive business environment. We have 289.
represented the supply chain and related problems Cohen, M. A. & Lee, H. L. (1998). Strategic analysis of integrated
production-distribution systems: Models and methods. Operations
through a unified, flexible, and scalable framework. Research 36 (2), 216.
The framework has two basic elements: object modeling Cohen, M. A. & Moon, S. (1990). Impact of production scale
of supply chain flows and agent modeling of supply economies, manufacturing complexity, and transportation costs
chain entities. We use three classes of agents: emulation on supply chain facility networks. Journal of Manufacturing and
Operations Management 3 , 269.
agents, query agents and project agents. This allows
David, R. & Alla, H. (1992). Petri Nets and Grafcet: tools for modeling
modeling the entire supply chain and the problem- discrete events systems . Prentice Hall International.
solving methodologies required for decision-support. Eo, S. Y., Chang, T. S., Shin, D. & Yoon, E. S. (2000). Cooperative
DSSs based on the framework can work in simulation problem solving in diagnostic agents for chemical processes.
and analysis modes. The user can configure supply chain Computers and Chemical Engineering 24 (2 /7), 729.
Fox, M. S., Barbuceanu, M. & Teigen, R. (2000). Agent-oriented
scenarios and compare the effect of different policies
supply-chain management. The International Journal of Flexible
and decisions using various KPIs. The framework Manufacturing Systems 12 (2 /3), 165.
enables decision support at cluster, intra-enterprise, Garcia-Flores, R., Wang, X. Z. & Goltz, G. E. (2000). Agent-based
and inter-enterprise levels and enables better decision information flow for process industries’ supply chain modeling.
support. We also illustrated the application of the Computers and Chemical Engineering 24 , 1135.
Gjerdrum, J., Shah, N. & Papageorgiou, L. G. (2001). A combined
framework using i COP, a DSS for the management of
optimization and agent-based approach to supply chain modelling
a petrochemical cluster. The same framework can be and performance assessment. Production Planning and Control 12
applied to provide decision support at the intra-enter- (1), 81.
prise level as illustrated in Part 2 using a refinery crude Goodwin, R., Keskinocak, P., Murthy, S., Wu, F., & Akkiraju, R.
procurement process. (1999). Intelligent decision support for the e-supply chain. In
Proceedings of artificial intelligence for electronic commerce */
AAAI workshop 99 (p. 77).
Han, C., Douglas, J. M. & Stephanopoulos, G. (1995). Agent-based
approach to a design support system for the synthesis of
continuous chemical processes. Computers and Chemical Engineer-
Acknowledgements
ing 19 (Supplement), S63.
Julka, N., Srinivasan, R., Karimi, I., Viswanadham, N., & Behl A.
The authors would like to acknowledge financial (2000) Enabling Framework for Decision Support Systems in the e-
support provided by The Logistics Institute */Asia Commerce Era. Presented at AIChE Annual Meeting, 12 /17
Pacific, National University of Singapore. November, Los Angeles, CA, USA.
 N. Julka et al. / Computers and Chemical Engineering 26 (2002) 1755 /1769 1769

Kim, M., Choi, D. H., Kim, M. H., & Lee, I. B. (1999). Scheduling Struthers, A. (1997). Use of intelligent software agents in an industrial
system for distribution of crudes to distillation columns in refinery pressure relief and blowdown design tool. Computers and Chemical
processes. In AIChE annual meeting . Dallas, TX. 211i Engineering 21 (Supplement), S83.
Lambert, D. M. & Cooper, M. C. (2000). Issues in supply chain Svoronos, A. & Zipkin, P. (1991). Evaluation of one-for-one replen-
management. Industrial Marketing Management 29 (1), 65. ishment policies for multiechelon inventory systems. Management
Lee, H. L. & Billington, C. (1993). Material management in Science 37 (1), 68.
decentralised supply chains. Operations Research 41 (5), 835. Swaminathan, J. M., Smith, S. F. & Sadeh, N. M. (1998). Modeling
Maguire, P. Z., Scott, D. M., Paterson, W. R. & Struthers, A. (1995). supply chain dynamics: a multiagent approach. Decision Sciences
Development of an advanced modelling environment. Computers 29 (3), 607.
and Chemical Engineering 19 (Supplement), S265. Towill, D. R. (1991). Supply chain dynamics. International Journal of
Maguire, P. Z., Struthers, A., Scott, D. M. & Paterson, W. R. (1998). Computer Integrated Manufacturing 4 (4), 197.
The use of agents both to represent and to implement process Towill, D. R., Naim, M. M. & Wikner, J. (1992). Industrial dynamics
engineering models. Computers in Chemical Engineering 22 (Sup- simulation models in the design of supply chains. International
plement), S571.
Journal of Physical Distribution and Logistics Management 22 (5),
McGreavy, C., Wang, X. Z., Lu, M. L., Zhang, M. & Yang, S. H.
3.
(1996). Objects, agents and workflow modeling for concurrent
Voudouris, V. T. (1996). Mathematical programming techniques to
engineering process design. Computers and Chemical Engineering
debottleneck the supply chain of fine chemical industries. Compu-
20 (Supplement), S1167.
Papageorgiou, L. G., Rotstein, G. E. & Shah, N. (2001). Strategic ters and Chemical Engineering 20 (Supplement), S1269.
Wikner, J., Towill, D. R. & Naim, M. (1991). Smoothing supply chain
supply chain optimization for the pharmaceutical industries.
Industrial and Engineering Chemistry Research 40 (1), 275. dynamics. International Journal of Production Economics 22 (3),
Perea-Lopez, E., Grossman, I. E., Ydstie, B. E., & Tahmassebi, T. 231.
(2001). Dynamic modeling and decentralized control of supply Williams, J. F. (1983). A hybrid algorithm for simultaneous scheduling
chains. Industrial and Engineering Chemistry Research , ASAP of production and distribution in multi-echelon structures. Man-
article (June 19, 2001). agement Science 29 (1), 77.
Sauter, J. A., Parunak, H. V. D., & Goic, J. (1999). ANTS in the Wooldridge, M. & Jennings, N. (1995). Intelligent agents: theory and
supply chain. In Proceedings of workshop on agents for electronic practice. Knowledge Engineering Review 10 (2), 115.
commerce at agents ‘99 . Yang, A., & Yuan, Y. (1999). An object/agent based system modeling
Shah, N. (1996). Mathematical programming techniques for crude oil framework for integrated process plant operations. In AIChE
scheduling. Computer and Chemical Engineering 20 (Supplement), annual meeting . Dallas, TX. 211f.
S1227.