Design of an intelligent supplier relationship management

system: a hybrid case based neural network approach

K.L. Choya,*, W.B. Leea, V. Lob

aDepartment of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom, Hong Kong, Peoples Republic of ChinabHoneywell Consumer Products (Hong Kong) Limited, Hong Kong, Peoples Republic of China

Abstract

In todays accelerating world economy, the drive to continually cut costs and focus on core competencies has driven many to outsource

some or all of their production. In this environment, improving supply chain execution and leveraging the supply base through effective

supplier relationship management (SRM) has become more critical than ever in achieving competitive advantage. It was found that the use of

artificial intelligence in the outsourcing function of SRM to identify appropriate suppliers to form a supply network has become a promising

solution on which manufacturers depend for products, services and distribution. In this paper, an intelligent supplier relationship

management system (ISRMS) using hybrid case based reasoning (CBR) and artificial neural networks (ANNs) techniques to select and

benchmark potential suppliers is discussed. By using ISRMS in Honeywell Consumer Product (Hong Kong) Limited, the outsource cycle

time from searching for potential suppliers to the allocation of order is greatly reduced.

q 2002 Elsevier Science Ltd. All rights reserved.

Keywords: Supplier relationship management; Supplier selection and benchmarking; Supply network; Case based reasoning; Artificial neural network

1. Introduction

The integration of customer/supplier relationship

management (CRM/SRM) to facilitate supply chain

management in the areas of supplier selection using an

artificial neural network (ANN) approach to validate the

search result using CBR technology during the retrieval

stage of the cycle in a real time base is a promising

solution for manufacturers to identify appropriate suppli-

ers and trading partners to form a supply network on

which they depend for products, components, services

and distribution. The result is the formation of an

integrated supply network that allows the most appro-

priate suppliers of the manufacturers to deliver competi-

tively priced, high quality products and services to their

final customers according to their demand effectively.

Choy, Lee, and Lo (2002a) designed a case based SRM

system using a help desk approach and it was then

applied in the purchasing department of an outsource-

type manufacturer in Hong Kong, which has greatly

improved the efficiency in the outsource cycle. Choy,

Lee, and Lo (2002b) also suggested and illustrated

the technique of using case based reasoning (CBR) and

ANN technologies in selecting and benchmarking

potential suppliers during the process of new product

development for manufacturers who outsource a signifi-

cant part of their business. In this paper, an intelligent

supplier relationship management system (ISRMS) using

a hybrid CBR technique to select potential suppliers from

a supplier list, followed by the benchmarking of the

potential suppliers using ANN technique under a CRM/

SRM platform, is discussed. By using ISRMS, manu-

facturers can shortlist and benchmark appropriate suppli-

ers according to the position the supplier is ranked,

resulting in the identification of preferred suppliers with

references to the suitability of the supplier attributes

selected. As a result, the outsourcing cycle time from

searching potential suppliers to the allocation of orders to

the most appropriate supplier can be greatly reduced with

high accuracy.

This paper is divided into seven sections. Section 2

introduces customer relationship management (CRM) and

SRM. Section 3 is about CBR and ANNs and their

suitability in SRM. Section 4 is the development of

ISRMS using a CBR system incorporating major tasks in

SRM to form a distinct intelligent supplier evaluation

system with the aid of the neural network (NN) shell, which

0957-4174/03/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved.

PII: S0 95 7 -4 17 4 (0 2) 00 1 51 -3

Expert Systems with Applications 24 (2003) 225237

www.elsevier.com/locate/eswa

* Corresponding author. Tel.: 852-2766-6597; fax: 852-2362-5267.E-mail address: mfklchoy@inet.polyu.edu.hk (K.L. Choy).

is important for manufacturers wishing to outsource

operations to reliable, suitable suppliers and business

partners. The procedures for constructing the CBR based

supplier selection module and NN based supplier bench-

marking module, which are the critical issue for the success

of ISRMS, are also detailed in this section. Section 5 is

about the application case study, results and benefits by

using ISRMS as an intelligent supplier relationship manage-

ment system in the purchasing department of Honeywell

Consumer Product (Hong Kong) Limited, to aid the

conventional human reasoning process of suppliers selec-

tion. Finally a conclusion of the application of ISRMS in

general is made in Section 6.

2. Customer relationship management and supplier

relationship management (CRM/SRM)

Customer Relationship Management (CRM) is a process

by which a company maximizes customer information in an

effort to increase loyalty and retain customers business over

their lifetimes. The primary goals of CRM are to (a) build

long term and profitable relationships with chosen custo-

mers, (b) get closer to those customers at every point of

contact, and, (c) maximize the companys share of the

customers wallet (Shaw, 2001). Simply stated, CRM is

about finding, getting, and retaining customers. It is at the

core of any customer-focused business strategy and includes

the people, processes, and technology questions associated

with marketing, sales, and service. CRM allows the

formation of individualized relationships with customers,

with the aim of improving customer satisfaction and

maximizing profits, identifying the most profitable custo-

mers and providing them with the highest level of service.

Moreover, in the Internet age, CRM accesses new markets

throughout the world wide web (www) to access world class

capabilities and consequently increase the commoditization

by shortening the product life cycle, and eroding margins. In

summary, CRM is focused on leveraging and exploiting the

interaction with the customer to maximize customer

satisfaction, ensure return business, and ultimately enhance

profitability for all.

It is found that the first wave of CRM processes focused

on sales force, customer and field service automation, with a

heavy emphasis on data access and transaction efficiency by

building out associated data infrastructures (data ware-

houses, segmented data marts, etc.). However, as each

division within global enterprises initiated customer-related

data projects on their own, they neglected acquiring

supports from the companys suppliers effectively, who

have become members in the companys supply chain.

As the trend toward use of technology to drive

competitive advantage has taken root, visionary manu-

facturers are starting to take advantage of a new

competitive opportunity called SRM. Herrmann and

Hodgson (2001) defined SRM as a process involved in

managing preferred suppliers and finding new ones whilst

reducing costs, making procurement predictable and

repeatable, pooling buyer experience and extracting the

benefits of supplier partnerships. It is focused on

maximizing the value of a manufacturers supply base

by providing an integrated and holistic set of manage-

ment tools focused on the interaction of the manufacturer

with its suppliers. In fact there is an interesting and

satisfying symmetry between the role of CRM and the

role of SRM in the manufacturing environment as

illustrated in Fig. 1, which shows an enterprise

applications architecture linking customers with the

supply bases. As companies recognize the value of

managing their supply base as a competitive weapon,

SRM becomes the single most important technology

investment they can make to ensure that supply chain

transformation they are driving is successful (Herrmann

& Hodgson, 2001). In fact, SRM improves the flow of

product demand and supply information throughout the

supply chain by four kind of activities. They are indirect

and direct procurement, sourcing, and trading exchange.

In fact, by integrating CRM with SRM properly through

the process of product design and development, and the

application of supply chain management under the

platform of an ERP system, SRM solutions can provide

significant competitive advantage by delivering value in

three important areas: (1) dramatic cost savings, (2)

increasing flexibility and responsiveness to customer

requirements, and (3) substantially faster cycle times.

Together these benefits can lead to meaningful faster

time to market share in the course of the product life

cycle based on customer demand with a maximum

degree of customization.

Fig. 1. Enterprise applications architecture.

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237226

3. Case based reasoning and artificial neural network

CBR is a plausible generic model of an intelligent,

cognitive science-based method by the fact that it is a

method for solving problems by making use of previous,

similar situations and reusing information and knowledge

about such situations (Kolodner, 1993). CBR combines a

cognitive model describing how people use and reason from

past experience with a technology for finding and presenting

such experience. It is a problem-solving paradigm that is, in

many aspects, different from other AI approaches CBR

provides a conceptual framework in which to store operator

experience and to later provide that experience to other

operators to facilitate situation assessment and solution

formulation processes. This is accomplished by providing a

context in which the human operator can view the current

state and recent activities of the system and gain easy access

to previous experience.

ANNs is an information-processing paradigm inspired by

the way the densely interconnected, parallel structure of the

mammalian brain processes information. Stock (1997)

concluded that a NN is a computer with an internal structure

that imitates the workings of the human brain and the

nervous system. In theory, the formation of NN is similar to

the formation of neural pathways in the brain as task is

practiced. NN is capable of learning from its mistakes by

adjusting the weight of the nodes so that repetitive tasks can

be accomplished more accurately the more number of times

it is trained (Dhar & Stein, 1997). In brief, it can be said that

a NN can lend itself to be a highly appropriate method with

its power on predicting the likely outcome based on what it

has learned before from previous historical data.

3.1. Process and applications of CBR

The processes of the CBR technique are case retrieval,

reuse, revise and retain (Aamodt & Plaza, 1994). The step

includes: (1) retrieve the most similar situation from a set of

cases, according to enquiry or request; (2) reuse the cases to

solve the problem in order to construct the solution for the

new problem. This solution becomes the output of a

proposed solution; (3) revise the suggested solution if there

is a difference between the new problem and the retrieved

case. This solution is verified and exported as a solution; (4)

retain the new solution in a case database for future usage.

Through the CBR cycle, it can be seen that if the best-

retrieved case is a perfect match, then the system has

achieved its goal and finishes. However, it is more usual that

the retrieved case matches the problem case only to a certain

degree. In this situation, the closest case may provide a sub-

optimal solution or the closest retrieved case may be revised

using some pre-defined adaptation formulae or rules.

Adaptation in CBR systems means that such systems have

a rudimentary learning capability, which can improve or

become more discriminatory, as the number of case

increases.

CBR applications can be broadly classified into two main

problem types, namely, classification tasks and synthesis

tasks (Watson, 1997). Classification tasks cover a wide

range of application that all share certain features in

common. A new case is matched against those in the case

base to determine what type, or class, of case it is. The

solution from the best matching case is then reused.

Classification tasks come in a wide variety of forms such

as planning (Kolodner, 1993), diagnosis (Watson &

Abdullah, 1994), design (Perera & Watson, 1995) and

process control (Koegst, Schneider, Bergmann, & Vollrath,

1999). Synthesis tasks attempt to create a new solution by

combining parts of previous solutions. Synthesis tasks are

inherently complex because of the constraints between

elements used during synthesis. CBR systems that perform

synthesis tasks must make use of adaptation and are usually

hybrid systems combining with other techniques. The

synthesis system is mainly used in planning and configur-

ation (Costas & Kashyap, 1993).

Besides the above applications in industry, there is a

growing trend for enterprises to apply CBR technology in

the customer service area. The history can be tracked back

to Acorn and Walden (1992), who discussed the develop-

ment of a CBR system, called support management

automated reasoning technology (SMART) system, which

was developed jointly by Compaq with Inference Corpor-

ation, to enhance the customer support service level in

Inference Corporation, in order to increase customer

loyalty. It was reported that by using this case based

oriented help desk approach in solving customers request,

companies can retain customer loyalty to its product. It has

been suggested that CBR is effective in make or

buy decision-making process. It has become clear

that CBR is useful in searching the knowledge, helping

users in comparing various tasks and items, automatically

notifying users with relevant new knowledge update,

and so on (Dutta, Wierenga, & Dalebout, 1997; Pawar,

Haque, Belecheanu, & Barson, 2000; Choy & Lee, 2000,

2001, 2002a).

In summary, it can be seen that the application of CBR

technique in the areas of SRM such as supplier selection is a

new approach, which can be used in integrating with CRM

through the process of new product development and supply

chain management. Since CBR is an advanced reasoning

technique simulating human reasoning to retrieve a relative

case, modify it and find a solution for the new coming

problem, it can be used to supplement the conventional

measures, which mainly rely on experts such as the

purchasing manager or procurement engineer, to make

decision on outsourcing matters.

3.2. Process and applications of ANNs

The basic element in a NN is a neuron. Each neuron is

linked to certain of its neighbors with varying coeffi-

cients of connectivity that represent the strengths of these

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237 227

connections. Learning is accomplished by adjusting their

strength so that neurons can then be grouped into layers. The

input layer consists of neurons that receive input from the

external environment. The output layer consists of neurons

that communicate the output of the system to the user or

external environment. There are usually a number of hidden

layers between these two layers. As shown in Fig. 2, the

hidden layer of a NN model acts as a black box to link the

relationship between input and output. When the input layer

receives the input, its neurons produce output, and

consequently they will become input of the other layers in

the system. The process continues until a certain criterion is

satisfied or until the output layer is invoked, it then passes

the output to the external environment. Generally speaking,

a NN acts as a black box, which matches sets of input

output pairs by adapting its internal degrees of freedom and

the weights.

ANNs have been widely used in many classification and

optimization situations, where history data are used to

train the network, automatically determining the most

appropriate configuration of the hidden networks (Gurney,

1997; Mehrotra, Mohan, & Ranka, 1997). Due to their

ability in modeling human associative memory, ANNs are

well suited on the application for product design (Zhang &

Huang, 1995) and cellular manufacturing systems (Rao &

Gu, 1995; Kusiak & Lee, 1996); for process planning

(Dimla, 1999); for monitoring and diagnosing (Chang &

Ho, 1999); for control (Hao, Shang, & Vargas, 1995) and

scheduling (Grabot, 1998).

So far, the research area on application of ANNs to the

supplier selection or supplier benchmarking is under-

developed. There are few examples concerning the usage

of NN approach to select or benchmark suppliers. Wei,

Zhang, and Li (1997) described a system that applied a

technique with NNs to select suppliers using the perform-

ance history, geography and price of a supplier as

determinant factors effecting the supplier selection. How-

ever, the criteria for selecting suppliers only fall into the

quantitative categories. There are not any criteria concern-

ing qualitative categories for selecting a supplier, meaning

that the system for selecting a supplier was not

comprehensive.

It can be seen that by using NN technique to incorporate

both quantitative and qualitative supplier attributes, the

process of supplier selection and benchmarking can be done

effectively by classifying suppliers into different categories

according to their capabilities, suitable to be used with

reference to the particular cases being considered. It is a new

and promising method suitable for manufacturers especially

for those who outsource a significant part of their business.

4. An intelligent supplier relationship management

system

The ISRMS proposed in this paper is used for an

organization to master its core competence by means of

subcontracting operations that cannot be used in building up

the knowledge inside the organization.

ISRMS uses one of the applications of CBR technology,

which belongs to the classification task, integrating CRM

with SRM to provide the prediction and assessment of

supplier capability. The architecture of the ISRMS is shown

in Fig. 3. It is divided into front end and back end. In the

front end, ISRMS is linked with the companys local and

overseas customers such as wholesalers and retailers to

acquire information related to their products.

The back end of ISRMS consists of a case database

containing cases stored in various departments within the

companys local and overseas offices as well as in other

plants. The CBR engine which uses the stored cases for

problem solving is also installed in the back end. By doing

Fig. 2. A NN model.

Fig. 3. Architecture of ISRMS.

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237228

so, CRM and SRM functions link the front and back end of

ISRMS.

The intelligent (AI) computational tool, the CBR engine,

embedded inside ISRMS is a new concept since there are no

similar systems in the market using the CBR technique to

manage and sort the potential supplier to form a customer

supplier integration strategies. The CBR engine, in which a

case base decision tree is constructed, arranges the past

practices in a systematical way for the retrieval process. For

the case base module, each case includes the name of a case,

a case description, a recommended solution and the further

reference linkages such as the inter-link between the

updated files or the front page on the Internet of the target

supplier. In order to implement the concept of the SRM, the

database in the form of case base is built on a web site for

collecting opinions from the customers.

As shown in Fig. 4, the two modules in the back end of

ISRMS responsible for matching the customer demand with

the respective supplier capability for a particular product are

the case based supplier selection module (CSSM) and the

NN based supplier benchmarking module (NNSBM).

4.1. Supplier selection module (CSSM)

A built-in supplier selection workflow (SSW) is pro-

grammed in the case base engine such that an authorized

supplier list is stored, consisting of three profiles: technical

capability, quality assessment and organization profile (Choy

et al., 2002a). Data for each supplier are stored in a case

structure, each consisting of a number of fields representing

the criteria in each category. The cases in ISRMS show the

relevant numerical performance values of the correspondent

criteria of suppliers. During the supplier selection process, an

external supplier list is converted to the format of a case and is

then imported to the case base of the authorized supplier list

when either the case retrieval is exhaustive or new potential

suppliers are desired in sourcing (Choy & Lee, 2002b, 2003).

In order to facilitate the operations of ISRMS, the CBR

engine analyzes the product data collected from the web site

such that supplier performance according to customer

feedback is transformed to the requirements into organiz-

ation when selecting appropriate suppliers in the business.

This approach enables the transfer of customer demand to the

related suppliers directly.

4.2. Supplier benchmarking module (NNSBM)

The supplier benchmarking function is done by the NN

based supplier-benchmarking module (NNSBM) in ISRMS

to choose and benchmark the most suitable supplier from

the list of potential suppliers generated by CSSM. The NN

technique is used in benchmarking suppliers as it out-

performs other methods in two ways.

1. In general, there are a large number of computation

activities to be performed at the same time during

supplier benchmarking process. A NN is inherently

parallel and naturally amenable to expression in a

parallel notation. Therefore, it is a superior method in

supplier benchmarking process.

2. A NN model is able to deal with problems requiring

highly complex quantitative and qualitative reasoning

such as the evaluation of suppliers attributes.

A large amount of input data is required for training the

NN model. The historical cases in the CSSM provide a large

training data set for the design of NNSBM. This would

enhance the accuracy level of the benchmarking result.

Fig. 4. CRM/SRMsupplier selection and benchmarking.

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237 229

Generally speaking, a NN model is capable of learning

from its mistakes by adjusting the weight of the nodes so

that repetitive tasks can be accomplished more accurately

the more times it is trained (Dhar & Stein, 1997). This

characteristic is relevant to use in the benchmarking process

of the SRM system for supplier evaluation, selection and

benchmarking processes in outsourcing manufacturing.

Here, a back-propagation NN that includes hidden layers

is used in the process of supplier benchmarking. The reasons

of selecting back-propagation type are because of its

generality and the ease of implementation for most of the

NN systems (Medsker & Liebowitz, 1994).

After all the potential suppliers for the particular new

product are determined by CSSM, NNSBM is then

responsible for benchmarking these potential suppliers in

order to find the most suitable one. This is done by

comparing the performance scores of each supplier to find

the most suitable one. An example of an input and output

model, which will be used in the later application case study

in Honeywell, shows the input and output nodes of NNSBM

in Fig. 5 for illustration purposes.

In this case, there are eleven input nodes each represents

a supplier attribute of a potential supplier recommended by

CSSM. Five output nodes are generated which represents

the recommendation of NNSBM after analyzing the input

attributes in the hidden layers through the NN mechanism.

The recommendation will be used in deciding the suitability

of this potential supplier to the requirement when out-

sourcing activities are required.

In brief, the first part of tasks undertaken by ISRMS relies

on the past record of the candidate companies for the short-

listing process. CSSM is adopted as the appropriate

approach for this purpose. The second part is related to

benchmarking assessment and is adopted by the NNSBM of

the system. In doing this, a benchmarking model is built to

select the most appropriate suppliers, providing relevant

facts and data about current as well as projected performance

based on various factors for comparison among the short-

listed suppliers. This model, based on the trends of past data,

predicts how well (or poor) a supplier will be performing

over the forthcoming period of time. The process of CSSM

and NNSBM of the system is shown in Fig. 6.

4.3. Construction of the key components of ISRMS

The steps for constructing the two key components of

ISRMS, which is the CSSM and the NNSBM are shown in

the following sections.

4.3.1. Case based supplier selection module

There are eight steps in constructing the generic supplier

selection mechanism using CBR technology to form the

CSSM.

Step 1: design the supplier attributes acceptance criteria.

The required attributes to be used in evaluating the

performance of suppliers are selected. They are grouped in a

hierarchical form.

Fig. 5. Mapping of input and output layers of NNSBM.

Fig. 6. The process of CSSM and NNSBM in selecting and benchmarking

suppliers.

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237230

Step 2: collect the information of potential suppliers,

analyze and add weightings into the criteria.

All suppliers are categorized into big, medium or small

classes, according to the size, number of workers within the

company. The information of potential suppliers such as the

name and past practices are recorded. This step can also be

used in establishing the best practice case for comparing the

retrieved case when the CBR engine is used in evaluating

potential suppliers.

Step 3: construct the number of layers, workflow of the

decision tree.

The number of layers required to construct the decision

tree, which depends on the way attributes are categorized

into tiers, is determined. In CSSM, five tiers are formed,

which are shown in Fig. 7. The selected attributes are

divided into the first and second tier of the generic system,

based on the classification of the supplier performance. The

number of attributes in tier 1 and tier 2 are adjustable,

according to the actual policy in different company, to suit

company strategies. In tier 3 and 4, a general and detail

categorized coding systems, which represents a companys

coding system, are designed. This function is to sort the

potential suppliers into the right product category. In tier 5,

suppliers are classified into high and low according to the

resulting scores.

Step 4: build the basic hierarchy of the decision tree

(Fig. 7).

This is the most important step in building the CSSM.

The logic in the decision tree represents the path for

the required solution. In order to do so, several steps are

carried out.

(a) Based on the requirement of each company, determine

the primary supplier selection attributes, e.g. delivery,

quality, etc. and take them from the attribute library

into tier 1.

(b) Select the secondary attributes from the library and

link them to the primary attributes.

(c) Evaluate the companys coding system; divide them

into different material categories such as plastic,

mechanical parts, and accessories. The coding system

is grouped in the form of a bill of materials (BOM).

Finally, a weight is added to each attribute.

(d) List all the outsourced material codes for the purposes

of linking potential suppliers (according to the score

they got from the selection process).

(e) Divide the score point into two classes; low score

(15 points) and high score (610 points). The

purpose of this classification is to maintain the case

base thus formed easily, as well as to speed up the

retrieval process.

By following these five steps, a logical hierarchy decision

tree for the case base engine to function accurately is built.

Step 5: build the cases into a CBR engine.

The information of each supplier is fed in the form of

individual cases to an embedded CBR engine called Case

Advisor. It is traded by Sententia Software Company at

Simon Frazier University, Canada. It is a CBR software

package modeled on a standard technique where human

uses in problem solving, which is composed of two

modules, namely, Case Authoring (CA) and Problem

Resolution (PR). CA is used in constructing the case base

while PR is used in finding the potential solution. The

following information is usually input in the two modules

for the CBR engine to function.

(a) In CA module:

case or supplier name:

case description, which includes:

(i) the existing supplier code;

(ii) the business nature (e.g. plastic, mechanical,

electronic parts manufacturers, etc.);

(iii) the major type of suppliers;

(b) In PR module:

case solution, which includes:

(i) the score points of each criterion;

(ii) the establishment of a link to the past practices;

Finally, the name of the material categories, which acts

as a medium in linking the different tiers of hierarchy tree to

form the case base of the CSSM in searching for the

appropriate suppliers is input.Fig. 7. Mechanism of constructing the generic CSSM.

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237 231

Step 6: the linkage of cases and relative questions.

All questions related to the attributes in tier 1 and 2 are

set, together with a list of preferred answers using a scale of

110 for scoring purposes. To link the material categories

of the company with the supplier, all questions are dragged

and dropped to each suppliers case database address to

form a decision tree, which describes the suppliers

capability in terms of scores for making a certain product.

Finally, the weighting of each material category is input.

Step 7: design the score point in different performance

measurement criteria of the target case.

Before starting the user interface for finding solutions,

the expected scores of the required supplier are prioritized.

Step 8: retrieve the solution by using CBR technology.

To search the potential suppliers, simply type in the key

word for the outsourcing material (e.g. PP resin). Then, it is

only needed to type in the expected score for each question

to allocate the potential suppliers.

In summary, steps 1 and 2 are the preparation stage. Steps

36 represent the processes to configure cases in the CA

module. Steps 7 and 8 are the processes in finding solutions

using PR of the CBR engine (Case Advisor) inside the CSSM.

4.3.2. Neural network based supplier benchmarking module

(NNSBM)

In the following section, a back propagation supervised

learning model is designed by using a NN package, Qnet

(2000), which is embedded in NNSBM.

The design flow of NNSBM is divided into three stages,

namely, the design, training and generalizations stage,

which are shown in Fig. 8. Each stage has its role to play in

the model design.

1. The design stage aims at selecting input and output

variables, training method selection and hidden layer

design.

2. The training stage is to select a training method, adjust

and verify the training parameters.

3. The generalization stage is the stage in which the NN

classifies any unseen data pattern countered during the

training stage to an acceptable pattern. In NNSBM, this

stage is to recall and validate the model.

4.3.2.1. Stage 1: design stage. In this stage, three kinds of

variable are to be determined. They are the input and output

attributes selection, the selection of the training method and

the design of the hidden layer.

(1) Input and output attributes selection. Input and output

selection is always a complex task for the NN model

developer as there is no formal method for selecting

variables for a model. Moreover, both under and over

specification of input variables will most often generate sub-

optimal performance of the NN model. The result is that on

one side, if there are too many input variables, it can bring

about poor generalization. On the other side, if there is

insufficient amount of information representing critical

decision criteria given to the model, then it is unable to

develop a correct and accurate model. In general, it is

important to design independent input variables since

correlated variables will degrade model performance by

interacting with one another to produce a biased effect.

(2) Training method selection. After selecting the

variables for the input and output, the next step is to

determine the kind of training to be employed to best fit the

problem. The learning method can be divided into two

distinct categories, namely, unsupervised learning and

supervised learning. Both require a collection of training

examples that enable the NN to acquire the data set and

produce accurate output values. For simplicity, the super-

vised learning method is adopted.

(3) Hidden layer design. The design of hidden layer is

dependent on the selected learning algorithm. For example,

unsupervised learning methods normally require the

quantity of nodes in the first hidden layer equal to the size

of the input layer. Supervised learning systems are generally

more flexible in the design of hidden layers. Barnard and

Wessels (1992) emphasized that an increment of the

number of hidden unit layers enables a trade-off between

Fig. 8. Design flow of NNSBM.

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237232

smoothness and closeness-of-fit. A greater quantity of

hidden layers enables a NN model to improve its

closeness-of-fit, while a smaller quantity improves the

smoothness or extrapolation capabilities of it. It was

concluded that the number of hidden layers is heuristically

set by determining the number of intermediate steps to

translate the input variables into an output value. According

to Patuwo, Hu, and Hung (1993), the number of hidden layer

nodes can be up to 2n 1 (where n is the number of nodesin the input layer). Several quantitative heuristics exist for

selecting the quantity of hidden nodes for an ANN such as

using 75% of the quantity of input nodes (Lenard, Alam, &

Madey, 1995), or using 50% of the quantity of input and

output nodes (Piramuthu, Shaw, & Gentry, 1994). In our

model, a compromise is made between these methods in

fixing the number of hidden nodes in order to minimize the

number of nodes without sacrificing the accuracy level.

4.3.2.2. Stage 2: training stage. The goal of the training

stage is to obtain an accurate NN model. Linilson (1999)

stated that whether the NN has a good generalization

capability or not depends on the resulting error level.

Basically, the lower the error, the better the model. In the

training stage, the selection of the transfer function, learning

rate, momentum and exit condition setting, the root mean

square (RMS) and correlation coefficient checking and

verification of the model are needed.

(1) Transfer function. A transfer function is needed to

introduce the non-linearity characteristics into the network.

The non-linear function will make the hidden units of multi-

layer network more powerful than just plain perception. The

transfer function used is a standard function for back-

propagation, that is, the sigmoid transfer function. The

sigmoid transfer function is chosen due to its ability to help

the generalization of learning characteristics to yield models

with improved accuracy.

(2) Parameters. The back-propagation training paradigm

uses three controllable factors that affect the algorithms

rate of learning. They are the learning rate coefficient (h ),momentum (a ), and the exit condition.

(i) Learning rate (h ). The learning coefficient governs thespeed that the weights can be changed over time,

reducing the possibility of any weight oscillation

during the training cycle.

(ii) Momentum (a ).The momentum parameter controlsover how much iteration an error adjustment

persists. There is no definitive rule regarding the

momentum, a. In general, it is set to 0.5, which ishalf of the maximum limit for training to reduce the

damping effect.

(iii) Exit condition. NNs use a number of different

stopping rules to control the termination of the

training process. For the training of the NNSBM

model, the rule of stop it after a specified number

of epochs was applied.

(3) Verification. The residual entropy of the trained

network is a measure of its generalization. When the

residual entropy increases, the performance of the general-

ization decreases, meaning that the model still needs

modification. The residual entropy is monitored during

training by means of the RMS error value. It is the error

between the networks output response and the training

target.

4.3.2.3. Stage 3: generalization stage. There are two steps to

be performed in order to accept the model. One is recall and

the other is validation.

(i) Recall. A validation data set is applied to check the

degree of the generalization of the trained model. By doing

so, the size of the generalization error is determined and

minimized. This step is called Recall.

(ii) Validation. A network is said to be generalized well

when the output is correct or close enough for an input. In

such cases, the model is ready for use.

5. Application case study and results

ISRMS was applied with the intention of strengthening

the SRM function in Honeywell Consumer Products (Hong

Kong) Limited. Honeywell is a multi-national based

manufacturer of consumer products such as fans, heaters,

humidifiers, air cleaners, etc. Its office is in Hong Kong and

the main manufacturing plant in Shenzhen, Mainland China.

Honeywell employs around 2800 workers and staff, 2500

are located at the Shenzhen office and 300 at the Hong Kong

office. Honeywell has around 50 core suppliers and over

2000 potential suppliers for sourcing. Its vision is to build

up a knowledge-based learning organization, which can

produce quality and innovative products with enhanced

services to their customers on time within budget, through

the collaborative product commerce via global networking

with their strategic suppliers and customers for operational

excellence. ISRMS can help to select appropriate suppliers

to develop new products according to customer demand

received from the global network through its embedded

CRM module.

(a) Potential suppliers selection. For example, when a

user wants to search the potential suppliers for supplying the

air cleaner cover, Case Advisor embedded in CSSM is used.

As illustrated in Fig. 9, the keywords of the problem

description for searching potential suppliers for the air

cleaner cover are entered in the case base in the space

provided in the upper left corner. After entering the

keywords, the CBR engine starts to retrieve appropriate

suppliers by using the nearest neighbor technique. A

question list is shown in the questions area where the

users select and answer those that are relevant to the task.

The rank of the suppliers change until it reaches the final

answer according to the questions provided. A list of

potential suppliers is shown in the left bottom screen.

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237 233

The detail information of each supplier is shown on the right

screen when double clicking the particular supplier in the

list at the left hand screen (Choy et al., 2002a).

This list of suppliers is then benchmarked with the past

best practice by NNSBM.

(b) Supplier benchmarking. After performing supplier

selection by using the CSSM, a list of potential suppliers is

generated. Supplier benchmarking is then carried out by

using Qnet (2000) embedded inside NNSBM. The past

performance of the potential suppliers in terms of scores are

then transferred from CSSM to NNSBM through the MS

Excel worksheet environment, where the scores of the

potential suppliers are pasted on the respective cells. After

the transfer of data from Case Advisor, the engine of CSSM,

supplier benchmarking is performed by activating the Qnet

Tool Execution Button, which has a macro written to copy

all the performance scores to the Qnet Tool, which is an

embedded module in Qnet (2000). Once all the data is

included, the trained model is triggered to perform the recall

function for benchmarking the selected potential suppliers.

After running the recall function in the trained model, the

final supplier benchmarking result is computed and shown

in Fig. 10, where data in the upper cell represent the input

nodes (in this case it contains 11 input nodes) and the result

of the output nodes (five output nodes) is shown in the lower

cell. The output data of NNSBM are then transferred back to

the excel worksheet to perform the last benchmarking step.

That is, to sort out the most competent supplier and rank

them in a descending order.

As seen from Fig. 11, the head row shows the potential

suppliers detail including their performance scores from

Score I to Score V, which represents the output nodes of

NNSBM. As the value of Score I represents the degree

of potential competence among suppliers (Fig. 5), there-

fore, the Score I of different suppliers are compared in

this case.

In Fig. 11, Logistic Industrial Supply Co. Ltd got the

highest score, 6.5137. As a result, it is recommended

automatically by NNSBM as the top rank supplier based on

past suppliers performance and matching with the require-

ment of the product characteristics. The supplier thus

recommended is the most suitable one according to ISRMS.

After implementing ISRMS for the selection of potential

suppliers in various projects, the performance is compared

with those using the experience-based approach. The

performance measurement criteria are the Honeywell

satisfactory rate, degree of delay in delivery, quality

below standard and the customer claims. As shown in

Table 1, while there are improvements using solely the CBR

approach in the process of supplier selection (Choy et al.,

2002b), results using the hybrid CBR NN approachoutperform the CBR approach and indicate that the adoption

of ISRMS has had a significant contribution to Honeywell

(Hong Kong) plant, which is shown by the increase of

Honeywell satisfactory rate. The chance of selecting the

right suppliers/partners is increased resulting in the

production of more reliable products produced. This can

be shown by the significant decrease in the percentage of

delay in delivery, quality below standard and the customer

claims.

By using ISRMS, other results and benefits are noticed.

Using the parallel processing of the new product develop-

ment process and SSW mechanism of ISRMS, the preferred

suppliers for a specific new product can be matched more

quickly. Furthermore, it is unavoidable that the modification

of the specification of a new product can trigger the change

of objective(s) and rule(s) of retrieval and weightings of

attributes in the supply management system simultaneously

in the early phase. For example, the changing of the design

when the material of a product is changed from metal to

plastic would result in critical changes of all parameters of

supplier sourcing development. ISRMS reacts by retrieving

similar cases and at the same time suggests the updating of

its case base by means of supplier adoption if the retrieved

case is unsatisfactory. Moreover, by using the CBR NNapproach, suppliers can be divided into different categories

Fig. 9. Finding potential suppliers by CSSM.

Fig. 10. The input data versus output data of potential suppliers in NNSBM.

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237234

such that the most competent suppliers can be identified

easily. According to this automatic categorization of

suppliers, follow up actions can be carried out if it is

indicated that further assessment to some suppliers is

required as suggested from this hybrid approach. In this

way, ISRMS can enhance the function of supply manage-

ment system in new product development by making it more

agile. In addition, ISRMS can learn from experience and

enrich its case base throughout the searching process.

As Honeywell is a multi-national manufacturer, its

suppliers and business partners number thousands. The

usage of ISRM to handle highly structured supplier cases

enables purchasing managers, engineers, and buyers to

investigate different scenarios with large finite supplier

cases effectively. This is due to the systems ability to

combine with a what-if analysis and the actionable

knowledge obtained from the case base in the case adviser.

The speed of sourcing the capable, preferred or potential

suppliers and decision-making in the related attributes is

increased. Consequently, quicker reaction in supplier

selection and management occurs, allowing a shorter new

product development cycle time to be achieved, and

accordingly, cost reduction becomes possible. Another

benefit is that the major part of ISRMS records the

knowledge of workflow and can be fully implemented

into the CRM module with little human involvement.

This can solve the problem of losing supplier selection

knowledge when experienced key staff leave the

corporation. In fact, new staff can allocate preferred

suppliers easily by the help of both CSSM and NNSBM

in ISRMS. In summary, ISRMS records the corporate

knowledge in the supplier case base, its algorithm and

workflow to support the CRM functions.

6. Conclusions

SRM involves the management of preferred suppliers

and finding new ones whilst reducing costs, making

procurement predictable and repeatable, pooling buyer

experience and extracting the benefits of supplier

partnerships, while CRM focused on leveraging and

exploiting the interaction with the customer to maximize

customer satisfaction, ensure return business, and ulti-

mately enhance customer profitability. It becomes crucial

for manufacturers to integrate the demand of customers

to their preferred suppliers as well as sourcing new ones

in a real time base during the new product development

cycle in order to remain competitive in business. The

major function of ISRMS is to integrate the customer

Table 1

Supplier selection performance by human and by ISRMS

By experience ISRMS

(CBR)

ISRMS

(Hybrid

CBR NN)

Honeywell

expected

Honeywell

satisfactory

rate (%)

65 90 95 99

Delay in

delivery (%)

20 10 10 10

Quality below

standard (%)

30 25 17 15

Customer

claims (%)

25 18 17 15

Fig. 11. Results of the supplier benchmarking process.

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237 235

requirements on product quality, delivery time, and

manufacturing cost by two advanced computational

retrieval technology called CBR and NNs, to evaluate

and benchmark suppliers through a single software

platform. With the implementation of ISRMS, an

organization can shorten the workflow of selecting and

benchmarking business suppliers on receiving a new

order. In addition, the potential suppliers retrieved from

CSSM are categorized into a different category by

NNSBM. In doing so, orders can be assigned to the

most competent suppliers appropriately. The workflow of

the new system minimizes the human involvement in

routine task in the system, and speeds up and enhances

the consistency of the decision making process. An

ISRMS solution can help manufacturers to reduce the

total production time and time to market through the

effective outsourcing its works to the most appropriate

suppliers. The computerization of the customersupplier

management process helps the enterprise to fully

implement a CRM strategy in each department, to

achieve a close relationship with suppliers/partners by

the integration with the SRM strategy, and consequently

increase the manufacturers own competitiveness, repu-

tation and revenue in the market. By using ISRMS, it is

possible for a manufacturer to build long term and

profitable relationships with chosen customers, and to

maximize the value of its supply base by increasing

flexibility and responsiveness to customer requirements

and substantially faster cycle times.

Acknowledgements

The authors wish to thank the Research Committee of the

Hong Kong Polytechnic University for the support of this

project.

References

Aamodt, A., & Plaza, E. (1994). Case-based reasoning: foundational issues,

methodological variations, and system approaches. AIComArtificial

Intelligence Communications, 7(1), 3959.

Acorn, T. L., & Walden, S. H. (1992). SMART: Support management

cultivated reasoning technology for compaq customer service. In S.

Scott, & P. Klahr (Eds.), Innovation Applications of Artificial

Intelligent 4. Proceedings of AAAI, 92, Cambridge, MA: AAAI

Press/MIT Press.

Barnard, E., & Wessels, L. (1992). Extrapolation and interpolation in neural

network classifiers. IEEE Control Systems, 12(5), 5053.

Chang, S. I., & Ho, E. S. (1999). A two-stage neural network approach for

process variance change detection and classification. International

Journal of Production Research, 37(7), 15811599.

Choy, K. L., & Lee, W. B. (2000). Task allocation using case-based

reasoning for distributed manufacturing systems. Logistics Information

Management, 13(3), 167176.

Choy, K. L., & Lee, W. B. (2001). Multi-agent based virtual enterprise

supply chain network for order management. Journal of Industrial

Engineering Research, 2(2), 126141.

Choy, K. L., & Lee, W. B. (2002a). On the development of a case-based

supplier management tool for multi-national manufacturers. Measuring

Business Excellence, 6(1), 1522.

Choy, K. L., & Lee, W. B. (2002b). A generic tool for the selection and

management of supplier relationships in an outsourced manufacturing

environmentthe application of case based reasoning. Logistics

Information Management, 15(4), 235253.

Choy, K. L., & Lee, W. B. (2003). A generic supplier management tool for

outsourcing manufacturing. Supply Chain Management, An Inter-

national Journal, in press.

Choy, K. L., Lee, W. B., & Lo, V. (2002a). Development of a case based

intelligent customersupplier relationship management system.

International Journal of Expert Systems with Applications, 23(3),

281297.

Choy, K. L., Lee, W. B., & Lo, V. (2002b). An intelligent supplier

management tool for benchmarking suppliers in outsource manufactur-

ing. International Journal of Expert Systems with Applications, 22(3),

213224.

Costas, T., & Kashyap, P. (1993). Case-based reasoning and learning in

manufacturing with TOTLEC planner. IEEE Transactions on Systems,

Man, and Cybernetics, 23(4).

Dhar, V., & Stein, R. (1997). Intelligent decision support methods: The

science of knowledge work. Englewood Cliffs, NJ: Prentice-Hall.

Dimla, S. (1999). Application of perceptron neural networks to tool-state

classification in a metal-turning operation. Engineering Applications of

Artificial Intelligence, 2(4), 471477.

Dutta, S., Wierenga, B., & Dalebout, A. (1997). Case-based reasoning

systems: from automation to decision-aiding and simulation.

IEEE Transactions on Knowledge and Data Engineering, 9(6).

Grabot, B. (1998). Objective satisfaction assessment using neural nets for

balancing multiple objectives. International Journal of Production

Research, 36(9), 23772395.

Gurney, K. (1997). An introduction to neural networks. London:UCL

Press.

Hao, G., Shang, J. S., & Vargas, L. G. (1995). A neural network approach

for online control of flexible manufacturing systems. International

Journal of Production Research, 33(10), 28352854.

Herrmann, J. W., & Hodgson, B (2001). SRM: Leveraging the supply base

for competitive advantage. Proceedings of the SMTA International

Conference, Chicago, Illinois, 1 October, 2001.

Koegst, M., Schneider, J., Bergmann, R., & Vollrath, I., (1999). IP retrieval

by solving constraint satisfaction problems. FDL 99, Second

International Forum on Design Languages, Ecole Normale Superieure

de Lyon, Lyon, France August 30September 3, 1999.

Kolodner, J. (1993). Case-based reasoning. San Mateo, CA: Morgan

Kaufmann.

Kusiak, A., & Lee, H. (1996). Neural computing-based design of

components for cellular manufacturing. International Journal of

Production Research Society, 34(7), 17771790.

Lenard, M. J., Alam, P., & Madey, G. R. (1995). The application of

neural networks and a qualitative response model to the auditors

going concern uncertainty decision. Decision Sciences, 26(2),

209227.

Linilson, R. P (1999). Comparison of neural networks performance for fault

classification in rolling bearing. Proceedings of the IASTED Inter-

national Conference Applied Modeling and Simulation, 13 Septem-

ber, 1999, Cairns, Australia, pp. 611.

Medsker, L., & Liebowitz, J. (1994). Design and development of expert

systems and neural networks. New York: Macmillan.

Mehrotra, K., Mohan, C. K., & Ranka, S. (1997). Elements of artificial

neural networks. Massachusetts, MA: The MIT Press.

Patuwo, E., Hu, M. Y., & Hung, M. S. (1993). Two-group classification

using neural networks. Decision Sciences, 26(6), 749779.

Pawar, K. S., Haque, B. U., Belecheanu, R. A., & Barson, R. J. (2000).

Towards the application of case based reasoning to decision-making in

current product development (concurrent engineering). Knowledge-

Based Systems, 13, 101112.

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237236

Perera, R. S., & Watson, I. (1995). A case-based design approach for the

integration of design and estimating. In I. Watson (Ed.), Progress in

case-based reasoning. Lecture notes in artificial intelligence 1020,

Berlin: Springer.

Piramuthu, S., Shaw, M., & Gentry, J. (1994). A classification approach

using multi-layered neural networks. Decision Support Systems, 11(5),

509525.

Qnet (2000). Vesta Services, Inc., Winnetka, IL 60093, http://qnetv2k.com.

Rao, H. A., & Gu, P. (1995). A multi-constraint neural network for

the pragmatic design of cellular manufacturing systems.

International Journal of Production Research, 33(4), 10491070.

Shaw, R. (1999). CRM definitionsdefining customer relationship

marketing and management. In SCN Education BV (Ed.), Customer

Relationship Management: The Ultimate Guide to the Efficient Use

of CRM. HOTT Guide, 2001, www.hottguide.com.

Stock, J. R. (1997). Applying theories from other disciplines to logistics.

International Journal of Physical Distribution and Logistics Manage-

ment, 27(9/10), 515539.

Watson, I. (1997). Applying case-based reasoning: Techniques for

enterprise systems. CA, USA: Morgan Kaufmann.

Watson, I. D., & Abdullah, S (1994). Developing case-based reasoning

systems: A case study in diagnosing building defects. Proceedings of

the IEE Colloquium on Case-Based Reasoning: Prospects for

Applications, Digest No. 1994/057, 1/11/3.

Wei, S., Zhang, J., & Li, Z (1997). A supplier-selecting system using a

neural network. IEEE International Conference on Intelligent Proces-

sing Systems, October 1997.

Zhang, H. C., & Huang, S. H. (1995). Application of neural networks in

manufacturing: a state of the art. International Journal of Production

Research, 33(3), 705728.

K.L. Choy et al. / Expert Systems with Applications 24 (2003) 225237 237

Design of an intelligent supplier relationship management system: a hybrid case based neural network approachIntroductionCustomer relationship management and supplier relationship management (CRM/SRM)Case based reasoning and artificial neural networkProcess and applications of CBRProcess and applications of ANNs

An intelligent supplier relationship management systemSupplier selection module (CSSM)Supplier benchmarking module (NNSBM)Construction of the key components of ISRMS

Application case study and resultsConclusionsAcknowledgementsReferences