DYNAMIC MULTI-AGENT BASED VARIETY FORMATION

AND STEERING IN MASS CUSTOMIZATION

Thorsten Blecker, Nizar Abdelkafi

Departement of Production/Operations Management, University of Klagenfurt, Austria

Gerold Kreutler, Gerhard Friedrich

Computer Science and Manufacturing, University of Klagenfurt, Austria

Keywords: Product Configuration, Mass Customization, V

ariety Formation and Steering, Multi Agent System

Abstract: Large product variety in mass customization involves a high internal complexity level inside a company’s

operations, as well

as a high external complexity level from a customer’s perspective. To cope with both

complexity problems, an information system based on agent technology is able to be identified as a suitable

solution approach. The mass customized products are assumed to be based on a modular architecture and

each module variant is associated with an autonomous rational agent. Agents have to compete with each

other in order to join coalitions representing salable product variants which suit real customers’ require-

ments. The negotiation process is based on a market mechanism supported by the target costing concept and

a Dutch auction. Furthermore, in order to integrate the multi-agent system in the existing information sys-

tem landscape of the mass customizer, a technical architecture is proposed and a scenario depicting the main

communication steps is specified.

1 INTRODUCTION

Mass customization is a business strategy that aims

at satisfying individual customers’ needs nearly with

mass production efficiency (Pine, 1993).The

development of mass customization is essentially

due to the advances realized in modular product ar-

chitectures and flexible manufacturing systems.

However, the progress in the fields of information

technologies and artificial intelligence for the sup-

port of Internet based customer-supplier interactions

can be considered as the most relevant enablers for a

successful implementation of the strategy. Rauten-

strauch et al. (2002) pointed out that information

systems provide the necessary support for enter-

prises pursuing mass customization.

The information which arises during the interac-

ion process between the customer and supplier

serves to build up a long-lasting individual customer

relationship (Piller, 2001). Due to high customer ori-

entation, mass customization induces a variety-rich

environment. However, customers generally do not

seek out variety per se. They do only want the

choice that fits to their needs.

The resulting variety in mass customization trig-

gers a

high complexity level that leads to additional

costs. Moreover, because of the limited human in-

formation processing capacity and lack of technical

product knowledge, excessive variety confuses cus-

tomers who are overwhelmed by the complexity of

the decision making process. Therefore, the main

goal should be to find an optimal product variety

which leads to the optimal cost-benefit-relation. For

example, Blecker et al. (2003) propose a key metrics

system to cope with the internal variety-induced

complexity and emphasize the importance of the in-

teraction systems to reduce the external complexity

experienced by customers during the buying process.

From this point of view, we can identify two

challenges. Fi

rstly, the mass customizer must be

supported by an information system to efficiently

cope with variety. Secondly, it is relevant to assist

customers with adequate information tools during

the individualization process in order to lead them in

a fast paced manner and with a low amount of effort

to their optimal choice. In this paper, after a short

description of the main variety problems in mass

customization, we formally define a multi-agent

based approach supporting dynamic variety-forma-

Blecker T., Abdelkaﬁ N., Kreutler G. and Friedrich G. (2004).

DYNAMIC MULTI-AGENT BASED VARIETY FORMATION AND STEERING IN MASS CUSTOMIZATION.

In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 3-13

DOI: 10.5220/0002608400030013

 SciTePress

tion and steering enabling mass customizers to face

both depicted challenges. Then, we suggest a techni-

cal infrastructure for the implementation of the

multi-agent system.

2 VARIETY PROBLEMS IN

MASS CUSTOMIZATION

Due to the high individualization level in mass

customization, final products are not manufactured

until a customer order arrives. This customer-pull

system improves the planning situation in dynamic

markets and avoids costs such as those due to final

products’ inventory and special offers to incur.

However, the huge variety induced in mass customi-

zation is associated with a high complexity and in-

volves costs which arise in the form of overheads.

Rosenberg (1997) mentions that product complexity

is essentially due to two main reasons which are (a)

the variety of product types and groups and (b) the

number of components being built in the products,

as well as their connections with each other.

An empirical study of Wildemann (2001) has

shown that with the doubling of the number of prod-

uct variants, the unit costs would increase about 20-

35% for firms with traditional manufacturing sys-

tems. For segmented and flexible automated plants

the unit costs would increase about 10-15%. Wilde-

mann concluded that an increase of product variety

is associated with an inverted learning curve. Fur-

thermore, Rathnow (1993) depicts a huge product

variety is not usually profitable and points out that

there is a point V

opt.

(optimal variety) from which the

cost effects of product variety overcompensate its

beneficial effects. Lingnau (1994) qualitatively ex-

amines cost effects which are involved when in-

creasing variety. He considers a functional organi-

zation structure and scrutinizes the effects of variety

on sales, production, purchasing and research and

development. Lingnau points out that variety gener-

ates additional costs in each function. For example,

when introducing new variants, new distribution

channels could be necessary. Increased variety also

complicates the production planning and control and

more setups leading to longer idle times in which are

required. With higher variety the work-in-process

inventory also increases and quality assurance

measures should be intensified.

The introduction or elimination of product vari-

ants are decisions which are made within the scope

of variety steering. Blecker et al. (2003) make the

distinction between variety management and variety

steering. Variety management embraces the con-

cepts that can be applied in order to increase compo-

nent and process commonality levels during a com-

pany’s operations such as part families, building

blocks, modular product architectures, etc. Unlike

variety management concepts, variety steering con-

cepts essentially deal with external variety, which

can be perceived by customers. In this paper, we as-

sume that the mass customizer has already imple-

mented a variety management concept and that the

main decisions concern variety steering.

The excess of variety and the resulting complex-

ity can endanger the success of mass customized

products whose prices should not dramatically differ

from the corresponding ones manufactured with a

mass production system. That is why, it is relevant

to efficiently cope with the internal effects of variety

in mass customization. In addition to high internal

complexity level during a company’s operations, va-

riety induces external complexity that has bad ef-

fects from a customer’s perspective.

Due to the limited human information processing

capacity, excessive variety could confuse customers.

Furthermore, customers are not aware of their needs

until they see them violated. By looking for suitable

products in huge assortments, customers can experi-

ence stress, frustration or regret. Iyengar/Lepper

(2000) also claim that in limited-choice contexts

people are engaged in rational optimization, whereas

in extensive-choice contexts people simply end their

choice-making when they find a choice that is

merely satisfactory, rather than optimal. Further-

more, Schwartz (2000) indicates that by adding new

options, the choice situation would be less rather

than more attractive and that some people would

look for the help of e.g. experts, who make the deci-

sion for them.

On the one hand, in order to avoid customers

getting lost in huge product assortments, the mass

customizer should support them during the interac-

tion process to help them find the product variants

corresponding to their optimal choice. On the other

hand, the mass customizer has to strongly consider

the internal complexity by efficiently steering vari-

ety. Therefore, a comprehensive solution approach

must integrate both customer’s and supplier’s per-

spectives in one information system.

3 A MULTI-AGENT APPROACH

FOR VARIETY FORMATION

AND STEERING

Common configuration systems for mass cus-

tomization necessitate specific product knowledge

and often overstrain customers. Therefore, we are

convinced that these systems should be improved to

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

better support customers during the elicitation proc-

ess. Blecker et al. (2004) opt for interaction systems

which are capable of assisting customers through

advisory. Thus, the interaction system should be able

to capture a customers’ preferences and profile in

order to display only the subset of relevant product

variants which would better fulfil customers’ re-

quirements. From the huge product assortment, only

the best variants succeed to be short-listed and dis-

played to customers. Consequently, in the long run

these will better contribute to a supplier’s success.

Those which are not short-listed will only trigger

high complexity and are not relevant for customers.

This would suggest that the product variants would

compete with each other. That is why, it is necessary

to define a mechanism setting the rules which or-

ganize the competition between variants. This leads

one to consider a market mechanism supported by

multi-agent technology. The complexity and fuzzi-

ness of the problem are further reasons for the use of

a multi-agent approach.

The multi-agent based system should dynami-

cally support each user during the interaction proc-

ess. This means that the system should iteratively

generate and refine product variants according to

specific customers’ needs. Concurrently, it supports

the long term supplier’s variety steering. This is re-

alized by the decentralization of variety decisions

which are supported by autonomous agents.

At first, we present the assumption and defini-

tions required to build up the multi-agent system.

Then, we conceptually describe how agents can

reach agreements in order to form product variants.

3.1 Assumption and Definitions

Pine (1993) pointed out that the best method to

achieve mass customization is to develop products

around modular architectures. Ericsson and Erixon

(1999) defined modules as building blocks with de-

fined interfaces. By combining only a few modules,

it is possible to construct a huge number of product

variants. The economies of scale are reached

through modules instead of products and economies

of scope are attained when modules are built in dif-

ferent products. That is why the assumption of this

paper is as follows:

Assumption: Modular product architecture

We assume that the complete set of product vari-

ants can be manufactured on the basis of modules.

The main idea is to consider the module variants

to be built in the different product variants as

autonomous rational agents. It is more reasonable to

consider the module variants as agents than the

product variants because with a few modules, one

can build up a very large number of product variants

which can go up to billions. Thus, by considering

modules as agents the problem remains manageable

and the computational resources are not over-

strained. Therefore, we provide the following defi-

nition:

Definition 1: Module agent

Let

be the set of all modules, M

{

}

MMMM , . . . , ,

. We call

a module class. A

module class

contains a set of module vari-

ants

)(21

, . . . , ,

iipii

MVMVMV

is a function associating

an index

of a module class with the index

referring to the number of module variants in a

module class. With each module variant

we associate an autonomous ra-

tional agent, called a module agent

which disposes of resources and is

able to perform tasks.

)(ip

)( , . . . ,1 , ipjMV =

)( , . . . ,1 , ipjMA

Modules can be classified in must- and can-

modules. Must-modules are indispensable for en-

suring the basic product functionalities, whereas

can- modules are optional. For example, an engine is

a must-module for a car. Without an engine a car

cannot ensure its basic functionality consisting of

mobility. In contrast to the engine, an air-conditioner

is a can-module because it does not disturb the main

functionalities a car must perform. In this context the

term platform is defined in the technical literature in

two distinctive views:

• A product platform is the set of all modules re-

quired for the manufacturing of all possible

product variants (e.g. Ericsson/Erixon, 1999).

• A product platform can also be the appellation of

a specific common module which is used in a

great range of product variants (e.g. Piller/Warin-

ger, 1999; Wildemann, 2003). This definition is

commonly used in the automobile industry.

The second definition of platforms will be

adopted in this paper because it considers the plat-

form as a module having an additional relevance in

comparison to other modules, which is mainly due to

its implementation frequency in product variants.

The corresponding agents are called platform agents

to make the distinction vis-à-vis other module

agents. Furthermore, the set of all platform and

module agents are grouped in an agent pool. All dif-

ferent agents are members of a multi-agent system

whose main goal is to present only a subset of prod-

uct variants, which would best fit customers’ needs.

Because only a subset is allowed to be displayed

to customers, the product variants have to compete

with each other. Due to the modular architecture of

products, we can argue that the module variants also

DYNAMIC MULTI-AGENT BASED VARIETY FORMATION AND STEERING IN MASS CUSTOMIZATION

compete to be existent in the set of the presented

product configurations. Being driven by a further

motivation of this work to support variety steering

decisions in mass customization, the module variants

which do not resist competition should be elimi-

nated. Therefore, it is legitimate to provide the sec-

ond definition:

Definition 2: Self-preservation

Each module agent

strives for ensuring its

existence by having enough resources to survive.

Definition 2 leads us to consider evolutionary

theory which sees evolution as the result of selection

by the environment acting on a population of organ-

isms competing for resources. The winners of the

competition, those who are most fit to gain the re-

sources necessary for survival, will be selected, the

others are eliminated (Kauffman, 1993). The re-

sources of an agent are stored in an account which is

defined as follows:

Definition 3: Module agent’s account

Each module agent

has revenues and ex-

penses that are summed up in an account

monetary units.

constantly diminishes in the

course of time.

Acc

It is relevant to mention that the monetary units

that a module agent has on its account do not relate

to the prices customers pay. The account only serves

as an internal steering mechanism for a multi-agent

system. The account surplus rather refers to the ac-

tual resources of an agent

. From definitions 2

and 3, we can conclude that each agent endeavors to

maximize its account surplus. A surplus of zero will

mean that the module agent risks death leading to

the elimination of the module variant. Furthermore,

the second part of definition 3 means that the agent’s

resources diminish in the course of time even if the

agent does not carry out any task. To explain what a

task is, we provide the following definition:

Definition 4: Module agent’s task

The task

of module agent is to form product

variants by joining coalitions

, . . . ,1 , nkC

The allocation of tasks to groups of agents is

necessary when tasks cannot be performed by a sin-

gle agent. The module agents on their own are not

able to provide a solution. They need to cooperate in

order to fulfill tasks. However, the autonomy princi-

ple of agents is preserved because each agent can

decide whether to take part or not in a product vari-

ant. By forming coalitions each agent strives for its

personal utility/account via cooperation. Module

agents follow the economic principle of rationality

and attempt to form a coalition which will maximize

their own utilities. Furthermore, because of the het-

erogeneity of customer requirements, module agents

may have different efficiencies in task performance

due to their different capabilities.

In order to participate in a coalition, the module

agent has to pay a certain fee. It is noteworthy that

as opposed to other work in multi-agent systems

(e.g. Shehory and Kraus, 1995), one agent may par-

ticipate in more than one coalition. Moreover, these

coalitions are dynamic and take place in real-time,

after capturing customers’ preferences. However, a

coalition may succeed or fail. This primarily de-

pends on the coalition’s result, which can be com-

plete or incomplete:

Definition 5: Complete vs. incomplete

coalitions

We say that a coalition is complete if the coali-

tion formed by the module agents builds up a salable

product variant. A coalition is incomplete if the coa-

lition formed by the module agents does not build up

a salable product variant.

Note that an agent will join a coalition only if the

payoff it will receive in the coalition is greater than,

or at least equal to, what it can obtain by staying

outside the coalition (Shehory and Kraus, 1995). In

our case a module agent that does not participate in

any coalition has a payoff of zero. Because the ac-

count surplus of an agent diminishes in the course of

time, each agent should be interested in participating

in beneficial coalitions to be able to reconstruct its

resources and thus better ensuring its existence.

However, there is no certainty about the success of

coalition results. As aforementioned, each agent has

to pay a certain fee in order to be allowed to partici-

pate in a coalition. But if the coalition that subse-

quently forms is incomplete or fails because it is not

powerful enough to be displayed to customers, then

the participation of an agent in a coalition is a waste

of resources. Therefore, each agent has to be capable

of estimating in advance the likelihood of the suc-

cess of the coalitions it joins. However, the module

agents should remain as simple as possible and

should not become very complex. The whole multi-

agent system has to contribute to problem solving

and one agent should only dispose of the intelligence

it requires in order to not waste computational re-

sources.

3.2 The Negotiation Process

Firstly, it is relevant to determine the mechanism

initiating a coalition. This being in reference to deci-

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

sions about (a) which module agents are able to be-

gin the formation of coalitions and (b) which reach-

ing agreement process should be implemented to co-

ordinate the module agents. We agree that platform

agents are most suitable for initiating coalitions. We

also assume that these agents dispose of an infinite

account surplus. Therefore, they do not have to care

about their existence. This is a legitimate assumption

because the development of product platforms is

generally cost-intensive. The development process

itself may last for a duration of many years. Plat-

forms are also created to be the basic module of a

wide range of product variants for a long period of

time. For example, by canceling a platform in the

automobile industry, this would mean to cancel all

models and the corresponding variants which are

supported by this platform. Thus, such a decision is

strategic and should not be allocated to automated

software agents. As in each decision in variety

steering it should be supported by human agents

who have the required competencies and informa-

tion. However, it is conceivable that each platform

agent strives for being successful as much as possi-

ble, e.g. by contributing to the most sales’ volumes.

On the basis of customers’ preferences, the type

and the number of product platforms to form coali-

tions are determined. Note that:

• a platform agent can be selected more than once,

• each product variant is based on one platform,

• each platform can be found in several product

variants and,

• the total number of the selected platform agents

is also the utmost limit of the product variants

which can be formed by coalitions.

The coalitions take place at a certain point in

time and form in order to fulfill the needs of one

customer. When all resulting coalitions are com-

plete, then the number of product variants will be

exactly equal to the number of selected platform

agents provided that no identical coalitions form.

The platform agents have the ability to steer the

formation of coalitions by (a) fixing the set of mod-

ule agents which could contribute to the fulfillment

of customers’ requirements and by (b) determining

the mechanism according to which module agents

can join coalitions.

We propose to base the coalition formation

mechanism on the target costing concept and a

Dutch auction. Target costing is based on the price

the customer is willing to pay. Starting from this

price, it is possible to determine the utmost limit of

the costs of each product function that is allowed to

incur by taking into account the contribution of each

function to the fulfillment of customer requirements.

Further on, because each product component or

module makes a certain contribution to the realiza-

tion of a product function it is possible to distribute

the function costs on the modules respectively com-

ponents (Seidenschwarz, 2001). Thus, the result of

target costing is an utmost limit for modules’ or

components’ costs. The platform agents which are

the auctioneers use these costs to initiate a Dutch

auction. The module agents which compete to join

the coalitions are the bidders. A Dutch auction is le-

gitimate in this case because each agent tends to de-

lay as much as possible in joining a sub-coalition in

order to (a) better evaluate whether to participate or

not in a hitherto sub-coalition and (b) minimize as

much as possible the fees to pay. But due to the

product configuration constraints, when a module

agent wins the bid, it may impose constraints on the

other bidding agents which intend to take part of the

sub-coalition. Thus, the intelligence of the module

agent should enable it to proficiently estimate when

and for which coalition it bids. These auctions will

continue until all coalitions are formed.

Although platform agents have an infinite ac-

count, we assume that they will also try to maximize

the revenues they receive from module agents. We

will describe in the next section how the product

variants resulting from the coalitions are filtered af-

ter their formation. Only the set of product variants

that are displayed to customers will receive revenues

which are distributed on modules. The total col-

lected monetary units from all module agents are

collected in an account and then distributed on the

module agents participating in the few successful

product variants by considering their contribution in

the fulfillment of the product functions and their

participation level.

Up to now, we have only described what module

and platform agents should perform and how the

whole multi-agent system can reach agreements in

order to form coalitions. But we have not mentioned

what are the abilities an agent should have, to effec-

tively carry out its tasks. Module and platform

agents have different tasks. Therefore, they have dif-

ferent abilities. Module agents strive for maximizing

their utilities (accounts). That is why, they have to

develop strategies in order to survive. Subsequently,

they should be able to evaluate in advance the suc-

cess of the coalitions by estimating the probability

that the formed product variants can be displayed to

customers. Furthermore, they have to know when to

bid and which coalition would be beneficial to join.

Generally, as intelligent agents module agents have

to update their knowledge from their own experience

and the behavior of the other module agents per-

taining to the multi-agent environment, which means

that they have to learn.

In opposition to module agents, platform agents

do not care about their existence due to their infinite

account surplus. Furthermore, they decide which

module agents are allowed to participate in the coa-

DYNAMIC MULTI-AGENT BASED VARIETY FORMATION AND STEERING IN MASS CUSTOMIZATION

litions. Therefore, they are more powerful than

module agents. Platform agents initiate and coordi-

nate the formation of coalitions. They are also capa-

ble of communicating with each other to avoid the

formation of identical coalitions. Platform agents

have the overview of the coalition while forming

and can forbid the further bidding of module agents

by considering the constraints imposed by module

agents which have already joined the coalition.

In the following we concentrate on module

agents. We assume that the product platforms are

capable of initiating the Dutch auction and that only

product constraints may restrain the participation of

module agents in coalitions. In order to represent the

module agents, we use a mathematical tool from de-

cision theory. Decision theory defines a rational

agent as one that maximizes the expected utility. The

expected utility

of an action is defined as (Rus-

sel/Norvig, 1995):

()

αωωα

)( PUEU

∑

Ω∈

{}

actions of eperformanc the

given outcomes possibleover on distributiy probabilit a :

real a

with outcomean gassociatinfunction utility a :

outcomes possible all ofset the:,,

agent,an

toavailable actions possible all ofset the: ;

IRU

AcAc

where

→Ω

′

=Ω

∈

ωω

Let the function

opt

take as input a set of possi-

ble actions, a set of outcomes, a probability distribu-

tion and a utility function and let this function return

an action. The defined behavior of

is defined as

follows (Russel/Norvig, 1995):

opt

() ()

()

αωω

max arg ,, ,

PUUPAcf

opt

∑

Ω∈

∈

=Ω

Wooldridge (2000) criticizes

opt

for building

rational agents because

opt

requires an uncon-

strained search which can be very expensive when

the space of all actions and their outcomes is very

wide. But, in our case this critic does not seem to be

strong enough because the action types that a mod-

ule agent can perform are (a) to participate in a coa-

lition (Participating is the action

) or (b) not to

participate (Not participating is the action

Thus, the number of action types a module agent

disposes of are only two. Furthermore, the outcome

of actions may be that either (a) the module agent is

a member of a product variant which is selected in

the final subset (Success of a coalition is the out-

come

) or (b) the module agent is a member of

a product variant which is not selected in the subset

to be presented to customers (Failure of a coalition is

the outcome

). That is why, we argue that it is

legitimate to consider

opt

to build module agents.

However,

opt

should be adapted to the requirements

of the multi-agent system problem that was pre-

sented above.

Suppose at a point in time t=0 the platform

agents initiate coalitions. Each platform agent

chooses which module agents are allowed to bid. For

each module agent, the platform agent communi-

cates a function to a module class i having the fol-

lowing form:

)()( tgKtK

iii

[]]]

timebidding maximal the:

and,function decreasingsteadily a :

;1)0( /1,0,0:

auction,Dutch

theof pricefirst thengrepresenticonstant a :

time,of course the

in decreasingfunction auction Dutch the:)(

gTg

where

=→

As aforementioned, when a module agent joins a

coalition, it may restrain the participation of other

module agents also intending to join the coalition.

Therefore, each module agent must be capable of

evaluating the behavior of the other agents that

could prohibit its participation. This behavior should

be captured by the function

which is a risk

function of a module agent

Risk

[

]

[

]

(

)(

RTRiskRiskTRisk ==→ and 00/1,0,0:

)

[]

agent module theof

ngnessrisk willi thereflectingconstant a is 1,0

function increaingsteadily a is

∈R

Risk

where

Note that the function

is a function that

leads the module agent to bid as early as possible in

order to increase its chances in being a participant of

a coalition. Let

be the function which

takes the value 0 when the agent is not a member of

the coalitions representing the product variants dis-

played to customers and the value

when the

product variants are displayed to customers.

Risk

venueRe

vRe

The utility function U of a module agent depends

on the risk function which is supposed to decrease

revenue during the auction process, the revenue

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

function and the Dutch auction function. The

adapted

for our case is:

opt

=Ω ),,,Re,,,( UPgvenueRiskAcf

iopt

[]

{}

()

{}

∑

∈

−−

1,0

)(Re)(1maxarg

αω

PtgKvtRisk

The adapted

opt

returns the point in time t at

which the module agent has to bid for the Dutch

auction to maximize its utility. But note that if

then

[

Tt ,0∈

]

and if there is no

[

]

Tt ,0

∈

maximizing the utility (

then )Tt >

and the

module agent intends on not participating in the

coalition.

Furthermore, suppose that a module agent

is allowed to participate in p coalitions:

. For each coalition the module

agent estimates

opt

and at the point in time t=0

where an auction begins, the module agent has to

develop a plan

{

pkC

,,1 , K∈

}

()

(

)

(

)

(

)

(

)

ppkk

tttPlan

00000

,,,,,,

ααα

KK=

which indicates whether or not and when to bid for

each coalition. For notation purposes, when

, the module agent allocates to

an infinite

value

(

)

∞=

, which is in accordance with the fact

that an agent will never bid and then not to partici-

pate. Moreover, by developing a plan the module

agent has to consider its account constraint. It is not

allowed to pay for coalitions more than the account

surplus. This means that

surplusAccountfees

∑

≤

. It

is also conceivable that the module agent wants to

allocate only a certain sum of monetary units for the

coalitions which should be formed to be presented to

one customer. This depends on the self-preservation

strategy the module wants to pursue.

However, the agent plan determined at

not fixed for the whole auction process. The module

agent has to adapt this plan according to the changes

which could occur in its environment. Suppose that

the tuples of

(

)

are arranged so that

1 pk

. Suppose that

and that at a point in time

an agent from the

same class wins the bid or an agent from another

class imposes participation constraints. At this point

in time, the module agent has to estimate once again

opt

for the remaining coalitions to determine

whether and when to bid. This is legitimate because

when the participation in one coalition fails the

module agent can allocate the resources he has

planed to expend differently. The resulting plan at a

point in time

Plan

ttt ≤≤≤≤

tand

000

∞≠∞≠

tt <

is:

(

)

(

)

(

)

(

)

(

)

ppkk

tttPlan

11111

,,,,,,

ααα

KK=

The application of the described process will

continue until different coalitions are formed.

Recapitulating, we can say that the main advan-

tages of the developed multi-agent approach are:

•

the easy maintenance of the system: when intro-

ducing new module variants or eliminating old

ones, it is sufficient to introduce or eliminate

module agents,

•

the dynamic variety generation during the inter-

action process and variety steering as well as,

•

the application of a market mechanism concept

which lets the intelligent agents themselves de-

cide according to the current situation about

their suitability to fulfill real customers’ re-

quirements. Such a market mechanism based

approach enables us to efficiently carry out the

coordination mechanism, even for a high num-

ber of involved agents (Shehory et al., 1998).

DYNAMIC MULTI-AGENT BASED VARIETY FORMATION AND STEERING IN MASS CUSTOMIZATION

4 TECHNICAL ARCHITECTURE

In this section we present a complete model for

variety formation and steering based on the multi-

agent system approach developed in the previous

section. We propose to interface the module agents’

pool to a customer advisory system to support dy-

namic variety formation during the real time cus-

tomer advisory process.

Advisory systems are software systems that

guide customers according to their profile and re-

quirements through a „personal”, customer oriented

advisory process to elicit their real needs from an

objective point of view (Blecker et al., 2004). Dur-

ing the advisory process, the system suggests the

customer product variants according to his profile

and refines them through the dialog. At the end of

the advisory process, the customer is supported with

product variants which fulfill his real needs.

At each advisory session the multi-agent system

dynamically creates coalitions of product variants

that can be recommended to the user. Therefore, we

aim at integrating the system into the existing infor-

mation system landscape. Figure 1 depicts the ar-

chitecture of such a system.

Beside the agents’ pool the architecture consists

of the following main elements:

•

an online interface to the data of the advisory

system that provides a customer’s preferences,

•

an interface to the supplier’s back office which

for instance comprises a CRM or OLAP system,

•

additional filtering and target costing data

sources,

•

librarian agents that have access to all back of-

fice systems and make proper data available for

the other components,

•

coordination agents that coordinate the variety

formation in the agents’ pool and,

•

a blackboard that supports communication be-

tween the module agents’ pool and its environ-

ment.

The system also supports variety steering. As

was mentioned in the previous section, the account

balance of the agents provides a measurement of the

success of a module variant which constantly can be

analyzed by a variety steering agent or humans.

Before we technically describe the system, we

will describe the main idea on the basis of a sce-

nario: During the advisory process the system cap-

tures the customer’s requirements according to his

knowledge level. During the advisory dialog the

system presents the user a proposal of several prod-

uct variants according to his profile and preferences.

These are refined through the advisory dialog which

leads to dynamically refined suggestions for product

variants. Finally, the system generates suggestions

of product variants that meet real customer needs.

The creation of a valid subset of product variant

coalitions is dynamically carried out by following

steps:

Module Agents' Pool

Platform

Agents

Must-

Module

Agents

Can-

Module

Agents

Coord

CRM-Datasource

OLAP-Datasource

Customer Interests-

Datasource

Target Costing-

Datasource

Lib

Valid.

Agent

Presented Subset

Var.

St.

Advisory

Online Data

Bla ckboard

Pro d. Mode l

Conf. Orders

Variety Steering

Filtering Component

Human Agent

Intelligent Agent

Customer Configuration Order

Data Exchange/Communication

External Datasource

(e.g. Webservice)

Agents:

MA (Module Agent)

PA (Platform Agent)

Coord. (Coordination Agent)

Lib. (Librarian Agent)

V.St. (Variety Steering Agent)

Valid. (Validation Agent)

Figure 1: Technical architecture for an agent based variety formation and steering

(1)

A so-called librarian agent obtains data from the

online advisory data source. These data can con-

tain both user data and product preferences – de-

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

pending on the knowledge level of the user. If

the user is familiar to the domain, he can make a

decision on the product level; otherwise the sys-

tem gathers his needs, which can be captured in

e.g. a language different form product specifica-

tion. For instance, data can contain personal data

such as the customer’s age or marital status, his

personal preferences or desired product attrib-

utes. In the automotive domain it could be about

a male customer with two children who is sporty,

but prefers an economical car.

(2)

The information about the customer is supple-

mented by the librarian agent: Depending on

whether the customer is recognized (e.g. by a

CRM data source where both the customer’s in-

terests and his past buying behavior are stored.

Otherwise the information can be provided from

the OLAP data source where traditional web

mining techniques such as clustering are used to

extend the customer’s profile. The result of this

process is an improved profile of the customer’s

needs.

(3)

In order to support the negotiation process in the

module agents’ pool, the librarian agent calls for

service from the filtering component in order to

convert the customer attributes to product attrib-

utes. For instance, this can be based on expert

rules or statistical methods. As an example the

attributes of (1) could be inferred that the car

should be a sedan with no automatic transmis-

sion that runs on diesel.

(4)

The target costing component is used to estimate

the costs of the product’s functions that the cus-

tomer probably will have to pay for. For in-

stance, this could be based on past sales of a

clustered customer group.

(5)

The data is passed on to the coordination agent

who monitors the load of the module agents’

pool. If the number of customer configuration

orders is below a certain limit, the coordination

agent sets forth a request for new product vari-

ants with the desired product properties onto the

blackboard. Note that these product properties

derived from the customer attributes only support

the negotiation process, they are not constraints.

Besides for that, the coordination agent selects

both appropriate platform agents who should

carry out the auction, and the number of coali-

tions of product variants they should form. This

decision is based on the customer’s profile and

the product model.

(6)

Now the negotiation process is carried out as de-

scribed in the previous chapter, until all coali-

tions are formed. The resulting coalitions are put

back onto the blackboard where they are re-

moved from the coordination agents and passed

to the validation agent.

(7)

The validation agent requests data from informa-

tion agents in order choose a subset of the avail-

able coalitions of variants to present them to the

customer. This task is performed on the basis of

validation which is a kind of „reverse mapping”.

That means that the properties of the selected

coalitions are mapped to customer attributes as a

kind of verification. The best coalitions are pre-

sented and rewarded with monetary revenue for

the accounts.

(8)

If the customer makes the decision to buy a cer-

tain product variant of the presented subset, it is

conceivable that an additional reward would be

sent back to the accounts of the corresponding

module agents.

(9)

Additionally, we propose the use of the account

level of each module agent as an indicator to

support variety steering decisions. In an inde-

pendent process the system makes suggestions to

eliminate module variants. If the account level is

negative or low in comparison with competing

module variants, this is an indicator to remove

the module variant. Furthermore, the introduc-

tion of new module variants can affect the inter-

nal and external complexities which can be esti-

mated by computing suitable key metrics

(Blecker et al., 2003).

In conclusion we can see that the complexity is

spread throughout all system components in the

multi agent system:

•

The module agents’ pool is responsible for carry-

ing out a negotiation for forming product variant

coalitions,

•

coordination agents manage the blackboard and

the general interface between the agents’ pool

and its environment,

•

librarian agents not only interface the back office

systems, they independently obtain data and

process them in an intelligent way to support the

other agents optimally,

•

a validation agent carries out the validation of the

module variants coalition independently of the

decisions of the other system components.

For the implementation of such, we propose to

base the system on Java technology. This not only

ensures platform independence, it also provides a

uniform framework for the implementation of all

components.

All back office systems such as CRM, OLAP or

other data sources must be connected via custom

interfaces, for example by XML. On the variety

formation system, data can be provided by web ser-

vices so that the agents can access the services. The

agents’ pool can be realized in one virtual machine.

Due to the decision to represent module variants in-

DYNAMIC MULTI-AGENT BASED VARIETY FORMATION AND STEERING IN MASS CUSTOMIZATION

stead of product variants as agents, this assumption

is admissible. This way we can lower the communi-

cation costs because this enables a direct interaction

between the agent instances. The coordination be-

tween the agents’ pool and the external agents is car-

ried out via a blackboard where all agents are regis-

tered. Coordination agents, validation agents and li-

brarian agents can be distributed for reasons of load

balancing. Communication between these agents can

be performed via Java’s RMI (Remote Method In-

vocation) or CORBA to support other systems.

5 CONCLUSIONS

In this paper, we have depicted the main problems

which are triggered by increasing variety in mass

customization. Variety involves an internal

complexity inside a company’s operations, as well

as an external complexity from a customer’s per-

spective. To mitigate both complexities’ problems,

the main idea is to provide an information system

solution which is capable of both supporting cus-

tomers during the interaction process by proposing

and refining product variants and simultaneously

supporting variety steering decisions. The agent

technology is able to be identified as a suitable ap-

proach to cope with this problem in a decentralized,

self-coordinating way.

The developed system integrates both customer’s

and supplier’s perspectives in one information sys-

tem. We outlined how module variants can be repre-

sented as intelligent agents that negotiate with each

other to ensure their survival within the scope of va-

riety steering. Based on the decision theory’s model

for rational agents, we formally define the function

that an agent strives to optimize. The negotiation

process between the intelligent agents is based on

the target costing concept and a Dutch auction. This

is also described in a formal way defining the possi-

ble functions which have to be determined. Because

we intend to carry out simulations of the entire sys-

tem, several functions which determine the intelli-

gence of the defined agents should be tested. Based

on these simulations we will decide which imple-

mentation will lead to a working prototype. Fur-

thermore, a technical architecture for the

agent-

based variety formation and steering in mass cus-

tomization is proposed.

The main advantages of the developed approach

are the easy maintenance of the system, the dynamic

variety generation and variety steering, as well as the

application of a market mechanism concept sup-

ported by agent technology. The adopted market

mechanism presents a relevant approach enabling

one to overcome the shortcomings of existing inter-

action systems and variety steering methods. Thus,

instead of building rigid rules in the interaction sys-

tem that map customer requirements into product

attributes, the proposed market mechanism approach

lets the intelligent agents themselves decide accord-

ing to the current situation about their suitability to

fulfill real customers’ requirements. Furthermore,

the market mechanism enables us to connect two

relevant concepts in mass customization, namely

which product variants should be retained in the

product assortment and which specific ones from

this assortment should be selected and offered to a

particular customer.

REFERENCES

Blecker, T., Abdelkafi, N., Kaluza, B., and Friedrich, G.,

2003. Key Metrics System for Variety Steering in

Mass Customization. In MCPC’03, 2

Interdiscipli-

nary World Congress on Mass Customization and

Personalization, Munich, October 6-8, 2003.

Blecker, T., Abdelkafi, N., Kreutler, G., and Friedrich, G.,

2004. An Advisory System for Customers’ Objective

Needs Elicitation in Mass Customization. In 4th

Workshop on Information Systems for Mass Customi-

zation, Madeira Island, February 29-March 3, 2004.

Ericsson, A., and Erixon, G., 1999. Controlling Design

Variants: Modular Product Platforms, Society of

Manufacturing Engineers. Dearborn, Michigan.

Iyenger, S. S., and Lepper, M. R., 2000. When Choice is

Demotivating: Can One Desire Too Much of a Good

Thing?, URL: http://www.columbia.edu/~ss957/

publications.html (Retrieval 10. Oct. 2003).

Kauffman, S. A., 1993. The Origins of Order: Self-Or-

ganization and Selection in Evolution, Oxford Univer-

sity Press, New York.

Lingnau, V., 1994. Variantenmanagement: Produktions-

planung im Rahmen einer Produktdifferenzierungs-

strategie, Erich Schmidt Verlag GmbH & Co. Berlin.

Piller, F. T., 2001. Mass Customization: Ein wettbewerbs-

strategisches Konzept im Informationszeitalter, Gabler

Verlag. Wiesbaden, 2nd edition.

Piller, F. T., and Waringer, D., 1999. Modularisierung in

der Automobilindustrie – neue Formen und Prinzipien,

Shaker Verlag. Aachen.

Pine II, J., 1993. Mass Customization: The New Frontier

in Business Competition, Harvard Business School

Press. Boston.

Rathnow, P. J., 1993. Integriertes Variantenmanagement:

Bestimmung, Realisierung und Sicherung der optima-

len Produktvielfalt, Vandenhoeck und Ruprecht.

Goettingen.

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

Rautenstrauch, C., Taggermann, H., and Turowski, K.,

2002. Manufacturing Planning and Control Content

Management in Virtual Enterprises Pursuing Mass

Customization. In Rautenstrauch, C., Seelmann-Egge-

bert, R., and Turowski, K. (Eds.). Moving into Mass

Customization: Information Systems and Management

Principles, Springer-Verlag. Berlin Heidelberg, pp.

103-118.

Rosenberg, O., 1997. Kostensenkung durch Komplexi-

taetsmanagement. In Franz, K.-P., and Kajueter, P.

(Eds.). Kostenmanagement: Wertsteigerung durch sy-

stematische Kostensenkung, Schaeffer Poeschel Ver-

lag. Stuttgart, pp. 225-245.

Russel, S., and Norvig, P., 1995. Artificial Intelligence: A

Modern Approach, Prentice-Hall, Inc. New Jersey.

Schwartz, B., 2000. Self-Determination: The Tyranny of

Freedom, URL: http://www.swarthmore.edu/SocSci/

bschwar1/self-determination.pdf

Seidenschwarz, W., 2001. Target Costing. Marktorien-

tiertes Zielkostenmanagement, Vahlen. Muenchen, 2

edition.

Shehory, O., and Kraus, S., 1995. Coalition formation

among autonomous agents: Strategies and complexity.

In Castelfranchi, C., and Muller, J. P., (Eds.).

From

Reaction to Cognition

, Springer Verlag. pp. 57-72.

Shehory, O., Sycara, K.P., and Jha, S., 1998. Multi-agent

coordination through coalition formation. In Singh,

M.P., Rao, A.S., and Wooldridge, M.J., (Eds.). Pro-

ceedings of ATAL’97, Springer Verlag. pp 143-154.

Wildemann, H., 2001.

Das Just-In-Time Konzept – Pro-

duktion und Zulieferung auf Abruf, TCW

Transfer-

Centrum GmbH.

Muenchen, 5 edition.

Wildemann, H., 2003. Produktordnungssysteme: Leitfaden

zur Standardisierung und Individualisierung des Pro-

duktprogramms durch intelligente Plattformstrategien,

TCW Transfer-Centrum GmbH. Muenchen, 2

edi-

tion.

Wooldrige, M., 2000. Reasoning about rational agents,

The MIT Press. Cambridge et al.

DYNAMIC MULTI-AGENT BASED VARIETY FORMATION AND STEERING IN MASS CUSTOMIZATION