A Self-organizing Multi-Agent System for Combining Method Fragments

elie Bonjean, Marie-Pierre Gleizes, Christine Maurel and Fr

eric Migeon

Institut de Recherche en Informatique de Toulouse (IRIT), Universit

e de Toulouse, Toulouse, France

Keywords:

Method Fragments, Adaptive Multi-Agent System, Method Process.

Abstract:

Software systems are becoming more and more complex. A common dilemma faced by software engineers

in building complex systems is the lack of method adaptability. However, existing agent-based methodologies

and tools are developed for particular system and are not tailored for new problems. This paper proposes an

architecture of a new tool based on SME for self-constructing customized method processes. Our approach is

based on two pillars: the process fragment and the MAS meta-model. These two elements are both deﬁned

and considered under a speciﬁc agent-oriented perspective thus creating a peculiar approach. Our work is

based on the self-organization of agents, making it especially suited to deal with highly dynamic systems such

as the design of an interactive and adaptive software engineering process.

1 INTRODUCTION

The MAS community has been proliﬁc to deﬁne

software engineering methods (Bergenti et al., 2004;

Henderson-Sellers and Giorgini, 2005) in order to

guide designers with respect to the wide range of

MAS properties. Facing the numerous methods, a

development team needs help to execute the relevant

process according to the development context which

is deﬁned by the system under study characteristics

as well as the team capabilities and preferences. Our

goal is to provide new tools for designing complex

systems where the method must be adapted to the de-

velopment context.

Coming from Situational Method Engineering re-

search (Brinkkemper et al., 1998; Henderson-Sellers

and Ralyt

e, 2010), the aim of decomposing processes

into pieces is to adapt the process to the characteristics

of the business problem and to the level of expertise of

engineer teams (Ralyt

e, 2004). A process can then be

deﬁned by assembling the pieces of methods, called

fragments, in order to suit the context (the situation)

changes. The Agent-Oriented Software Engineer-

ing (AOSE) community contributed to this research

splitting up methods into fragments and providing

precise descriptions of them (INGENIAS

, PASSI

ADELFE (Bernon et al., 2005), TROPOS

...).

We present in section 2 the aim of the SCoRe

http://ingenias.sourceforge.net/

http://www.pa.icar.cnr.it/passi/Passi/PassiIndex.html

http://www.troposproject.org/

model (Self-Combined method fRagments) that we

propose and its architecture in section 3 to automat-

ically build a self-adaptive design process where each

fragment is encapsulated in an autonomous agent.

This approach relies on the self-organization of its

agents, making it especially suited to deal with highly

dynamic systems such as the design of an interactive

and adaptive Software Engineering Process (SEP).

Section 4 brieﬂy presents tests and explains the results

obtained. Finally, section 5 describes related works

before concluding.

2 REQUIREMENTS AND

CHARACTERISTICS OF SCoRe

The contribution of the work lies in the self-

adaptive multi-agent system implementation for self-

composition and self-organization of method frag-

ments. This section presents the challenges.

2.1 Adaptation

While the demand for speciﬁc, complex and varied

system continues to grow, current methods in the

MAS domain remain limited and sometimes not well

adapted e.g. (Cossentino et al., 2008). The need for

well-deﬁned guidelines that will make the develop-

ment process more efﬁcient and more effective has

become crucial. Currently, there is no single method-

ology that can be uniquely pointed as ”the best”. Until

288

Bonjean N., Gleizes M., Maurel C. and Migeon F..

A Self-organizing Multi-Agent System for Combining Method Fragments.

DOI: 10.5220/0004261402880293

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 288-293

ISBN: 978-989-8565-38-9

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

now method adjustments to the speciﬁc requirements

and constraints are mixed in ”local” adaptations and

modiﬁcations of existing one. In order to succeed

in creating good situational methods, i.e., methods

that best ﬁt given situations, fragment representation

and cataloguing are very important activities. In par-

ticular, the fragments (sometimes addressed as pro-

cess fragments, method fragments or chunks) have

to be represented in an uniform way that includes all

the necessary information that may inﬂuence their re-

trieval and assembling.

2.2 Fragment Standardisation

Method fragments are ﬁrst identiﬁed by examining

existing methods. These method fragments are made

according to templates deﬁned by repository design-

ers and standardisation working groups

. Therefore

the choice of fragments granularity relies on design-

ers. According to the Rational Uniﬁed Process, meth-

ods are deﬁned following different levels of granular-

ity (phase, activity and step) which is an important

factor. The ”step” level involves a speciﬁc and ﬁd-

dly task but also requires perfect knowledge of meth-

ods and long work. This fragmentation is very ﬁne-

grained and provides a greater number of fragments.

This low level of granularity is useless and inaccurate

when the steps are related and interdependent. On

the other hand, the ”phase” level of granularity could

form huge complete fragments. The coarse-grained

granularity promotes the redundancy issue. The du-

plication of activities or steps may occur with high

granularity. By consequences, an activity or step may

be included in different fragments. The risk that this

happens grows up with the level of granularity. In ad-

dition, the assembling possibilities are therefore min-

imized.

2.3 Complexity

Currently, ten AOSE methods are fragmented, each

one composed of approximately twenty fragments.

Such fragments constitute the root constructs of the

method itself and they have been extracted by consid-

ering a precise granularity criterion: each group of ac-

tivities (composing the fragment) should signiﬁcantly

contribute to the production/reﬁnement of one of the

main artefacts of the method (for instance, a diagram

or a set of diagrams of the same type). Following this

assumption, fragments obtained from different meth-

ods are based on a similar level of granularity.

Besides, to design a process manually means

studying for the compatibility of each fragment with

http://www.pa.icar.cnr.it/cossentino/ﬁpa-dpdf-wg/

the others i.e. approximately twenty thousand possi-

ble combinations. Although this number can be de-

creased by the knowledge and the know-how of pro-

cess engineers, the work remains long and irksome.

It is why we propose to design the system SCoRe to

self-combine and self-organize fragments.

2.4 User Requirements

In our approach, a new complete process is self-

designed contingent on a situation. A complete pro-

cess enables engineers to visualise all activities and to

have a whole view of the process. SCoRe focuses on

adaptation during process execution. At every step,

the development team is advised by SCoRe on its next

fragment choice according to the running features. If

the features evolve, this advice may therefore differ

from the following fragment initially suggested.

The studied solution is based on the fragments

agentiﬁcation in order to self-design an adaptive pro-

cess which can deal with the complexity mainly due

to the huge number of fragments. Indeed, a com-

plex system cannot currently be designed without bug

caused by designers. Assisting the designer during

the system utilization would reduce the number of

bugs and make the system most suitable to the cur-

rent situation. The adaptation is therefore required.

As for components assembling, the fragments com-

bination needs features. In our approach, they corre-

spond to MAS Metamodel Elements (MMME). Two

fragments are therefore combined if one produces the

MMME required by another one.

3 ARCHITECTURE OF SCoRe

The general structure of the Self-Combined method

fRagments (SCoRe) proposed is described in this sec-

tion, before detailing the behaviors and the interac-

tions of the agents composing it.

3.1 SCoRe Components

We consider a method process as a set of assem-

bled method fragments which are linked through their

own required or produced MMMEs. Establishing a

method process consists in combining some of the

fragments taking into account the user-deﬁned objec-

tives and knowledge.

The main goal of SCoRe is to suggest a tailored

process. For that, SCoRe learns the context to ap-

ply on fragments, in order to sustain this evolution.

SCoRe acts without relying on a model of the pro-

cesses,i.e. it is only able to take into account the de-

ASelf-organizingMulti-AgentSystemforCombiningMethodFragments

289

signers’ knowledge and objectives, and to observe the

evolution of the running process on which MMMEs

are available, in order to decide which fragments to

add. The best possible running process is therefore

designing according to a context.

3.1.1 Users’ Objectives and Knowledge

In order to design a tailored process, SCoRe requires

some information about the wanted system and about

the users. Actually, these informations enable to se-

lect fragments which make up the suggested pro-

cess. On the one hand, the user has to give his/her

knowledge about the known technologies, methods

and paradigms. On the other hand, the user has to de-

scribe brieﬂy the intended system by deﬁning the ﬁeld

of application, the phase corresponding to the initial

and ﬁnal work product and the type of system. Figure

1 shows an example of the both kind of characteristics

but it is not an exhaustive list.

Figure 1: Example of characteristics.

3.1.2 Context

The context is a set of elements external to the activity

of an entity. It describes the environment wherein the

entity evolves. Moreover, the context has an inﬂuence

on the process of fragments selection.

In processes under construction, the context is

made up of users’ objectives and knowledge, avail-

able fragments and the elements included in the run-

ning process.

3.1.3 Agents

SCoRe is composed of four distinct kinds of

agents, following a perception-decision-action lifecy-

cle, which cooperate according to the adaptive multi-

agent systems theory described in (Capera et al.,

2003). The basic idea underlying this cooperation

consists, for every agent, in always trying to help

the agent which encounters the most critical situation

from its own point of view.

Figure 2 gives the structure of SCoRe designing

a method process. The different types of agents in-

volved are shown, as well as the links modelling the

existing interactions between them. Actually, SCoRe

is made up of the following agents:

• MAS Metamodel Element (MMME): required

or produced by a Running Fragment, its aim is

to choose which fragment it will be linked to. A

MMME is connected with all the waiting frag-

ment and the Running Fragments which are in-

cluded in the running process and which produce

or consume it.

• Waiting Fragment (WF): its purpose is to inte-

grate its instances (running fragment) in a process

once it is in an adequate situation. They are linked

to the MMMEs and a set of context.

• Running Fragment (RF): it aims at ﬁnding its

localization inside the running process. A RF is

linked with the MMMEs encompassed in the run-

ning process that it produces or requires.

• Context (C): related to a fragment, it aims at

evaluating its relevance according to the MMMEs

already involved in the running process and the

users’ objective and knowledge. The context

agent is related to a fragment for which it eval-

uates its relevance to be added in the running pro-

cess.

Figure 2: Example of agents and their relationships in

SCoRe.

3.2 General Behavior of SCoRe

3.2.1 Prerequisite

Some prerequisites are necessary for the execution of

SCoRe. Actually, in order to help the fragments se-

lection in SCoRe, the user has to ﬁll in two forms

about its objectives and its knowledge. Firstly, the

user completes its knowledge that will be mainly used

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

290

to select fragments. Secondly, the user characterizes

the wished system. From the ticked ﬁelds, the initial

and ﬁnal MMMEs are extracted. Actually, according

to user’s knowledge and the initial and ﬁnal phase se-

lected, MMMEs are included in the running process

by SCoRe.

3.2.2 Starting of the Running Process

Construction

When the user provides the prerequisite items, Wait-

ing Fragment agents and their Context agents are cre-

ated. They know each other; however the Waiting

Fragment don’t know the others Waiting Fragment,

and their associated Context don’t know each others.

Besides, the MMMEs corresponding to users’

problem are created with the acquaintance of the

available Waiting Fragments. They are then included

in the running process. The construction of the run-

ning process starts therefore with the MMMEs. These

MMMEs are said initial and ﬁnal. While a MMME

aims at being linked to at least two fragments, i.e. one

which produces it and one which consumes it, an ini-

tial, respectively ﬁnal, MMME aims at being linked

to at least one fragment which consumes it, respec-

tively produces it. The initial and ﬁnal MMME start

the running process construction by interacting with

all the waiting fragments.

3.2.3 Running Process Construction

As it is shown in ﬁgure 2, an agent communicates

as follow. The ﬁrst agent interactions correspond to

the initial and ﬁnal MMME looking for a fragment.

When a waiting fragment receives a message from a

MMME agent, it solicits their own contexts. The con-

text self-evaluates itself. Then it answers by giving its

relevance to the waiting fragment which sends it to the

MMME in turn. According to the different answers,

one of the requesting MMME selects a waiting frag-

ment. The selected waiting fragment is then ready to

create a running fragment. The created running frag-

ment is added in the running process. Being related to

input and output MMMEs, when a running fragment

is added in the running process, it links itself to the

MMMEs already present in the running process. If

one of its MMMEs is missing, the running fragment

creates it. Then, the created MMME agent is added in

the running process. The MMME agents request the

waiting fragment agents until to be satisﬁed.

Next sections will provide a more detailed de-

scription of these agents behaviors and interactions.

3.3 General Behavior of the Agents

An agent that intervenes in an AMAS is composed

of different parts that produce its behavior: skills,

aptitudes, the interaction language, world represen-

tations, Non Cooperative Situations, and criticality

and/or conﬁdence.

For an agent, criticality represents the degree of

non-satisfaction of its own goal. It enables an agent to

determine the relative difﬁculty of agents in its neigh-

borhood. Evaluation methods and calculation of the

criticality are speciﬁc to each type of agent. The con-

ﬁdence of an agent is an internal measure that pro-

vides information on the reliability of the decision on

actions intended.

These notions guide the SCoRe agents behavior

and will be presented in the following subsections.

3.3.1 The MMME Agents

The MMME agents represent the links between run-

ning fragment agents. Their individual goal is to be

incorporated in the running process. The MMMEs

behavior is represented by an automaton with two

states: non incorporated and incorporated. The non

incorporated state corresponds to a MMME linked to

at least one running fragment which respectively pro-

duces or consumes it. In this state, it is looking for

a relevant fragment to which it can be linked by re-

questing all waiting fragments. The relevant waiting

fragments answer it by giving their conﬁdence. The

conﬁdence of the waiting fragment agent provides in-

formation on its reliability proposition. The most rel-

evant fragment will be chosen by the MMME agent

for being a potential fragment to be added in the run-

ning process and the MMME agent updates its conﬁ-

dence.

The MMME agents evaluate their own criticality

and cooperate to choose the most relevant fragment

according to the criticality of the ones suggested.

The incorporated state is reached when the

MMME agent is linked with at least two fragments:

one consumer and one producer. The initial or ﬁnal

MMMEs provided by the designer have only to be

linked respectively to at least one producer and one

consumer.

3.3.2 The Waiting Fragment Agents

Each waiting fragment agent has an associated set of

context agents. Their goal is to be integrated in a

process once it is in an adequate situation. For that,

when a waiting fragment agent receives requests from

MMMEs which are looking for a fragment, it for-

wards the request to its context agents, if the wait-

ASelf-organizingMulti-AgentSystemforCombiningMethodFragments

291

ing fragment agent considers itself as a potential so-

lution. A waiting fragment agent considers itself to be

solution if the requesting MMME belongs to its own

required or provided MMME. Then, the waiting frag-

ment agent waits the answer from its context agents.

It updates its conﬁdence and sends it to the MMME.

Moreover, a waiting fragment agent can be se-

lected to be inserted in the running process. Informed

by the MMME agent, the waiting fragment agent cre-

ates a running fragment agent and sends the informa-

tion to its context agents as a feedback.

3.3.3 The Running Fragment Agents

A running fragment agent is created by a waiting frag-

ment agent and is introduced on time in the process.

Its aim is to be incorporated in the running process.

Its behavior changes according to its current state and

its perception. The current state of a running fragment

agent corresponds to non incorporated and incorpo-

rated. Actually, a running fragment agent is said in-

corporated when all the required MMMEs are in the

incorporated state and at least one of the provided

MMMEs is incorporated. Otherwise its state is non

incorporated and the running fragment agent makes

links with each MMME agent existing in the running

process on which a link is physically possible. More-

over, when the running fragment agent is inserted in

the running process, it adds the required or produced

MMMEs which are missing in the running process.

3.3.4 The Context Agents

The goal of the context agents is to represent a situa-

tion leading to a speciﬁc method process. They aim at

selecting the fragment to add in the current situation

to reach the objectives. When such an agent ﬁnds it-

self in its triggering situations, it notiﬁes the waiting

fragment agent, by submitting its conﬁdence accord-

ing to its own knowledge.

In order to know when the fragment is relevant, a

context agent relies on two different sets of informa-

tion. First, a collection of input values represents the

set of user and system characteristics. This element

enables the context agent to know if it has to be trig-

gered or not. Besides, a context agent possesses a set

of forecasts, which describes the impact of the action

proposed on the satisfaction of the both user and sys-

tem characteristics. Moreover, a context agent pos-

sesses a set of metrics, which describes the impact of

the action proposed on the running process (Bonjean

et al., 2012). Those input values are modiﬁed during

the life of a context agent. According to its behav-

ior, from different feedback that it receives, a context

agent adjusts its conﬁdence.

Finally, the behavior of a context agent is repre-

sented by an automaton. Each state relates its current

role in the MAS. A total of three different states exist:

disabled, enabled and selected. The context agent can

switch from a state to another thanks to the messages

it receives from other agents in the system. A disabled

context agent considers itself non-relevant in this spe-

ciﬁc situation. An enabled context agent thinks that it

is relevant and potentially deserves to be selected. It

then computes its conﬁdence and sends it to the corre-

sponding waiting fragment agent. Finally, a selected

context agent is validated by a waiting fragment agent

and its associated fragment is added in the running

process.

4 RESULTS AND ANALYSIS

The conducted tests focus on the functional adequacy

and the dynamic adaptation to speciﬁc situation. The

ﬁrst test has been carried out with known method pro-

cesses to show the correctness of SCoRe. We show

that Score enables to compose a known method back

from its fragments. Concerning adaptability test, we

conducted them with ﬁctive processes. Combining

known methods are very complex task. Until now, the

inter-operability and semantic matching of fragments

from different known methods stay an important chal-

lenge. Actually, in this problem, some works base

on standardisation of fragments notion and of their

description. For this reason, we simulate methods.

Therefore we used ﬁctive processes which have been

simpliﬁed. Actually, they are deﬁned in the following

way. They are composed by four fragments while a

known method is made up of approximately twenty

fragments. These tests show how SCoRe is able to

adapt itself: on the one hand, how SCoRe adapts it-

self to a context and on the other hand, how SCoRe

adapts itself to design a new tailored method process.

5 RELATED WORKS

Apart from application ﬁeld, several recent works

exploit the lessons of adaptive self-organizing natu-

ral and social system to enforce self-awareness, self-

adaptability, and self-management features in dis-

tributed system.

Some approaches of components agentiﬁcation

have been developed. They aim at allocating com-

ponents to agent properties such as autonomy and in-

teraction. (Hara et al., 2000) proposes an extension

of ContractNet protocol for ﬁnding components in li-

braries. A function which deals with request as mes-

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

292

sage is added in the components stored in libraries as

an agent. The agents have knowledge about speci-

ﬁcity of the component and its ability to answer a

need. Our proposition has the same goal with a view

to distribute fragments research and their adaptation

and their composition. This allocation of reuse pro-

cess enables some assembling strategies for instance.

Besides, Web Services represent today’s reference

standard technology for the set up of distributed sys-

tems that need to support machine-to-machine inter-

action among heterogeneous applications distributed

over a network. The automatic composition and adap-

tation of services has been explored using a variety of

AI planing engines (Rao and Su, 2005).

In (Thomas et al., 2009), a set of workﬂow frag-

ments are composed in ad hoc wireless mobile envi-

ronments. This approach designs dynamically con-

struction of custom, context-speciﬁc workﬂows in re-

sponse to unpredictable and evolving circumstances

by exploiting the knowledge and services available

within a given context. For that, a graph made up

of all workﬂow fragments is built up before explor-

ing and pruning it. As presented approaches, ours

is based on current data base of fragments and on

MAS metamodel elements. The way to integrate the

method fragment in the process is different because

in running development, our approach can take into

account process adaptation according to development

context.

6 CONCLUSIONS AND FUTURE

WORKS

This paper presents SCoRe architecture, an adaptive

multi-agent system, which designs a tailored process

by combining fragments together. Each agent com-

posing the adaptive multi-agent systems follows a lo-

cal and cooperative behavior, driven by the use of

their conﬁdence. The four different kinds of agents,

composing the SCoRe system, were deﬁned in order

to self-design and self-combine a tailored method pro-

cess without relying on the method engineer. The re-

sulting behavior of the SCoRe system is the ability

to design process and adjust the proposed process ac-

cording to the characteristics of application domain

and users proﬁle. This ﬁrst prototype allowed to en-

hance our experience on practical problems such as

metamodel compatibility, parameters composition or

fragments adaptation to speciﬁc ﬁeld.

However, there is still room from improvements

for incoming interoperability and for evaluating the

designed process.

REFERENCES

Bergenti, F., Gleizes, M., and Zambonelli, F. (2004).

Methodologies And Software Engineering For Agent

Systems: The Agent-oriented Software Engineering

Handbook.

Bernon, C., Camps, V., Gleizes, M.-P., and Picard, G.

(2005). Engineering Adaptive Multi-Agent Systems:

The ADELFE Methodology . In Agent-Oriented

Methodologies , pages 172–202.

Bonjean, N., Chella, A., Cossentino, M., Gleizes, M.-P.,

Migeon, F., and Seidita, V. (2012). Metamodel-Based

Metrics for Agent-Oriented Methodologies. In Inter-

national Joint Conference on Autonomous Agents and

Multiagent Systems (AAMAS).

Brinkkemper, S., Saeki, M., and Harmsen, F. (1998). As-

sembly Techniques for Method Engineering.

Capera, D., Georg

e, J.-P., Gleizes, M.-P., and Glize, P.

(2003). The AMAS Theory for Complex Prob-

lem Solving Based on Self-organizing Cooperative

Agents. In International Workshop on Theory And

Practice of Open Computational Systems, pages 389–

394.

Cossentino, M., Fortino, G., Garro, A., Mascillaro, S., and

Russo, W. (2008). Passim: A simulation-based pro-

cess for the development of multi-agent systems. In-

ternational Journal on Agent Oriented Software Engi-

neering (IJAOSE).

Hara, H., Fujita, S., and Sugawara, K. (2000). Reusable

software components based on an agent model. In

Proceedings of the Seventh International Conference

on Parallel and Distributed Systems: Workshops, IC-

PADS ’00.

Henderson-Sellers, B. and Giorgini, P. (2005). Agent-

oriented methodologies.

Henderson-Sellers, B. and Ralyt

e, J. (2010). Situational

method engineering: State-of-the-art review. J. UCS,

pages 424–478.

Ralyt

e, J. (2004). Towards situational methods for infor-

mation systems development: Engineering reusable

method chunks. Procs. 13th Int. Conf. on Information

Systems Development. Advances in Theory, Practice

and Education, pages 271–282.

Rao, J. and Su, X. (2005). A Survey of Automated Web

Service Composition Methods. pages 43–54.

Thomas, L., Wilson, J., Roman, G.-C., and Gill, C. (2009).

Achieving coordination through dynamic construc-

tion of open workﬂows. In Proceedings of the

10th ACM/IFIP/USENIX International Conference on

Middleware, Middleware ’09.

ASelf-organizingMulti-AgentSystemforCombiningMethodFragments

293