A Neural Network-Based Agent Framework for Mail Server Management

Copyright © 2005, Idea Group Inc. Copying or distributing in print or electronic forms without writtenpermission of Idea Group Inc. is prohibited.

International Journal of Intelligent Information Technologies, 1(4), 35-51, Oct-Dec 2005 35

A Neural Network-BasedAgent Framework for

Mail Server ManagementCharles C. Willow, Monmouth University, USA

ABSTRACT

Amidst the era of e-economy, one of the difficulties from the standpoint of the informationsystems manager is, among others, the forecast of memory needs for the organization. Inparticular, the manager is often confronted with maintaining a certain threshold amountof memory for a prolonged period of time. However, this constraint requires more thantechnical and managerial resolutions, encompassing knowledge management for thegroup, eliciting tacit knowledge from the end users, and pattern and time series analysesof utilization for various applications. This paper proposes a framework for building anautomated intelligent agent for memory management under the client-server architecture.The emphasis is on collecting the needs of the organization and acquiring the applicationusage patterns for each client involved in real time. Due to the dynamic nature of thetasks, incorporation of a neural network architecture with tacit knowledge base issuggested. Considerations for future work associated with technical matters comprisingplatform independence, portability, and modularity are discussed.

Keywords: automata; automatic intelligent agent; computer-supported collaborationwork; human-computer interaction; information resource management;knowledge base; knowledge management; memory management; neuralnetworks; tacit knowledge

INTRODUCTION

Integrated information systems fordistributed organizations comprising theinformation technology (IT) infrastructure

and generic business applications, such asenterprise systems (ES), supply chainmanagement (SCM), customer relation-ship management (CRM), and knowledgemanagement system (KMS), are by far

36 International Journal of Intelligent Information Technologies, 1(4), 35-51, Oct-Dec 2005


one of the vital assets to sustain the com-petitive edge in e-economy. At the sametime, however, administrators of these in-formation systems often are confoundedwith a vast array of management prob-lems, which, in large, may be classified intothe following:

• Selection and upgrades of hardwareplatform, middleware, and applicationscombination;

• Information resource management;• Integration of applications; and• Security management.

This paper discusses problems as-sociated with information resource man-agement (IRM) and suggests a develop-ment framework for reconfiguring theserver as well as clients for moderatelylarge-scale information systems.

Among others, one of the difficultiesconcerning IRM from the standpoint ofthe information systems manager is thecorrect forecast of overall memory needsfor the organization. In particular, the man-ager often is confronted with maintaininga certain threshold amount of memory fora prolonged period of time. One may ar-gue that the cost of memory is decliningrapidly, and its management may not be afactor affecting IRM. Contrary to this com-mon misbelief, however, a number of au-thors suggest that there should be a cer-tain threshold for memory management(Applen, 2002; Kanawati & Malek, 2002;Kankanhalli et al., 2003; Lansdale, 1988;Mathe & Chen, 1998; Pinelle & Gutwin,2002; Pinelle et al., 2003; Roos et al.,2003) within a prescribed time window,analogous to budgetary considerations. Inessence, memory management affects

Figure 1. General information systems architecture

Clients (Graphical User Interface)

Middle Tier

Back-end Server (Demilitarized Zone: Safety Zone for Databases)

...

DMZ

E-Mail Server

Web Server

Applications Server

Narrow-band

Broad-band



overall performance for both client-serverand peer-to-peer architectures of the in-formation system.

Under multi-tiered server configura-tion, which seems to be the norm for mostorganizations, access to the very back-end server(s) requires high bandwidth net-works (Willow, 2005a). As a conse-quence, the impact of effective memorymanagement is amplified further. Figure 1represents a general three-tier architec-ture of information system.

The focus of this paper is on build-ing a conceptual framework for develop-ing intelligent memory management sys-tem for the mail server. To tackle the IRMproblem more comprehensively, a modu-lar approach is employed. As e-mails be-come one of the preferred methods ofcommunications for most organizations,mail server management indeed requiresattention with top priority.

The major contribution of this paperlies in the development of a frameworkfor building multiple agents for the mailserver, with an emphasis on memory man-agement. Agent-based autonomous sys-tems (i.e., automata) in recent years havebeen adopted as one of the better meth-ods for managing virtual organizations invarious applications (Flenner et al., 2002;Kanawati & Malek, 2002; Mathe &Chen, 1998; Murch, 2004; Taylor, 2004).They range from consumer applications,such as online travel arrangements, to sys-tem diagnostics, including online data qual-ity audits and remote troubleshooting.

The organization of this paper fol-lows. In the second section, knowledgemanagement for eliciting, building, andmanaging end-user preference and e-mail

usage patterns is discussed. The third sec-tion follows to illustrate the core system— multiple agents. Suggestions for (hands-on) construction of the proposed frame-work are made in the fourth section, fol-lowed by a section devoted to conclusions.

KNOWLEDGEMANAGEMENT

The key to maintaining accuracy ofthe proposed multi-agent system lies inmanaging highly subjective knowledge forsharing and customizing or personalizingend-user memory usage patterns acrossthe organization. Information technology(IT) may support knowledge management(KM) in two classes: codification and per-sonalization (Kankanhalli et al., 2003). Inessence, the codification approach man-ages structured knowledge, whereas per-sonalization manages unstructured, tacitknowledge. Because e-mail usage patternsfor end users may entail both types ofknowledge, a separate knowledge baseor repository is suggested for the frame-work of this research. That is, there maybe, on the one hand, common patterns ofe-mail management among end users, suchas removing messages which are morethan 36 months old, organizing e-mail fold-ers every 30 days, and so forth. On theother hand, each user may have highlysubjective patterns that may not be con-sistent with those codified knowledge.Lansdale (1988) emphasized in his earlyresearch the need for Cognitive InterfaceTools (CITs) to collect, organize, build,and share both types of knowledge forthe office system. However, not many lit-eratures to date have been dedicated tosolving this problem.



Markus (2001) describes the gen-eral stages of the KM process: knowl-edge elicitation, knowledge packaging,distributing or disseminating knowledge,and reusing knowledge. To this end, build-ing knowledge bases in conjunction withautomatic agents is considered one of thebetter methods for managing user knowl-edge associated with e-mail usage in realtime.

Attributes forKnowledge Management

This section illustrates the necessaryset of attributes to be incorporated intoKM. Note that the values of these at-tributes will be collected in real time fromend users for generating patterns of e-mailusage.

As noted in Lansdale (1988), theprocess of information retrieval in the hu-man mind is fundamentally different froma filing or library system, in which itemsare accessed by location rather than bytheir meaning. The first notion is that peoplerecall chronological information about in-formation (i.e., what else was happeningat roughly the same time). Consequently,

time stamp of e-mails may be a goodsource of structured information or knowl-edge. Association is another means bywhich humans retrieve information. Eache-mail message is associated with fourpieces of tacit information: recipient orsender, event or subject, attachment(s),and significance of the message. Table 1summarizes the attributes for KM con-cerning e-mail usage.

A set of six generic attributes asso-ciated with e-mails, as described in Table1, is to be employed in the suggestedframework of this paper. Notice that thefirst two attributes — time stamp and size— are structured information, which maybe available for both the server and cli-ents. By contrast, each client may man-age his or her e-mail messages, basedon one or more of the four tacit at-tributes: recipient/sender, event/subject,attachment(s), and significance. Givena certain restriction of memory size, say100MB per e-mail account holder, oneclient may choose either to remove or toarchive (on local memory store) e-mailsbased on recipient/sender, event/subject,attached file(s), significance, or any com-

Table 1. Attributes of knowledge management for e-mail usage

Attributes Type of Knowledge

Time Stamp Structured

Size (KB) Structured

Recipient/Sender Tacit/Unstructured

Event/Subject Tacit/Unstructured

Attachment Tacit/Unstructured

Significance Tacit/Unstructured



bination of the four, so far as tacit knowl-edge management is concerned. Alterna-tively, he or she simply may choose toarchive or to remove e-mails with regardto structured information, such as timestamp and/or size. It is precisely thisknowledge associated with each e-mailclient that is expected to be elicited by theautomated multiple agents proposed in thispaper, preferably in real time.

INTELLIGENTAUTOMATED AGENT

Under a certain memory constraint foreach e-mail client, the administrator of theinformation system (i.e., mail server) maychoose to adopt a brute-force approach,based on structured information, such astime stamp or size of the message. As aconsequence, clients (without their consent)often may realize that their e-mails are un-available at times, once they have reachedthe memory quota set by the system. How-ever, this aggressive method is not effec-tive, due to its user service and, perhaps,legal implications. Thus, an automated sys-tem that may advise clients in real time abouttheir e-mail usage patterns is consideredan attractive alternative for information re-source management. Once the client islogged on to the system, the automatic in-telligent agent generates a list of e-mailmessages that are to be removed as wellas those that are highly likely to be candi-dates for local archives. In addition, an-other agent system is suggested in orderfor the server to advise the administrator(s)of potential preventative measures. In es-sence, a conceptual framework for a multi-agent system is proposed in this paper.Similar ideas are being incorporated at

present into the Web and applications serv-ers in Willow (2005a, 2005b).

Neural networks (NN) are em-ployed as the inference engine for the pro-posed multi-agent system. An NN typi-cally processes large-scale problems interms of dimensionality, amount of datahandled, and the volume of simulation orneural hardware processing (Willow,2002). It emerged as an area of artificialintelligence (AI) to mimic human neuronsin both perception and learning. It is inter-esting to note, however, that a conceiv-ably disparate area within information sci-ence classified as knowledge representa-tion brought the attention of researchersto pursue classes of computing and pro-cessing, such as neural networks. An ob-ject-oriented paradigm emerged as oneof the better models for knowledge rep-resentation. In fact, the motivation for NNresearch was to seek an improved meth-odology in machine learning and, morespecifically, in the area of planning algo-rithm, thereby augmenting the techniquesavailable at the time. However, as the re-search progressed, more obstacles toemulating human neurons were realized.Toward this end, the jargon NN at presentwould be more appropriate if it were re-placed with parallel, distributed simulation.Figure 2 illustrates taxonomic views of NN(Willow, 2002). Notice that it is not com-prised of an exhaustive list of available NNmodels to date.

The concept of feedback plays acentral role in learning for NN. As illus-trated in Figure 3, two different types oflearning are to be distinguished: learningwith supervision (i.e., training) vs. learn-ing without supervision (Willow, 2002).



In supervised learning (see Figure3(a)), the desired response d of the sys-tem is provided by the teacher at each in-stant of time. The distance ρ [d, o] be-tween the actual and the desired responseserves as an error measure and is used tocorrect network parameters externally.Since adjustable weights are assumed, theteacher or supervisor may implement areward-or-punishment scheme to adaptthe network’s weight matrix, W. This modeof learning is pervasive and is used in manysituations of natural learning. A set of in-put and output patterns, called a trainingset, is required for this learning mode.Often, the inputs, outputs, and computedgradient are deterministic; however, theminimization of error proceeds over all itsrandom realizations. As a result, most su-pervised learning algorithms reduce to sto-chastic minimization of error in multi-di-mensional weight space.

In learning without supervision (Fig-ure 3(b)), the desired response (d) is notknown; thus, explicit error informationcannot be used to improve network be-

havior. Since no information is availableas to correctness or incorrectness of re-sponse, learning somehow must be ac-complished, based on observations of re-sponses to inputs of marginal or, at times,no knowledge. Unsupervised learning al-gorithms use patterns that typically areredundant raw data having no labels re-garding their class membership or asso-ciation. In this mode of learning, the net-work must discover for itself any pos-sible existing patterns, regularities, sepa-rating properties, and so forth. While dis-covering these, the network undergoes achange of its parameters, which is calledself-organization. Adaptive ResonanceTheory (ART) is a good example of sucha class.

Adaptive Resonance TheoryAdaptive Resonance Theory

(ART), as illustrated in Zurada (1992),is a unique unsupervised class of neuralnetwork algorithm. It has the novel prop-erty of controlled discovery of clusters.Further, the ART network may accom-

Figure 2. Classification of neural network models

Neural Networks

Network Architecture

FeedBack (F/B) dynamic

FeedForward (F/F) static

Perceptron Hopfield Functional Link Associative Memory

Hopfield Bidirectional Assoc. Memory (BAM) Temporal Assoc. Memory (TAM): Time Series Adaptive Resonance Theory (ART) 1 & 2

Training Algorithms

Backpropagation rule Widrow-Hoff’s Delta/Least Mean Square (LMS) rule Perceptron rule Hebbian rule Correlation rule Outstar rule Winner-takes-all rule: Greedy algorithm



modate new clusters without affecting thestorage or recall capabilities for clustersthat were already learned, fit for the

scope of the problem of this paper. Fig-ure 4 illustrates the ART architecture(Zurada, 1992).

Figure 3. Learning modes for neural networks: (a) supervised (left diagram) vs.(b) unsupervised (right diagram)

Adaptive network

W

Distance generator

x o

d

Learning signal

[d, o]

Distance measure

Adaptive network

W

x o

Legends

x : input pattern vector

o : output vector of a neuron layer

W : weight matrix

d : desired output vector of a trained network

: vigilance test level (error level)

Figure 4. Neural network architecture for Adaptive Resonance Theory



Nomenclature of the model follows:

Subscripts and Superscripts

i Subscript for input variable, x; i = 1,..., n.

j Subscript for output clusters, j = 1, ...,M.

m Subscript for output neuron, y or neu-ron of hidden layer; m = 1, ...j, ..., M.

k Superscript for neuron y at layer k; k ≥0.

Parameters

M Total number of clusters set by the de-cision maker.

n Total number of variables for input vec-tor/tuple, x = [x

1, ..., x

n] = < x

1, ...,

xn>.

Variables

x Input vector; x = [x1, ..., x

n].

w Weight of the input vector; w = [w1,

…, wn].

y Output vector; yk = [y1, ..., y

M].

r Controlled vigilance factor indicatingcloseness of input to a stored clusterprototype to provide a desirable match;0 < ρ < 1. The ART net will seek aperfect match for ρ =1 and looselycoupled matches for lower values of ρ.

v Weight vector for verifying cluster ex-emplar proximity; v = [v

1, ..., v

n].

t Update index for weights, w and v.

Algorithm for ART is summarized asfollows (Zurada, 1992):

Step 1: InitializationThe vigilance threshold, ρ, is set.Weights are initialized for n-tuple in-

put vectors and M top-layer neurons.(M ××××× n) matrices W and V each

are initialized with identical

W =

+ n1

1(1)

V = [1] (2)

0 < ρ < 1 (3)

Step 2: Input Neuron ProcessingBinary unipolar input vector x is pre-

sented at input nodes, xi = 0, 1 for i = 1,

2, …, n.

Step 3: Matching Score ComputationAll matching scores are computed

as follows:

∑=

=n

iiim

om xwy

1 for m = 1, …, M. (4)

In this step, selection of the bestmatching existing cluster, j, is performedaccording to the maximum criterion, asfollows:

)(max,...,1

om

Mj

oj yy

== (5)

Step 4: ResonanceThe similarity test for the winning

neuron j is performed as follows:

∑=

>n

iiij xv

x 1

1 ρ (6)



where the norm is defined as:

∑=

≡n

iixx

1(7)

If the test as illustrated in equation(6) is passed, the control is passed on toStep 5. Upon failing the test, Step 6 isfollowed only if the top layer has morethan a single active node left. Otherwise,Step 5 is followed.

Step 5: Vigilance TestEntries of the weight matrices are

updated for index j passing the test of Step4. The updates are only for entries (i, j),where i = 1, 2, ..., M, and are computedas follows:

∑=

+=+ n

iiij

iijij

xtv

xtvtw

1

)(5.0

)()1(

(8)

vij(t + 1) = x

iv

ij(t) (9)

This updates the weights of the j-thcluster, newly generated or existing. Thealgorithm returns to Step 2.

Step 6: Cluster GenerationThe node j is deactivated by setting

yj to 0. Thus, this mode does not partici-

pate in the current cluster search. The al-gorithm goes back to Step 3 and will at-tempt to establish a new cluster differentfrom j for the pattern under test.

Clusters are generated by the net-work itself, if such clusters are identifiedin input data, and store the clustering in-formation about patterns or features in theabsence of a priori information about the

possible number and type of clusters. Inessence, ART computes the input-pattern-to-cluster matching score (y), which rep-resents the degree of similarity of thepresent input to the previously encodedclusters. The vigilance threshold, ρ, where0 < ρ < 1, determines the degree of re-quired similarity, or match, between a clus-ter or pattern already stored in the ARTnetwork and the current input in order forthis new pattern to resonate with the en-coded one. If no match is found, then anew class or cluster is created.

Applications of ART to the pro-posed multi-agent system follow in the nexttwo subsections.

Client Agent Architecture with ARTThis section describes the ART ar-

chitecture for the suggested client agent.The object of the client agent is to pro-vide real-time knowledge regarding e-mailmanagement for each client end user. Fig-ure 5 follows to illustrate.

For each e-mail message, an inputvector comprised of the following seven-tuple attribute is produced; x = <x

1, ...,

x7>:

x1Age of the message. It is automaticallycomputed as system-generated time(TNOW) – (Time_Stamp).

x2Size of the e-mail, generally measuredin kilobytes (KB).

x3Recipient information in SMTP addressformat (To: [email protected]).

x4Sender information in SMTP addressformat (From: [email protected]).Note that x

3 and x

4 are mutually exclu-

sive, and may have null values associ-ated. That is, each message is either



received from or strictly sent to anSMTP address.

x5Subject of the message (characterstrings).

x6Attachment to the message. A uniquefour-digit code encompassing the num-ber of attachments between 0 and 99(first two digits) and their file types isassigned for x

6. To simplify the data

structure, file types are restricted to thethree most common on the Internet;ASCII .txt (1), Microsoft .doc (2), andAdobe .pdf (3). Examples of x

6 values

are:

0000 No attachments.0103 One attachment in .pdf format.9923 Ninety-nine attachments in mix-

tures of .doc and .pdf files.

x7Significance of the e-mail message, setby the end user. It ranges from 1 to 5,5 being the most significant and 1 being

the least. Note that this value is appli-cable exclusively for the body of thee-mail. For instance, a message withx

7 = 1 does not warrant automatic re-

moval from the mailbox. Instead, theuser may choose to archive it due tothe importance of its attachment(s), x

6,

for example.

A simplified numerical example fol-lows to illustrate. Consider the followingthree messages for a client:

Message #1 = <x1, …, x

7>

= <60, 2000,[email protected], null, “publica-tion consideration”, 0203, 5>

Message #2 = <02, 20, null,[email protected], “Greetings”,0000, 1>

Message #3 = <14, 250,[email protected], null, “gettogether”, 0000, 5>

Figure 5. ART for the client agent

Time Stamp

Size Reci-pient

Sen- der

Sub-ject

Attach- ment

Signi- ficance

x1 x2 x3 x4 x5 x6 x7

Remove

Archive

Keep

y1 y2 y3

. . .

w11

v11

w12

w13

w73 w72 w71

… …

v73

1

-

- - 1



Thus, each message forms an inputpattern vector, x. Tacit input values areconverted into a utility scale of 1 to 5 forneural processing, based on interactionswith the knowledge base. This pertains tothe attributes, x

3, x

4, and x

5. As a conse-

quence, the input vectors are:

x1 = <60, 2000, 4, null, 5, 02, 5>x2 = <02, 20, null, 5, 1, 00, 1>x3 = <14, 250, 5, null, 1, 00, 5>

When x1 is presented, the steps ofthe ART algorithm are:

M = 3; n = 7

1

1

+=

nwij = 1/8 = 0.125;

vij= 1, i = 1, …, 7, j = 1, 2, 3 for Re-move, Archive, Keep.

Given a standard vigilance value ofρ = 0.5, the left term in inequality (6) is ofunity in the first pass, allowing the similar-ity test to be passed. This results in un-conditional definition of the first cluster, thedefault being the “Archive.” Equations(8) and (9) of Step 5 produce:

2424.0]52504[5.0

41)2(32 =

+++++×=w

3030.0]52504[5.0

51)2(52 =

+++++×=w =

w72

for the tacit variable set only. Noticex

3, x

5, and x

7 had significant values of 4 or

above. The remaining weights wi2 = 0.125,

as initialized in Step 1. In addition,

v32

= v52

= v72

= 1,

as initialized, while the remainingweights are recomputed as v

i2 = 0.

For the second input pattern vectorx2, there are no significance values, andthe similarity test of equation (6) yields

∑=

>7

1

1

iiij xv

xρ = 0 < 0.5

Due to the failure of the vigilance testand the absence of other nodes for fur-ther evaluation and for potential disabling,pattern x2 is treated as another new clus-ter. Further, a null value for the left-handside of equation (6) is classified as the“Remove” cluster.

In essence, the ART-based neuralnetwork processing is expected to advisethe e-mail clients of possible action(s) foreach (e-mail) message.

Server AgentArchitecture with ART

The Adaptive Resonance Theory(ART) model also may be employed foranother set of agent systems dedicated toassisting the e-mail server administrator(s).The two systems of agents, which targetindividual clients as well as administrators,then may communicate in real time by ac-cessing the integrated knowledge base.However, the server agent architecture isrelatively more complicated due to differ-ences in network protocols. Two majormethods employed are: Internet MessageAccess Protocol (IMAP) and Post Of-fice Protocol (POP). IMAP integratesmessages on the shared mail server andpermits client e-mail programs to accessremote message stores as if they were lo-cal. Thus, the memory burden on the



server is far greater for IMAP than forPOP. However, complete monitoring ofclient e-mails is possible under the IMAPscheme. That is, a single server-side agentsystem may suffice for the IMAP, whereasa single client-agent may fit systems withPOP being implemented. See Table 2 foran inllustration of this.

A multi-agent system, therefore, isexpected to be highly useful for managingclient knowledge under the IMAP (fourthquadrant in Table 2) and for managingserver knowledge under the POP struc-ture (second quadrant).

IMPLEMENTATIONAGENDA

Architecture of the proposed multi-agent system, followed by description ofits functions and the challenges encoun-tered during systems analysis, are dis-cussed in this section.

A generalized Multi-Agent System(MAS) architecture called RETSINA hasbeen presented by Sycara et al. (1996).Three classes of agents are proposed inSycara et al. (1996): interface, task, andinformation. The major objective of theinterface agents is to interact with the cli-ents/users in order to receive user specifi-cations and to deliver the results. They

acquire, model, and utilize user prefer-ences to guide system coordination in sup-port of the user’s tasks. Task agents per-form the majority of autonomous prob-lem solving and, thus, are regarded as theinference schema of RETSINA. It exhib-its higher levels of sophistication and com-plexity than either an interface or an infor-mation agent. Information agents provideintelligent access to a heterogeneous col-lection of information sources depicted atthe bottom of Figure 6.

Having gathered an intuitive under-standing of a generalized Multi-Agent Sys-tem (MAS) architecture in Sycara et al.(1996), the architecture of the proposedMAS for mail server management is builtsimilar to the RETSINA architecture. Fig-ure 6 shows our proposed architecture,and its components are briefly describedin subsequent paragraphs.

In the application domain of mailserver memory management, it is interest-ing to note that the users or clients act asinformation source. That is, the input pat-tern vector x, to be incorporated into theAdaptive Resonance Theory (ART) algo-rithm used by the four task agents and rep-resented by shaded hexagons, is collectedby each corresponding interface agent fora user. In essence, the interface agents

Table 2. Classes of information availability based on e-mail application protocols

E-Mail Protocols

Information Process

IMAP (central)

POP (distributed)

Server-based Server only Mixed

User-based Mixed Client only



function as both interacting and informa-tion agents. There may be as many as mnumber of interface agents for the n num-ber of clients, where n ≥ m, since someusers may choose to decline the auto-mated service that tracks and managestheir e-mail messages. Instead, they willbe liable for managing their memory quotamanually. Dashed arrows indicate accessto the knowledge repository, in which pat-terns of e-mail usage for each client areretained in the form of rules.

Three task agents — Input WeightVector, Output Weight Vector, and Vigi-lance Factor — interact dynamically withknowledge bases in order to adjust asyn-chronously in real time. In effect, the neu-

ral network based on ART learns withoutsupervision, and a unique cluster is gener-ated for each e-mail message. Possibleclusters were illustrated in Figure 5.

Implementation of the proposed ar-chitecture has been initiated with DellPowerEdge™ 1850 server with Intel Xeonprocessor at clock speed up to 3.0GHzand 1.0GB RAM. At present, a closedproprietary network with two clients is be-ing tested for building the prototype multi-agent system. Network Operating System(NOS) of choice is Linux, with Visual C++as the major development platform.

In building a technical infrastructure,the following obstacles are expected,among others:

Figure 6. Multi-agent system architecture for mail server management

Clients

. . .

User 1 User 2 User n

Interface & Information Agents

Interface Agent 1

Interface Agent 2

Interface Agent m

...

Knowledge Repository

Knowledge Base

KB1

KB2

KBi

. . .

Inference/ Task Agents

Input Weight Vector Agent (w)

Output Weight Vector Agent

(v)

Vigilance Factor Agent ( )

Cluster Generation

Agent



• Difficulty of Data Mining. Executionof a spyware is inevitable on the clientmachine, which may develop legal im-plications. At present, cookies are be-ing considered as the quick implemen-tation vehicle.

• Network Utilization. Running a multi-agent system in real time may decreasethe network performance in terms ofbandwidth utilization and reliability to acritical level.

• Portability/Scalability. There is a con-stant portability problem of this pro-posed agent system with respect to op-erating system and/or hardware plat-form. Platforms running operating sys-tems other than Linux have to be simu-lated and tested for, once this proto-type completes its pilot run.

CONCLUSIONA conceptual framework for a real-

time multi-agent system built with neuralnetwork and knowledge base has beenpresented in this paper, with an emphasison information resource management(IRM). Managing client as well as serverknowledge concerning e-mails was se-lected as the scope of this research due toits significance as a major communicationvehicle in the e-economy.

Adaptive Resonance Theory (ART)was the primary algorithm of choice forthe neural-network engine due to its ca-pability to achieve unsupervised learning.A simplified numerical example was pro-vided to illustrate the effectiveness of ARTapplied to the problem domain.

Marked differences were discoveredfor the two major e-mail protocols for theserver: IMAP and POP. As a conse-

quence, the suggested multi-agents areexpected to be most effective for manag-ing client knowledge under the IMAP andfor managing server knowledge under thePOP structure.

Challenges of implementing the pro-posed framework include but are not re-stricted to data mining, network utilization,portability, and security.

REFERENCESAhuja, M.K. (1998). Network structure in

virtual organizations. Journal of Com-puter-Mediated Communication, 3(4).

Applen, J.D. (2002). Technical commu-nication, knowledge management, andXML. Technical Communication,49(3), 301-313.

Cohen, F. (2004). JAVA testing and de-sign: From unit testing to automatedWeb tests. New York: Prentice-Hall.

Curran, T.A., & Ladd, A. (2000). SAPR/3 business blueprint: Understand-ing enterprise supply chain manage-ment, 2/E. New York: Prentice-Hall.

Davenport, T.H., & Prusak, L. (1998).Working knowledge: How organiza-tions manage what they know. Bos-ton: Harvard Business School Press.

Dutta, S. (1997). Strategies for imple-menting knowledge-based systems.IEEE Transactions on EngineeringManagement, 22, 79-90.

Eales, R.T.J. (2004). A knowledge manage-ment approach to user support. In Pro-ceedings of the Fifth Conference onAustralasian User Interface Confer-ence (AUIC2004), Dunnedin (pp. 33-38).

Flenner, R. et al. (2002). Java P2P un-leashed: With JXTA, Web Services,XML, Jini, JavaSpaces, and J2EE.



New York: Prentice-Hall.Gallivan, M.J., Eynon, J., & Rai, A.

(2003). The challenge of knowledgemanagement systems: Analyzing thedynamic processes underlying perfor-mance improvement initiatives. Infor-mation Technology and People,16(3), 326-352.

Gold, A.H., Malhotra, A., & Segars, A.H.(2001). Knowledge management: Anorganizational capabilities perspective.Journal of Management InformationSystems, 18(1), 185-214.

Grudin, J. (1988). Why CSCW applica-tions fail: Problems in the design andevaluation of organizational interfaces.In Proceedings of the 1988 ACMConference on Computer SupportedCooperative Work, New York (pp.85-93).

Kanawati, R., & Malek, M. (2002). Amulti-agent system for collaborativebookmarking. In Proceedings of theFirst International Joint Conferenceon Autonomous Agents and Multi-agent Systems: Part 3, Bologna, Italy(pp. 1137-1138).

Kankanhalli, A., Tanudidjaja, F., Sutanto,J., & Tan, C.Y. (2003). The role of ITin successful knowledge managementinitiatives. Communications of theACM, 46(9), 69-73.

Lansdale, M. (1988). The psychology ofpersonal information management. Ap-plied Ergonomics, 19(1), 55-66.

Laurillau, Y., & Nigay, L. (2002). Cloverarchitecture for groupware. In Pro-ceedings of the 2002 ACM Confer-ence on Computer Supported Coop-erative Work, (pp. 236-245).

Majchrzak, A., Rice, R.E., Malhotra, A.,

& King, N. (2000). Technology adap-tation: The case of a computer-sup-ported inter-organizational virtual team.MIS Quarterly, 24(4), 569-600.

Malek, M. (2000). Hybrid approachesintegrating neural networks and casebased reasoning: From loosely coupledto tightly coupled models. In K.P.Sankar, S.D. Tharam, & S.Y. Daniel(Eds.), Soft computing in case-basedreasoning (pp. 73-94). New York:Springer.

Markus, M.L. (2001). Toward a theoryof knowledge reuse: Types of knowl-edge reuse situations and factors in re-use success. Journal of ManagementInformation Systems, 18(1), 57-93.

Markus, M.L., & Connolly, T. (1990).Why computer supported collaborativework applications fail: Problems in theadoption of interdependent work tools.In Proceedings of the InternationalConference on Computer SupportedCollaborative Work (CSCW ’90),New York (pp. 371-380).

Mathe, N., & Chen, J.R. (1998). Orga-nizing and sharing information on theWorld-Wide Web using a multi-agentsystems. In Proceedings of ED-ME-DIA ’98 Conference on EducationalMultimedia and Hypermedia,Freiburg, Germany.

Murch, R. (2004). Autonomic comput-ing. New York: Prentice-Hall.

Nonaka, I., & Takeuchi, H. (1995). Theknowledge-creating company: HowJapanese companies create dynam-ics of innovation. New York: OxfordUniversity Press.

Pinelle, D., & Gutwin, C. (2002).Groupware walkthrough: Adding con-



text to groupware usability evaluation.In Proceedings of the ACM SIGCHIConference on Human Factors inComputing Systems: Changing ourWorld, Changing Ourselves.

Pinelle, D., Gutwin, C., & Greenberg, S.(2003). Task analysis for groupwareusability evaluation: Modeling shared-workspace tasks with the mechanics ofcollaboration. ACM Transactions onComputer-Human Interaction,10(4), 281-311.

Roos, L.L., Soodeen, R-A., Bond, R., &Burchill, C. (2003). Working more pro-ductively: Tools for administrative data.Health Services Research, 38(5),1339-1357.

Roseman, M., & Greenberg, S. (1996).Building real-time groupware withGroupKit, a groupware toolkit. ACMTransactions on Computer-HumanInteraction, 3(1), 66-106.

Sycara, K., Pannu, A., Williamson, M.,Zeng, D., & Decker, K. (1996). Dis-tributed intelligent Agents. IEEE Intel-ligent Systems, 11(6), 36-46.

Taylor, W.A. (2004). Computer-medi-ated knowledge sharing and individualuser differences: An exploratory study.European Journal of InformationSystems, 13, 52-64.

Wiesenfeld, B.M., Raghuram, S., &Garud, R. (1998). Communication pat-terns as determinants of organizationalidentification in a virtual organization.Journal of Computer-MediatedCommunication, 3(4).

Willow, C.C. (2002). A feedforwardmulti-layer neural network for machine

cell formation in computer integratedmanufacturing. Journal of IntelligentManufacturing, 13(2), 75-87.

Willow, C.C. (2005a). Neural network-based multiple agent system for Webserver management. Monmouth, NJ:Monmouth University.

Willow, C.C. (2005b). Neural network-based multiple agent system for ap-plication server management.Monmouth, NJ: Monmouth University.

Ye, Y., Fischer, G., & Reeves, B. (2000).Integrating active information deliveryand reuse repository systems. In Pro-ceedings of the Eighth ACMSIGSOFT International Symposiumon Foundations of Software Engi-neering: 21 Century Applications,San Diego, California (pp. 60-68).

Zurada, J.M. (1992). Introduction toartificial neural systems. St. Paul,MN: West Press.

ACKNOWLEDGMENTSMy deepest gratitude is extended to

Drs. Edward Christensen, Bob Smith, andDiana Sharpe of Monmouth University fortheir time and invaluable suggestions whileacquiring ideas and intuition for this pa-per. Gratification is also expressed to themembers of the Computing Society forINFORMS and Dr. Ramesh Sharda ofOklahoma State University in particularfor thoughtful insights.

This paper was supported in part bythe 2004 Summer Research Grant from theBusiness Council of the Monmouth Uni-versity, West Long Branch, New Jersey.



Charles Willow (www.monmouth.edu/~cwillow/), PhD, is an assistant professor ofmanagement information systems (MIS) and management of technology (MOT) inthe School of Business Administration at the Monmouth University, West Long Branch,New Jersey, USA. As an engineer-and-computer-scientist-turned-management-faculty,Dr. Willow’s research agenda has been extensive, ranging from information systemsdevelopment to case-driven strategic management issues, systems engineering, andoperations research. His current research interests include information systems andnetwork security, neural network applications, intelligent software agents, computer-generated graphics, cost-model analysis of Internet-based business, and strategicmanagement of technology. Dr. Willow’s papers have appeared in journals such asACM Transactions on Information Systems, Journal of Intelligent Manufacturing, andIEEE Transactions on Systems, Man, and Cybernetics, among others. He is a memberof the IEEE Computer Society, ACM, INFORMS Computing Society, AIS, and theNational Society of Professional Engineers (NSPE). At present, Dr. Willow is on theeditorial board for the Journal of Management Information Systems (JMIS), MISQuarterly (MISQ), and Information Systems Research (ISR). Outside academia, hehas been an active member of the International Simultaneous Interpreters Society(ISIS) and the International Judo Federation (IJF).

A Neural Network-Based Agent Framework for Mail Server Management

Documents