intelligent agents in network management a state-of-the-art

Intelligent Agents in Network Managementa State-of-the-Art

Morsy M. Cheikhrouhou — Pierre Conti — Jacques Labetoulle

Corporate Communications DepartmentInstitut EurécomBP 193, 06904 Sophia Antipolis CedexTel: + 33 4 93 00 26 94, Fax: +33 4 93 00 26 27email: cheikhro, conti, [email protected]

ABSTRACT. Networks will soon be the keystone to all industries. However, the NetworkManagement Systems (NMS) that are currently available are not adapted to the wide spectrumof network installations and configurations. Emerging technologies like CORBA for example,do not seem to be able to solve the problems of complexity, cost and scalability. Differentstudies are on the way to distribute the intelligence to the different network components asa logical answer to these issues. Among them, the Intelligent Agent (IA) paradigm seems tobe the most promising solution. This paper presents a synthesis on current research on theIA for NM. It begins by discussing the management by delegation paradigm, and then goeson to discuss the different ways mobile agents are being proposed for NM frameworks. Theremainder of the paper addresses IA based NM approaches according to agent architectureand cooperation. Finally, the paper presents the perspectives and the potential benefits of IAsin NM.

KEY WORDS: Network Management, Intelligent Agents, Mobile Agents, Management by Delega-tion.

Networking and Information Systems. Volume 1 - no 1/1998, page 1 to 29

2 Networking and Information Systems. Volume 1- no 1/1998

1. Introduction

In the near future, all computers will be inter-connected. We will find networksnot only in world-wide companies as at present, but also inside intelligent buildingsand even inside houses where they will be used to work toasters or washing ma-chines [HUI 96]. Like other systems, these networks will need to be controlled andmanaged. Furthermore, it is obvious that personal networks(e.g., house networks) orsmall enterprise networks, will not require the same Network Management Systemsin terms of complexity, cost and necessary resources. At present, if almost all networ-kable devices are manageable (mostly via SNMP), only a few Network Managementplatforms exist, and all of these have been designed forlarge networks. As we are loo-king for scaleable, flexible and economic solutions, the emerging Intelligent Agentparadigm would seem to be a solution. Yet while hundreds of papers on IntelligentAgents are available on the Internet, reflecting the incredible untidiness of the AgentWorld, only a few of these papers have dealt with the use of Intelligent Agents forNetwork Management. This situation is now changing and the network managementcommunity is realising that “As we move further and further into the information age,any information-based organisation which does not invest in agent technology may becommitting commercial hara-kiri” [NWA 96].

In this paper, we present a State-of-the-Art report on the use of Intelligent Agentsin Network Management. This report comprises three parts. The first part consists oftwo sections, the first of which (Section 2) gives a review of network managementprinciples and issues. The second section(Section 3) provides an introduction to In-telligent Agent principles. It states the properties that IAs might have as a way ofperceiving what IAs are.

The second part presents research findings on IAs in the Network Managementdomain. The first section of this part (Section 4) deals with an important trend inNM called Management by Delegation (MbD). It focuses on workwhich has usedthe concepts of MbD to build Intelligent Agents, and discusses the improvementsneeded to reach the actual concepts of IAs. The next section (Section 5) deals with theuse of Mobile Agents for Network Management purposes. Section 6 focuses on thearchitectural aspects of IAs that are used in NM. After presenting the approaches usingeach agent architecture, the benefits and the weaknesses of the latter are presentedfrom an NM viewpoint. Subsequently, Section 7 discusses thecooperation aspect ofIAs and how it is handled in the different agent-based approaches to NM. Finally,Section 8 presents a unique application of an interface agent for network supervision.

The third and final part (Section 9) of this report is a generalsynthesis in whichsome important aspects of agency which have not been thoroughly investigated arepresented, after which new trends in Network Management consideration of Intelli-gent Agents are discussed.

Section 10 concludes this report.

Intelligent Agents in Network Management 3

2. Network Management Issues

Network Management systems have to evolve quickly in order to satisfy custo-mer requirements. One has only to read the proceedings of international symposia(e.g., [IFI97]) on Network Management to realise that thereis a strong need to moveto flexible and scaleable platforms to fulfil market expectations. The Intelligent Agentparadigm, as stated in the introduction, would seem to be a solution to these pro-blems, but while most, if not all Intelligent Agents use networks, only a few are usedfor Network Management. Only two percent of the official Agent product and researchactivities referenced by the Agent Organisation are related to Network Management.Why is this?

We believe that there are two reasons:

— Firstly, NM is still strongly influenced by OOP (Object-Oriented Program-ming), and the huge development effort made by the companiesdeveloping networksolutions is still under way. Most of these companies are focusing their strength on theuse of CORBA, which is a logical extension of their object-oriented programming.

— Secondly, if agent-oriented programming has a strong potential in all of thecomputer sciences, as object technology had ten years ago, it marks a clean breakwith classical development techniques, and has still to reach maturity.

In this context, we ask the question: is it a good idea to spendtime studying a newdevelopment paradigm? To answer this question, let us, firstof all, remind ourselvesof the main objectives of Network Management, and try to understand the reasons forthe user’s dissatisfaction with current Management Systems.

2.1. Network Management Objectives

At a first glance and without undertaking complex research, it is obvious that asnetwork end users, we expect fast, secure and reliable connections; as network mana-gers we would like to easily configure and control network access and resources; andas corporate managers we expect a low usage cost. To go further, let us look at thebreakdown of the users’ network management requirements proposed by [TER 92]and reported in [STA 96]:

— Controlling Corporate Strategic Assets: Networks and distributed compu-ting resources are increasingly vital resources for most organisations. Without effec-tive control, these resources do not provide the pay-back that corporate managementrequires.

— Controlling Complexity : The continued growth in the number of networkcomponents, end users, interfaces, protocols, and vendorsthreatens management withloss of control over what is connected to the network and how network resources areused.

— Improving Services: End users expect the same or improved service as theinformation and computing resources of organisations growand are distributed.


— Balancing Various Needs: The information and computing resources of anorganisation must provide a spectrum of end users with various applications at givenlevels of support, with specific requirements in the areas ofperformance, availability,and security. The network manager must assign and control resources to balance thesevarious needs.

— Reducing Down Time: As the network resources of an organisation becomemore important, minimum availability requirements approach 100 percent. In additionto proper redundant design, network management has an indispensable role to play inensuring high availability of its resources.

— Controlling Costs: Resource utilisation must be monitored and controlled toenable essential end-user needs to be satisfied with reasonable cost.

Because of its more or less natural heterogeneity, networksare difficult and com-plex to manage. Management systems must handle a wide and increasing number ofdevices, resources, protocols over larger and larger areas. This is why the WorkingGroup ISO/TC 97/SC 21/WG4 was formed within the ISO in the 1980s [PRA 95].But the work was not easy, and in 1987, IETF, taking a more pragmatic approach,wrote Network Management proposals for Internet management. Protocols (SNMP)and MIBs were standardised and quickly implemented on all network devices, andbecame a de facto standard. But unlike OSI, IETF does not provide Management Ar-chitecture, which makes these standards difficult, if not impossible, to apply to themanagement of wide networks.

2.2. Network Management and ISO

Network management has always been studied by the International Standard Or-ganisation which defines an architecture and a set of standards. These standards aredivided into three main parts:

— Management Communication (CMIS/CMIP) (ISO/IEC 9595...)

— System Management Functions (ISO/IEC 10164-*)

— Management Information (ISO/IEC 10165-*,10589,...)

The Five OSI Management Functional Areas:

The functional side of OSI Management is split into five strongly linked areas.Functional units are defined and used to cover these functional areas.

— Configuration Management: startup, shutdown, setting, reading, modifyingconfiguration

— Fault Management: detection, trace, location, analysis, prediction, correction

— Performance Management: measurement, data collection, analysis, tuning,testing

— Security Management: controlling access, transfer , testing


— Accounting Management: set quota, verification, planning, billing

The OSI defines how a “managing system” may communicate and act directly ona managed system which may be either another managing systemin charge of a partof the network, or directly an agent.

The management is viewed through the five areas of management(FCAPS), whichuse “management functions” to perform management activities. These functions useservices over management protocols (CMISE/CMIP) to access the managed objectwithin the managed system. The managed system is represented by a managementagent, CMIP Agent, which, in turn, is in charge of performingon the objects what hasbeen requested by the managed system, or of sending event reports when necessary.

In comparison to the Internet standards of network management which is SNMP,OSI management standard are complex and costly to implement.

2.3. Internet Network Management

Internet Network Management is a flat management based on protocols of com-munication between managers and agents to access a Management Information Basedevice (MIB), which identifies the variables that can be managed at the agent level.This simple and direct access to devices to be monitored and controlled, explains whythere is no need for special standards to describe the Internet Management Architec-ture. But the other side of the coin is that SNMP cannot be usedto manage widenetworks for the following reasons: an intensive use of polling which consumes lotof bandwidth; the problem of packet size for retrieving MIB data; lack of security;the impossibility of having an intermediary network manager (manager to manager).Therefore, IETF has been obliged to propose an improved version of SNMP to solvethese problems. The main extensions of the far more complex SNMPv2 are as follows:

— Protocol: New primitives GetBulkRequest and Inform Request

— SMI extensions: New data types, table categories

— Manager to Manager capability: Inform Request, Get Response

The second improvement proposed by IETF is an RMON specification developedto perform pro-active LAN monitoring on local or remote segments. Both SNMPv2and RMON2 are a step towards distributed management, but theprinciple is still tohave a centralised management platform. Even if there is no architecture provided,and only low level operations, the Internet NM is far and awaythe most widely usedbecause of its simplicity.

2.4. Network Management Systems: The Market

All networks are different. The different components - topology, media, architec-tures, hardware and software have together resulted in an explosion of network im-plementations. That is why it is almost impossible to find an evaluation of NetworkManagement systems available on the market.


Some trials have been done [GRE 96] but they are limited and even if the test bedis relatively simple, the results have not been very encouraging. Several products arebased on the Open View platform completed by ISV (Independent Software Vendor)partner applications that may be executed from the GUI (Graphical User Interface).

In his conclusion, I. G. Ghetie [GHE 97], who provides us witha study of the mainmarketed NM products, observes that there is a lack of cooperation and integrationbetween Network Management applications even within the same platform.

ConclusionWithout going into detail, one may retain the followingpoints:

— None of the products are able to fully cover all the network management areas.Their main focus is on network monitoring and event reporting.

— None of the products are able to easily manage heterogeneousnetworks (be-cause of the lack of a useful standard).

— Scaling is difficult due to the inadequacy of the applicationdevelopment tools.

— The systems are very resource consuming (e.g., SNMP polling).

— All are expensive.

Customers are still looking for greater simplicity, for better integration and col-laboration between the different applications, and for scaleable and inexpensive plat-forms.

2.5. Intelligent Agents and Network Management

It is therefore necessary to review the way we design NetworkManagement Ar-chitecture, bearing in mind that we will have to use intelligent entities to do the work.Some people have proposed new ways of developing Network Management Architec-tures: for example Aiko Pra [PRA 95] who presents a kind of prototyping approachto Network design, or D. Stevenson [STE 95] who starts from the “Help Desk pointof view” to gather the most important NM Requirements and whoemphasises thatArtificial Intelligence needs to be applied at all management levels.

Studying the great potential of Agents, we feel closer to this last approach, as webelieve that we have people who are able to perform activities, or who may acquireskills to perform them, and that we will have to organise the NM activities aroundthem. The Network Management Architecture becomes then a problem of organisa-tion, where the problems encountered by the field services, or help desk are taken inaccount.

Several studies have been carried out around what is called the Management Po-licy, which is based on the Object-Oriented paradigm. However, the prototypes thathave been developed are too much influenced by OOP and seem restrictive comparedwith what may be done with Intelligent Agents. For example, the Department of Com-puting at the Imperial College in London, is working on “Authorisation-Obligationpo-licies” [MAR 96], and a framework for distributed systems management based on roletheory [LUP 97]. The interest of these studies is the possibility they present of pro-viding an architectured approach of Network Management with Intelligent Agents,while maintaining the simplicity supplied by the Internet management protocols.


3. Intelligent Agent Principles

3.1. Origins

There is no doubt that we are approaching the time when we can expect helpfrom software systems that are capable of taking decisions and working independently,without being under our permanent supervision. During our research, we found a lotof different Agents: Software Agents, Intelligent Agents,Mobile Agents, or PersonnelAgents also called Personal Digital Assistants (PDA). At the time where the mainsoftware corporations are focusing their strengths on the Object World, why is theremore and more interest in these Agents? Reading the many publications and studieson this subject, we can find three main domains to explain thisphenomenon.

1. Artificial Intelligence (AI), seen for a long time as a marginal software activity,providing complex and difficult software to develop, has been somewhere influencedby the Object-Oriented approach [LAB 93]. Peolpe working in AI have understoodthe interest of working on simpler, more flexible and re-usable software entities. “Thegroup of agents is more than the sum of the capabilities of itsmembers” says J. Mullerin [MUL 96].

2. Object World after the integration of the network environment with the deve-lopment of distributed Object Architectures, had a strong need to go a step further byproviding a certain “autonomy” and “mobility” to its Objects. New object orientedlanguages have been conceived, giving mobility to the objects, and the name agentappears with the following definition: “An agent is a computer program that acts au-tonomously on behalf of a person or organisation.” [OMG 97].

3. Networks, which have been using software called “agents”for Network Mana-gement purposes for a long time, even if these agents do not have strong relations withthe current definitions of Intelligent Agents. The SNMP and CMIP agents are softwaremodules that gather and record management information for one or more network ele-ments and communicate the information on request from a monitor using SNMP orCMIP protocols. There are experiments to replace simple SNMP or CMIP agents byintelligent substitutes, [PAR 95] but without a global review of management prin-ciples this is unlikely to resolve the current issues. On theother hand the expansionof networks has given a new orientation to the classical way of using computers. Theresources are now distributed among different and, most of the time, heterogeneoussystems, and may be shared. In the other direction, information is now spread out onnetworks, and should be gathered. The standard client/server architecture based on asingle server should be reviewed. The agent approach is now being taken into accountand is seen as THE solution by companies like Oracle with their client/agent/serverarchitecture.

Even if it is not yet recognised, all these domains should converge, because they areusing the same networked environment, and because economical reasons will dictatethat what has already been designed does not need redesigning.


So far, the multiple origins of the Agents make the task of reaching a consensuson an Agent definition a difficult one. As an example, during the ATAL’96 workshop“Agent Theories, Architectures, and Languages” a lot of time was spent discussing thedefinition and taxonomy of Agents. Mario Tokoro gave us a resolute definition in thetitle of his paper “An Agent is an Individual that has Consciousness”. But the generalconclusion is perhaps that there is no urgent need to come up with a unique definitionof the Intelligent Agent.

Thus, in this study we will not join this interesting debate,but on the contrary wewill be using the Intelligent Agent as a set of capabilities,also called properties, andwe will use the word “Agent” as an abbreviation for “Intelligent Agent.”

3.2. Agent Properties

The first aim of this approach by properties was to make a classification of Agentsthat we met in the course of our research. However, it soon became clear that this willnot be useful as we explained in the previous section. Nevertheless, it is useful to knowwhat properties Agents may own to understand why they will change our life, and wepropose below a description of these properties.

— Autonomy“Self-government, independence: Branch managers have full autonomy in their ownareas” (Oxford Advanced Learner’s Dictionary). The agent decides himself when andunder which condition he will perform what actions. “An autonomous agent is a sys-tem situated within and as a part of an environment that senses that environment andacts on it, over time, in pursuit of its own agenda so as to effect what it senses in thefurture” [FRA 96]. M. Wooldrige disagrees with Franklin andGraesser’s definition ofagent, and in “A response to Franklin and Graesser” the autonomy is explicitly requi-red as property, but also the reactive, pro-active and social behaviour. One questionthat we have debated for a long time is: “What is the autonomy without pro-active orreactive or deliberative behaviour?”. Someone suggested that viruses and “networkedviruses” (also called worms) were example of autonomous, and mobile agents.

— CommunicationOne of the key properties of agents is the ability to speak with a peer, with a human(Interface Agents), or with a device. The following communication between agents,called languages, are often used:� Blackboard: Agents read and write messages in a shared location (blackboard).� KQML: Knowledge Query and Manipulation Language is a language and proto-col for exchanging information and knowledge using “performatives”.� KIF: Knowledge Interchange Format.� COOL: structured conversation, KQML-based, used for the coordinationof agents[BAR 96].

— Collaboration/Cooperation Agents are collaborative when they are able towork together. Also the agent is able to communicate and negotiate with others, he


is deliberative and may coordinate his actions with others.Collaborative agents areparticularly useful when a task involves several systems onthe network. Negotiationis the main issue for collaborative agents. While coordination can occur without col-laboration, collaboration needs negotiation.

— DeliberationKnow rules, and apply them without waiting for instructions. M. Wooldrige and N.Jennings define a deliberative agent as “one that contains anexplicitly represented,symbolic model of the world, and in which decisions (...) aremade via logical (or atleast pseudo-logical) reasoning, based on pattern matching and symbolic manipula-tion” [WOO 95].

— MobilitySince Java appeared, we may find a lot of mobile Agents, but there are different kindsof mobility:� The mobility which allows the agent to move from one system toa similar one� The mobility which allows the agent to move to another different system� The mobility which allows the agents to suspend their actionon one system, moveto another and go on.� The mobility which allows the agent to move himself, rather than being transpor-ted.� The mobility which is a duplication of the agent to another system (cloning).� The agent carries his knowledge to another system.

Generally, mobility turns out to be a mixture of these definitions. One of the mainissues around mobility is the potential security weakness of mobile agents.

— LearningThere are at least two definitions of learning:� The bad one: an agent is said to learn if he is able to acquire knowledge (data).� The good one: an agent is said to learn if he is able to use his new knowledge tomodify his behaviour.

Despite the fact that learning is an important factor of intelligence, only a few agentsare able to learn. Most of the time they have fixed (pre-compiled) rules and knowledgebases.The objective of learning is for the agent to perform new tasks dynamically withoutbeing stopped. Different ways of learning are studied and experimented:Generalisation: you observe your environment and deduce rules. Instruction: you ob-tain knowledge and rules from others (transfer).

— Pro-activeness“Pro-active actions are intended to cause changes, rather than just reacting to change”.Pro-active agents generally follow plans, or at least execute rules when the environ-ment reaches a known threshold.Sometimes pro-active is used with the same meaning as deliberative, but an agent maybe pro-active, because he has been requested to perform pro-active tasks, as opposedto deliberative agents, who decide themselves to be pro-active.


— ReactivityDo something when an event occurs.

— SecurityBe able to discriminate friends from enemies and contaminated elements.

— PlanningThe agent organises by priorities the actions to perform during his life. For manyresearchers planning is one of the most important properties for an intelligent agentto possess. Planning is used by deliberative and pro-active agents according to theirknowledge of the environment and the possible actions that they can apply to it.

— DelegationAn agent may ask someone else to perform one of his goals or tasks. This capacity isvery important for balancing resources.

Conclusion

The above descriptions or definitions have been subject to interesting internal de-bates. As may be seen from the published papers, because there is a lack of a world-wide consensus, it is of pri me importance to have an internalagreement and unders-tanding on the terminology used within any projects.

3.3. Agent Technologies

The development of agents requires languages to build the agents, protocols andlanguages to give software the possibility to communicate,and languages to describeand transfer agents’ Knowledge, Beliefs, Goals, Desires, and Intentions [MUL 96].

3.3.1. The Birth of an Agent

The main idea for our discussion on Network Management, is the possibility foragents to exist in different heterogeneous environments. Thus, they require a standardinfrastructure on each system where they need to be hosted. Then agents may be deve-loped as if they will always be on the same machine - the Virtual Machine (VM). Theeasiest languages to create an “agent” are script languagesthat may then be executedon remote systems, because they are interpreted as shell languages on Unix systems.The most known are Telescript from General Magic (replaced now by Odyssey whichis 100% Java), Perl and Tcl (Tool Command Language) from Sun, and, of course,Java which is an Object-Oriented scripting language. The second class of languagescontains languages coming a classical programming approach as LALO (Agent Orien-ted Language) which generates C++ sources, or AI oriented languages as Obliq fromDEC. Depending on the languages and tools used to create them, Agents are born withmore or less usable “intellectual capabilities”.


3.3.2. The Agent’s Voice

Once the agents are created, they need to communicate with humans or with fel-low agents. Depending on the techniques used to create them,different solutions areavailable, generally TCP/IP based, communication which mayhave several layers andwhich at the end are used to transfer the information. KQML (see Section 3.2) is themost often used language to encapsulate requests and answers between agents overprotocols like IIOP, HTTP, and SMTP. We may here compare the different ways weuse to communicate ourselves, e.g., mail, e-mail, phone,. ..

3.3.3. The Agent’s Knowledge

Now the knowledge transported is described using another layer of language, thatwe may compare with our respective languages such as French,English, Portuguese,German,. . . A commonly used knowledge language is KIF (Knowledge InterchangeFormat), but specific languages are usually developed within projects for specific pur-poses (e.g., GDL for MANIA).

It is sometimes not well understood that the Intelligent Agent paradigm uses ahuman view and speech (Anthropomorphism) to present what isonly software. Thisapproach is nevertheless normal and accurate, because we are now at the stage wherewe can use Intelligent Agents as if they were human workers, giving them responsi-bilities and autonomy in their actions. In the next sections, we review the experimentsusing agents to solve the problems encountered by the classical Network Managementapproaches:

— ScalabilityWhen systems are growing or new services are added, Agents may take in charge thenew environment, by changing theirs roles, beliefs, organisation.

— RobustnessThe fully distributed versus centralised Network Management is more robust becauseNM operations may continue even if part of the network is disconnected from theMaster management site.

— UpgradabilityA new Agent may replace the oldest one without stopping theiractivities.

— PerformanceCloser to the systems to manage, distributed management allows faster reactivity andeven pro-activity, and less resource consumption.

4. Intelligent Agents Through Management by Delegation

4.1. Preamble

The management by delegation (MbD) paradigm has been developed without ini-tial consideration of intelligent agents. However, we havedecided to include MbD inthis report for the following reasons.


— Recently, a lot of work which reconsiders MbD is being conducted under thebanner of intelligent agents.

— We believe that MbD is an important issue that can foster further discussionabout intelligent agents in network management.

— Some researchers believe that intelligent agents could be developed by usingMbD concepts.

Next, we will respectively present the reasons for introducing MbD, its concepts andprinciples, the works on “intelligent agents” using these concepts, and finally a syn-thesis and discussion.

4.2. Background

People working in the NM field have quickly identified deficiencies in classicalNM architectures and protocols. Management data is distributed among agents withvery limited processing capabilities, while data processing to extract management in-formation is centralised in managers. The latter consequently suffer from a heavyprocessing load. Furthermore, management operations mustbe decomposed to veryprimitive functions mostly limited to SET- and GET-like primitives and then com-municated to the agents. This is referred to as micro-management [YEM 91]. Micro-management leads to heavy management traffic in the network.

To overcome this uncomfortable situation, and in order to define a more flexibleand scalable architecture with real-time management capabilities, it is tacitly admittedthat management functions and operations should be dynamically carried out close towhere the managed objects are. Management by delegation is designed following thisprinciple.

4.3. Management by Delegation

The main concept on which MbD relies is that of anelastic server[GOL 93]. Anelastic server is an enhanced server whose functionality can be extended and contrac-ted dynamically during run-time. A delegation protocol allows a client to pass newfunctions to an elastic server and then asks the server to execute these functions byinstantiating them. The protocol also allows control to be kept on each instance sothat it can be stopped then restarted for example. New functionality is prescribed in ascript with no restrictions on the language in which it is written.

Applied to network management, elastic servers can be interpolated between ma-nagers and agents (in a similar way to middle-level managersin organisations) withsimilar roles to proxy agents [STA 96]. These elastic seversare then called MbDagents [YEM 91], MAD agents (Manager Agent Delegation) [YEM91], managingagents [TRO 97], delegation and flexible agents [MOU 96] or elastic agents [GOL 95].


An alternative approach would be to make classical SNMP/CMIPagents evolvetowards elastic agents. But this approach has been rather discredited. Indeed, manu-facturers usually tend to keep the legacy systems as they areand to design ad hoccomponents instead.

The manager uses the delegation protocol to upload management scripts to theflexible agents. Using the same protocol, it asks permissionto instantiate some de-legated script and to run it. Therefore, the management operation prescribed in thisscript can execute “autonomously” in the same environment as the managed objects.The manager is able to control its execution via dedicated primitives in the delegationprotocol.

Later, Goldszmidt [GOL 93] called these scripts “delegatedagents”. Under thebanner of intelligent agents, some research is being conducted to explore the MbDparadigm for NM purposes. For example, [KAL 97] suggests a spreadsheet scriptingenvironment for SNMP. The spreadsheet scripting language allows a manager to pres-cribe computations that can be carried out by the agent. Eachcell in the spreadsheet isdefined by an expression that computes a value from other given data such as values inthe MIB. Expressions can be inserted, updated and deleted according to the managerneeds.

A further improvement is depicted in [SUZ 97]. Each agent contains functional ob-jects with network management functions allowing access tothe management objectsand communication with other agents. Therefore, a manager needs only to delegatescript skeletons to invoke these management functions. When running an instance, themanagement functions’ invocation is bounded to those implemented in the functionalobjects. The same management function could be implementeddifferently on differentagents according to the network device specifics for example.

Other works may use different scripting environments such as the SQL-like onepresented in [ZNA 96], event-driven scripts presented in [KOO 95] in its divide andconquer approach and the Tcl based scripting language in [GRI 97].

Alternatively, Trommer and Knonopka [TRO 97] chose to add a rule processingunit to agents. Rules can be transmitted from managers to agents, and can endowthe latter with intelligent behaviour. Rules can be dynamically loaded, updated andremoved, making the agents flexible.

Referring to the agent definition and properties in [WOO 95],Mountzia [MOU 96]suggested evolving delegation agents into intelligent agents by making them autono-mous, pro-active, reactive, mobile and able to learn. As they are presented in thatpaper, these ideas are still at an early stage of development.

4.4. Synthesis and Discussion

The implementations of the Management by Delegation concepts presented aboveare far from the agency’s concept though they are presented under the banner of in-telligent agents. With reference to the properties of software agents presented in Sec-tion 3.2, none of these really exist in these implementations. The potential properties


that are claimed to exist are autonomy, delegation and cooperation. However, delega-tion agents are not autonomous since management scripts must specify exactly whatthe agent should do. Agents cannot perform tasks without being ordered to do thesetasks. They cannot decide on their own what, when, and how to do. These same criti-cisms apply to delegation: only low-level actions can be delegated. Finally, agents arenot cooperative, in that they do not dynamically decide to share an important goal inorder to achieve it in a coordinated manner.

Nevertheless, the management by delegation concept is tacitly powerful and can beexploited in several ways. It can be the basis for supportingintelligent agents conceptsin one of two possible ways.

In the first way, one may follow the direction of [MOU 96] and conceive delegationagents as software agents instead. The second way consists in delegating intelligentagents as management scripts. This leads to remote-execution mobile agents. In bothcases, further efforts, improvements and considerations must be addressed. In fact, anagent-oriented view of delegation will shift from the delegation of scripts and func-tions to the delegation of goals which is a high-level delegation.

5. Network Management with Mobile Agents

5.1. Generalities

Mobile agents (MA) are able to hop from node to node during their activities.Therefore, they can use the information they found during previous visits to sites inorder to adjust their behaviour accordingly. They are also able to carry out activitiesclose to where the needed resources are found. In the following sections, we presenthow these properties could be used in Network Management andthen discuss theresearch presented.

5.2. Advantageous Ways of Using MAs

In [MAG 95], Thomas Magedanz has presented issues for the deployment of MAsfor the purpose of NM activities. MAs can encapsulate management scripts and bedispatched on demand where needed. An agent can be sent to a network domain andtravel among its elements collecting management data, and return with the data filteredand processed. Sending a MA for this task is a substitute to performing low-level mo-nitoring operations and processing them centrally. Then ifthe agent is able to extractuseful and concise information from raw data collected on each element, the agent’ssize can remain small enough to save bandwidth usage. In fact, this is one of the mostadvanced arguments for promoting MAs.

Similarly, an MA can also be used to carry out management operations. Insteadof remotely invoking management operations, the administrator can encapsulate theseoperations into a mobile agent and send it where needed. If these operations involve


a lot of message exchange along with human decision making, then the mobile agentcan also save bandwidth by locally taking the necessary decisions while performingthose operations.

These ideas have been implemented by FTP Software [SOF 97]. The IP Auditoris an application that dispatches mobile agents in the target network in order to col-lect management information such as configurations and network elements’ operationstates. The IP Distributor is another application that dispatches agents that distributemanagement payload.

Another original and concrete application of MAs is presented in [APP 94]. Mo-bile agents are used to control traffic congestion in a circuit-switched network. A firstclass of MAs, called parent agents, randomly navigate amongthe network nodes andcollect information about their utilisation. By keeping track of this information, theyare able to gather an approximate utilisation average of thenetwork nodes. Therefore,they are able to identify which nodes are congested relativeto this average. When acongested node is identified, a load-balancing MA is created. It has to update the rou-ting tables of the neighboring nodes (using an optimal algorithm) so as to reduce thetraffic routed by the congested node.

In a comprehensive study on code mobility, [BAL 97] presentsa consequent ad-vantage of using mobile agents; when the network administrator can only be connec-ted via an unreliable, costly or lossy link, he can create itsMA off-line, connect tothe network to dispatch the agent, close the connection and then reconnect later toget his agent back with the results. This principle is actually implemented in Astrolog[SAH 97] to support mobile managers; in order to have a network flexibly managed,an administrator can connect to the network using his portable computer. The connec-tion could even be established on the basis of a GSM link. If a hand-over occurs, thedispatched MA can retrieve its sender’s new location and return to him.

[MAG 95] and [MAG 96] suggest other possibilities of applying MAs. A serviceor a network provider can send MAs to user end-points in orderto adapt his equipmentto new services. These agents can also achieve other tasks such as user accounting andcapturing user requirements. [HJÁ 97] also suggests how to use MAs for immediateservice creation and customisation.

5.3. Discussion

The use of MAs seems to be rosy for NM Applications. However, some precau-tions must be taken before heavily investigating in this direction. Concerning band-width saving, it is not obvious that carrying an agent with high capabilities consumesless bandwidth than carrying the desired information and processing it centrally. Ac-tually, management information processing (e.g, for eventcorrelation or fault sourcedetection) is often complex and difficult to achieve. Unlessthe agent is described ina high-level and concise language, it will be of a large volume. This leads to twooptions. Either the Agent run-time environment would be complex and heavy, or theagent has only limited capabilities. The first option leads to costly implementation of


an MA platform while the second may simply be useless. Yet thedesign concept ofa run-time environment for mobility which supplies the agents with a homogeneousinterface across heterogeneous systems is not a straightforward one. The Java virtualmachine is already a step forward, but this is still far from the agent concept.

Finally, security is one of the most challenging problems that MAs may face. In asensitive domain such as NM, security must be carefully studied. [SOF 97] has alreadyclaimed that the IP Auditor’s and IP Distributor’sagents (see the previous sub-section)support security mechanisms.

Therefore, MAs, though presenting an attractive approach,must be carefully in-vestigated and deployed. In particular, using MAs for predetermined tasks such asin [APP 94] is needless though elegant. Instead, a protocol for exchanging utilisationinformation and traffic routing could have been used substantially.

6. Reactive, Deliberative and Hybrid Agents in Network Management

6.1. Preamble

According to its architecture, an agent might be reactive, deliberative or hybrid[WOO 95]. While a reactive agent acts in situation-reactionor event-reaction manner,a deliberative agent acts according to decisions taken after a reasoning process aboutthe environment. A hybrid agent is a combination of both. Thefollowing three sec-tions will respectively treat these types of agents, and thelast section will provide asynthesis and a discussion.

6.2. Reactive Agents

The application presented in Section 5 used Mobile Agents for congestion mana-gement. The agents used therein were reactive agents [APP 94]. Both the parent agentsand the load agents do not use any model of the network they aremanaging. Instead,they only observe the state of the network nodes and react accordingly. The use ofReactive Agents is particularly advantageous for this application because MAs needto have a light weight so as not to overload the network by their mobility. The parentagents do not react unless they discover an overloaded node.The decision to react bycreating a load agent is taken by simply comparing the currently visited node utilisa-tion with the average utilisation computed during its traveling. The load agents whichupdate the routing tables decide on the basis of an algorithmwhich doesn’t imply anyreasoning. Each agent taken alone, be it a parent or a load agent, would not be ableto manage the network. The application is therefore a good example that shows howthe overall multi-agent system goes beyond the capabilities of all the reactive agentswhen taken individually.

Other research concentrates on building agent-based frameworks for network ma-nagement and use the reactive architecture mainly for its simplicity. In [STE 96], in-


telligent agents are applied to implement distributed management policies. Within thisframework, agents are composed of a communication mechanism, a rule base, a solu-tion base and an inference engine. The rule base contains rules that matches perceivedevents (corresponding to anomalies) with the reactions to be taken. Reactions (calledservices in the paper) are also described by rules within thesolution base. Therefore,the reactive rules are of the form<event-Id,reaction-Id>. This frameworkwas used to achieve usage-based accounting.

Another framework is described in [SKA 96]. A generic agent architecture is de-fined and built using April++, an Object-Oriented extensionof April [MCC 94]. Theagent is organised in units which communicate (inside the agent) via a blackboardmechanism. The units are organised vertically which means that they execute concur-rently. The local information of the agent is centralised inan information database. Atthe same time, the agent has a link mediator which is an interface with the physicalcomponents of the network, and a head that insures the communication with otheragents. New functions can be added to the agent by acquiring new units. This modelhas been applied to provide agents suitable for multi-service network management.Indeed, for this to be achieved, the agent is endowed with a customer unit to interactwith the user, a service unit that honours service requests and a fault handling unit thatdeals with the faults that may occur.

6.3. Deliberative Agents

Two applications can be stated as deploying purely deliberative agents. The firstis that described in [WEI 92]. The paper describes an interesting framework appliedto control a VPN service. Here, agents are used to automate thenegotiation that usedto occur between the service provider and the service users when these ask to changethe service parameters or to repair network faults. A customer agent has knowledgeabout the logical structure of the VPN and its usage, while theprovider agent knowsboth the logical and physical implementation of each customer’s VPN. Thus, agentshave a model of the external world on which they act. Going into detail, a customeragent knows for each trunk (or logical link) its capacity andits utilisation, whereasthe provider agent knows for each trunk the physical links onwhich it is implemen-ted. When, let us say, a network fault occurs, the VPNs based onthe faulty networkelement are affected. In this case, the customer agent first asks the provider agent torepair the fault. In many cases, a complete and immediate repair is not possible anda negotiation-based cooperation is started. Customer agents try to find intermediatesolutions based on their knowledge of the utilisation of their respective trunks, andsuggest partial solutions to the provider agent by updatingthe logical structure oftheir VPNs. The provider agent coordinates these solutions and makes the necessaryupdates in order to reach an accepted configuration. Both customer agents and pro-vider agents use logical reasoning based on their models of the network and on theirbeliefs on its elements. The implementation of this reasoning is, in fact, based on aplanner called PRODIGY.


The second application of deliberative agents in NM is the MANIA project (Ma-naging Awareness in Networks with Intelligent Agents) [OLI95]. An agent-based ap-proach is used to manage the quality of service in the network. Agents are structuredin beliefs, desires, intentions, goals and commitments. Anagent’s beliefs express itsperception of the environment. A first part of the beliefs describe a real-time state ofthe network, e.g., the printing service is highly solicitedat the current time. A secondpart contain the historical behavior of the network. The historical behaviour is used toachieve some kind of learning about the dynamics of the network such as deducingthat the printing service is highly solicited every day from10 to 11 am. A third partof the beliefs translate the states of the services providedover the network, e.g., theminimal response time the NFS server can ever have, or the maximum client numbera web server can handle at a time. Finally, the agent may have beliefs on the userapplication contexts i.e., user requirements in terms of QoS.

The agent’s wishes consist of two parts. The first part corresponds to the requeststhat the agent could not satisfy, such as when a certain user requires a video connec-tion while there is no available bandwidth (according to theagent’s belief). The secondpart consists of motivation policies. The network administrator may want to “motiva-te” the agent to give a certain priority to some project members because they have aconstraining deadline.

When receiving a user-application context (which describes the user needed re-sources and the required quality of service parameters), the agent translates it into aset of goals. Goals are independent from the system’s state.For example, a goal maystate to “actively monitor the NFS server”. These abstract goals are then mapped intointentions. Intentions take into account the current stateof the system, the availablemeans to achieve the goals (e.g., MIBs, testers, etc.) and the possible constraints thatmay apply.

Finally, and as the administrator could motivate the agent,it is also able to specifyobligation policies which form the commitment part of the agent.

6.4. Hybrid Agents

Here also two applications are worth presenting. The first ispresented in [GIB 97].The overall project addresses the application of intelligent multi-agent systems to ma-nage connection admission control in ATM. An ATM network is very dynamic andmanaging it centrally will not be appropriate because managed data goes quickly outof date. But local management is not appropriate because it lacks an overall view ofthe network. Here, the hybrid agents are particularly interesting. They may combineboth local real-time management via the reactive behaviour, while the deliberative be-haviour is concerned with cooperating with other agents in order to achieve globalplanning and coordinate high-level tasks.

However, since the project is still in its infancy, no further implementation detailsare given.

In contrast, the second application is already well designed and the agent’s ar-


chitecture is well defined and is based on the concept ofVivid agents. Vivid agents[WAG 96] comprise both a reactive and a deliberative part. However, the reactive partof the agent is not hardwired within the agent. Instead, the deliberative part may dy-namically change the reactive behaviour.Reagentsare particular vivid agents with noplanning capabilities.

Reagents are applied to perform a distributed diagnosis on distributed systems[FRÖ 97]. The network is partitioned into physical domains,each domain having itsown diagnostic agent. The agent has a detailed knowledge model of its domain andminimal information about the neighboring domains, mainlythe address of their res-pective agents. The multi-agent system applied a distributed version of the ModelBased Diagnosis. The principle is that when an agent detectsa fault, a lost connectionfor example, it starts by performing a local diagnosis of itsdomain. This is the delibe-rative behaviour of the agent, since it performs reasoning according to a model of itsdomain. If it finds no local fault, then it sends the resultingobservations to the neigh-bor agent. The latter performs the same procedure. If it findsthe faulty element(s), itreports it to the first agent. If the first agent itself had received the observation fromanother agent, then it must ensure that the other agent is informed by forwarding thereport. This is a reactive behavior of the agent where no reasoning is performed.


The first thing to note is that few papers supply the reader with information aboutthe agent architecture. There are several possible reasonsfor this. The first reason isthat many papers only present scenarios and have not yet reached the point of desi-gning multi-agent systems. An example of such papers are [MAG 95] and [MAG 96],[RIZ 95] (though the reader may implicitly understand that agents may have goals)and [INT 97]. The second reason is that some frameworks support both architectures,and that implementation has not yet started. An example of this is [SOM 96]. Finally,this could be due to ignorance of agents’ architectures which is rather common forexample in the mobile-agent community.

What is interesting to note, is that reactive agents are suitable for local and real-time management operations. They can be used efficiently in fixing well-known faultswithin a relatively small portion of a network. An approach based on relating symp-toms to faults, and faults to fixing methods is broadly adequate for this purpose.

On the other hand, deliberative agents are much stronger from a problem-solvingand logical reasoning point of view than reactive agents. This makes them requiremuch more resources, and makes their response time rather long. For this reason, theyare suitable for complex but not time-constrained tasks. A good example would bethat of analysing the origin of a security failure in a large corporate network.

Furthermore, the idea of [GIB 97] of using a hybrid agent to take advantage ofboth architectures is very interesting. The concept of vivid agents is a good basis forimplementing this idea. Vivid agents offer an increased flexibility when designing anagent since their reactive behaviour could be updated during run-time. This could be


applied for example to support complex learning algorithmswhere the agent, still ha-ving timely reactions, has its behaviour improved in time. However, one may keep inmind that the same benefits could be obtained using a heterogeneous architecture, i.e.,a multi-agent system with both reactive and deliberative agents with the latter gover-ning the behaviour of the former (See [NWA 96] for details andfurther references).

7. Cooperation in Multi-Agent Systems for Network Management

Cooperation is one of the most important properties in MASs [DOR 97]. In NM,cooperation could have large benefits in Distributed Network Management Systems.Though the term “cooperation” (or “collaboration”) is not clearly used, we will tryin the following subsections to describe cooperation in agent-based network mana-gement approaches, progressively starting from no, or simple, coordination towardsmore elaborated cooperation.

7.1. Coordination Without Communication

As described in Section 5, the MAS used in [APP 94] is based on agents whichdo not directly communicate. An agent may be aware of the other agents only byobserving the system’s state. Each time a parent agent visits a node, it leaves thereinsome information such as its identity, its age and the date ofits visit. By reading thisinformation in every visited node, the parent agents are aware of their number in thenetwork. If their number is excessive, then the youngest parent agent will terminate.Furthermore, in determined nodes, static processes continuously check if all the parentagents are still alive and did not crash. If a very long periodhas elapsed from the lastvisit of a certain parent agent, then the static process willconclude that this parent hascrashed and will replace it.

In addition, when a load agent is launched on a node, it creates a record with itsidentity and its start time. When it finishes updating the routing tables, it returns tothe first node and registers its same start time again and thenterminates. Therefore,any parent agent that visits that node is able to know whetherthe load agent is stillactive or not. It is also able to detect that the load agent hascrashed if the start timehas not been written for the second time after a long period. The parent agent becomesresponsible for creating another load agent to replace the first.

When there is no direct communication between agents, a lot of coherence pro-blems are avoided. Firstly, an agent would not send messagesto another agent thathas crashed. Besides, all the agents equally share the same information and no mecha-nisms are needed to preserve coherence. Secondly, the agents are able to coordinatetheir tasks in a simple way. These properties gave the MAS thus designed a highdegree of robustness and graceful degradation which are rare (actually never) to beconsidered in the other applications.


7.2. A Primitive Cooperation

In Section 6, we presented an agent-based system implementing a distributed ver-sion of Model-Base Diagnosis [FRÖ 97]. In case of a fault, an agent performs localdiagnosis and if it finds no faulty element, it asks the next agent to perform diagnosisin its corresponding domain. One may consider this as a simple cooperation for atleast two reasons. The first reason is that the underlying communication mechanismis itself primitive. The second is that this cooperation is hard coded, and the agentshave not decided how to carry it out. However, and referring to [DOR 97], agents arecooperating since they participate in achieving a global (though implicit) goal whichis to identify the faulty element in a large network.

7.3. Self-interested Negotiation Agents

Self-interested agents do not need to know about the other agents’ plans and goals.Instead, they work in order to achieve their own goals in the best way without caringabout the others. However, in order to achieve a goal, a self-interested agent mayneed the services of some other agents. A natural approach would be based uponnegotiation. Here, when requiring certain tasks to be done by another, the agent triesto maximise its profit by launching a kind of auction. A good model for this is theeconomic market place [WEL 93].

Self-interested negotiation agents seem to be very suitable for some network ma-nagement activities, especially for service management. Most of the papers dealingwith service management refer to this kind of agents. In [MAG95], Thomas Mage-danz suggests that these agents are adequate for the provision of high-level servicesin an open electronic market of telecommunications services. A good basis for co-vering this trend is the FIPA application draft presented in[INT 97]. The applicationdefines three kinds of agents, namely the Personal Agent (PA),the Service ProviderAgent (SPA), and the Network Provider Agent (NPA). The latteris responsible forthe provision of the network resources and elements which are necessary for the ser-vice implementation. Based on these network resources, theSPA is responsible forthe provision of the services with the expected quality. ThePA is a kind of PersonalDigital Assistant [MAE 94] that assists the user in defining his requirements for theapplication’s needs with regard to the user’s preferences.It then has to negotiate theserequirements with different SPAs in order to find out the bestprovider with the bestservice in terms of maximum quality and minimum cost. When a communication isabout to be established, the PA has to commit the local resources of the user and toconfigure his equipment according to the established connection.

The SPA has to catch the user’s requirements and to identify the necessary servicesand to map them into these services’ parameters. It then negotiates with the NPAs toselect the best network provision, again in terms of maximumquality with minimumcost. Feedback with the PA may occur in order to conclude withthe best arrangementboth with the Network Provider and the User.


Finally, the NPA gets the SPA’s specifications and translates them into networkrequirements in terms of bandwidth, jitter, etc. It may alsohave to negotiate withother NPAs, mostly in the case where multiple network domains are implied.

Other works on the subject use the same, or otherwise similar(restricted) ap-proaches. Mike Ruzzo and Ian A. Utting [RIZ 95] have also introduced User Agentsto allow for the definition of customised services. New services can be customised bysupplying the agent with the user’s policy. The UA is then able to make the best choicebetween what can be achieved as suggested by the fall-back ofthe called agent, andwhat the service provider offers. The caller and the called agents may also negotiatein order to reach an agreement that satisfies both users’ policies i.e., preferences andconstraints. To support user mobility, end-point agents are also introduced. Mobile-user agents have to negotiate with end-user agents to get thepermission to access thecorresponding end-points and establish communications.

Other scenarios suggest implementing new services in agents. [HJÁ 97] suggeststhat while services are user-customised by PAs, new services can be created withinservice agents in a service-level agent environment. Service agents encapsulate net-work specifics and each one is dedicated to one service that can be constructed overother service agents. Similarly, [MAG 96] presents a scenario where mobile agentsare sent into an agent environment on the services’ host, where they can offer new ser-vices by using the legacy services and/or by exploiting other mobile agents’ facilities.In both scenarios, the agents that implement new services have to negotiate with thealready existing agents in order to find the agents on which the new service is bestimplemented.

In [GIB 97], auction-based negotiation is suggested to match resource providers toconsumers in ATM networks. This is a simple yet efficient protocol by which decisionsare taken in a timely fashion by “matching highest bidding buyers to lowest biddingsellers”.

Negotiation is also adopted in [SOM 96] for managing an ATM network. Agentsare organised within a hierarchy of authorities. Each authority is responsible for somedelegated resources and exports “performance indices”. Therefore, when an authority(particularly, a service agent in the authority) receives auser request, it commits thenecessary directly managed resources and then decides which of the sub-authorities isbest eligible to route the request through. This decision isbased on a tendering processusing the above-mentioned performance indices.

7.4. Delegation-based Cooperation

As opposed to the MbD paradigm presented in Section 4 which only deals withscript (or at the best functionality) delegation, we address here the delegation betweenintelligent agents. This is a higher-level delegation because agents delegate goals andmotivations instead of low-level operations.

The agent framework described in [STE 96] (see Section 6.2) makes use of policiesexpressed in logical rules. High-level policies are delegated to agents. These policiesare then mapped into lower-level policy rules and stored in the agent’s rule base.


The MANIA project (see Section 6.3) also makes use of high-level delegation ofgoals. An agent may fail in achieving a goal mainly because itlacks the necessaryresources. It, therefore, delegates that goal to another agent. What is important to noteis that only the goal and not the associated intentions is delegated. The latter agentrebuilds the entire intentions for that goal and these intentions may differ from thosebuilt by the delegating agent due to different beliefs, commitments and skills of thetwo agents.


Although it is tacitly admitted that cooperation could be very beneficial to distri-buted network management architectures, it is not yet well exploited in current agent-based approaches. The non-communicative approach presented in [APP 94] has provi-ded robustness and made it easy to reach the graceful degradation of the whole MAS.Negotiation is a very interesting-to-use technique for theprovision and managementof telecommunication services. A market place model is suitable to implement agentsthat work to provide high-level services with the best quality and the lowest price. Forthis, an agent has to negotiate with others in order to maximise its owner’s profit.

The underlying communication mechanism is not equally addressed in each paper.While in [FRÖ 97] the communication protocol is not emphasised (simply becausethere is no need for more than simple message exchange), [RIZ95] has already distin-guished between message syntax and semantics. In [WEI 92], two types of messages,namely INFORM and REQUEST messages were sufficient, whereasin [SKA 96],messages are identified using message patterns. Most of the other related works haveopted for KQML such as in [INT 97] and [STE 96]. In the MANIA project [OLI 95],agents use KQML for the message syntax and a KIF-like Goal Definition Languageto exchange management goals. Finally, in [SOM 96], an extended version of KQMLis used. The added performatives are PROPOSE, COUNTER-PROPOSE,ACCEPT,REJECT, PROBLEM and REJECT. They are introduced to support the negotiationand the tendering processes.

From the Distributed Problem Solving point of view, none of the presented coope-rative agents has reached the degree of elaborating global planning and collaboratingto achieve shared goals. Also, according to [DOR 97], a cooperative MAS may eitheremerge from agents which give priority to their own internalgoals or from agents thattry first to better achieve the global goals of the system. Most of the cooperative agentsdescribed above fall into the first category.

8. An Interface Agent

An application of an interface agent for network supervision is presented in [ESF 96].The agent has to process a large amount of alarms and events byfiltering them andrelating each notification to its context.


An on-line knowledge acquisition is adopted. Using the chronicle model (see[GHA 94]), the agent is able to perform temporal reasoning onthe notified eventsand to automate management tasks. A chronicle is a set of (correlated) events termi-nated with an action that must be taken when the chronicle is matched. A chronicleis expressed using two formula types. Thehold formula expresses that an attributeholds its value over some interval. Theevent formula expresses a discrete change ofan attribute’s value.

While “looking over the shoulder” of the human administrator, aChronicle Re-cognition Systemidentifies chronicles and associates each one to the taken action.A Learning Systemevaluates the identified chronicle by comparing it to the knownones. A first step would be to store the chronicle in an unconfirmed chronicle base.After reaching a certain threshold (or being confirmed by thehuman), the chronicle istransferred to a confirmed chronicle base.

In this application, the agent learns from the user either byobserving him or byreceiving explicit instructions. Cooperating with other interface agents is not yet sup-ported but it is planned for future study (see the conclusionof [ESF 96]).

9. A General Review

This section discusses the achievements presented above from an agent-orientedview. It states the agent properties that were not investigated, and that consequentlyrequire more study to be carried out in the network management field. Afterwards, thetrends in the introduction of agent technology by network management communityare presented.

9.1. An Agent-oriented view

9.1.1. Pro-activeness

As stated in Section 3.2, pro-activeness is the ability for an agent to anticipatechanges in the environment and act in a way to prevent potential problems from oc-curring. In the NM context, pro-activeness makes an agent able to foresee faults orQoS performance degradation and to avoid such situations before they even occur.

Pro-activeness is the most ignored property in existent agent-based approaches toNM. Only two approaches have referenced it.

The first is [STE 96] which states when talking about pro-activeness: “Policiescould be used to allow the IA to divide a problem up into sub-problems, and then tosolve each one of the sub-problems. . .”. Clearly, this statement has nothing to do withpro-activeness.

The second approach is that used in the MANIA project [OLI 96]. The pro-activebehaviour is achieved in two possible ways. In the first way, the agent observes the ap-plication contexts sent by the users to declare their QoS requirements. It then deduceswhat network resources would soon be highly solicited. The second way consists


in learning the evolution of the network’s behaviour over time. The agent could forexample deduce that the NFS server is overloaded each day from 9 to 10 a.m. Theagent may decide to launch another server just before 9 a.m. (i.e., before even thecongestion occurs) or to forbid any logging-in on the server’s machine during the rushperiod.

9.1.2. Learning

Learning Network Management knowledge or expertise shouldbe distinguishedfrom learning the network’s behaviour. The first kind of learning enables the agentto acquire new capabilities in management, mainly to handlenew types of networkproblems. The unique example of this is the interface agent presented in [ESF 96] andreported in Section 8.

The second kind of learning enables the agent to acquire knowledge on the net-work it manages in order to apply the very same expertise but more efficiently. Inthe previous section, we saw how this kind of learning was applied to provide pro-activeness. In [APP 94], one may consider that the agents keeplearning the system’sstate in order to deduce a sufficiently accurate node utilisation average.

9.1.3. Robustness and Graceful Degradation

These two properties concern the whole multi-agent system.The graceful degra-dation means that the MAS does not fail drastically at the boundaries [NWA 96]. Ro-bustness is the ability to recover from faults that may occurdue to the underlyingenvironment (network failure or system halt).

The unique application where both these properties were studied is that in [APP 94].The mechanism that detects crashed agents and replaces themhas already been pre-sented in section 7.1. Let us recall that if a parent agent crashes, it will be detected andreplaced by static processes whose role is to maintain a minimum number of agents.Meanwhile, the other agents continue their tasks without stopping the system (grace-ful degradation). If a load agent crashes, it will be detected and replaced by the parentagents. Therefore, the MAS keeps on performing its role and is able to recover frompossible crashes (robustness).

9.2. Network Management Trends

The number of network management firms and network-device manufacturers whoare becoming interested in Intelligent Agent technology israpidly increasing. Manymanufacturers are trying to incorporate IAs within networkdevices such as routers andATM switches. The idea behind this is to endow the network devices with a level ofintelligence that allows them to acquire the properties of autonomy and self-healing.Besides, this intelligence at the lowest level of the network will have benefits on ma-nagement traffic reduction and network management automation.

Other trends are focusing on the service level. These trendsare mostly led by


telecommunications operators who are aiming for a high degree of flexibility in esta-blishing new services and for means to control and satisfy the quality of service thatthe end user requires.

Network management platforms are also affected by IAs. Manymanufacturers areannouncing products supporting IAs, although it is not yet specified how IAs will beintegrated into their platforms. For example, the whole NM platform can be basedon IAs, or it can include a framework to allow network administrators to developtheir own NM intelligent agents, or it can simply use static preprogrammed intelligentagents. Their hope is for the NMS to be much more reactive (or even pro-active),flexible and scaleable.

In fact, a lot of research is still needed to reach an acceptable usage of IA tech-nology in NM platforms. Indeed, a vast number of languages, environments and tech-niques are available for designing agents to define their behavior and to work out howthey can cooperate. Besides, there is still a great desire inthe NM field to find solutionsto integrate heterogeneous and vendor-specific equipment.

Regardless of these issues, efforts are being directed towards Java-based solutions.People are relying on the promise of a portable code due to the Java virtual machineand to Java-enabled chips which have been announced and evenrecently commercia-lised.

10. Conclusion

This paper has presented a state-of-the-art on Network Management approachesbased on Intelligent Agents. The reader should be aware thatmuch research remainsunpublished due to (industrial) confidentiality constraints. We are informed of suchprojects only via personal contacts, projects’ home pages and from the individual par-ticipation of members of such projects on mailing lists.

The papers that we were able to collect fall into three classes. The first class ofpapers only describe scenarios where some NM activities could be achieved by usingagents. These papers do not offer any architectural detailswhich makes them difficultto classify. The second class of papers present an agent-based solution to a specificNM issue. These solutions are well detailed but could not be used for an integrated ap-proach to NM. The last class of papers describe general architectures based on agents,each with an example of an application to a specific management function. In fact,the most researched functional areas are fault, configuration and performance mana-gement. Accounting and security management are much less researched - indeed, thelatter having never been studied.

The use of IAs has provided the NM approches presented with a high degree ofautomation and flexibility. As a result, the human administrator has been moved outfrom low-level and routine tasks. He is thus able to interactwith the agents in a high-level manner using policies, motivations and goals. Another new interesting feature isthat agents are closely aware of the detailed state of the network elements and theirusage. Corrective operations can be appropriately taken ina timely fashion and pro-active actions much more easily planned.


For these reasons, many companies have already begun heavily investing in IAs,and some are even participating in standardisation effortsas a way of promoting theirown research in this domain.

Acknowledgments

This report is established within the framework of the project F&E 391 which is acooperation between SwissCom and the Eurecom Institute. Itis financially supportedby SwissCom, and any further information about the project may be addressed to ourcorrespondant M. Hans Peter Gisiger (e-mail: [email protected]). Theauthors would like to thank Raul Oliveira for his precious collaboration.

References

[APP 94] APPLEBY S. and STEWARD S., Software Agents for Control. In COCHRANE P.and MEALTHEY P., Eds.,Modelling Future Telecommunication Systems. Chapman & Hall,1994.

[BAL 97] B ALDI M., GAI S. and PICOO G. P., “Exploiting Code Mobility in Decen-tralized and Flexible Network Management”. InProceedings of the First Inter-national Workshop on Mobile Agents, Berlin, Germany, April 1997. Available athttp://www.polito.it/˜picco/papers/ma97.ps.gz.

[BAR 96] BARBUCEANU M. and FOX M. S., The Design of a Coordination Language forMulti-Agent Systems. InIntelligent Agents III. Agent Theories, Architectures, and Lan-guages, p. 341–355. Springer, 1996.

[DOR 97] DORAN J. E., FRANKLIN S., JENNINGSN. R. and NORMAN T. J., “On Coopera-tion in Multi-Agent Systems”.The Knowledge Engineering Review, vol. 12, no 3, 1997.available at http://www.elec.qmw.ac.uk/dai/pubs/fomas.html.

[ESF 96] ESFANDIARI B., DEFLANDRE G. and QUINQUETON J., “An Interface Agent forNetwork Supervision”. In?, 1996.

[FRA 96] FRANKLIN S. and GRAESSERA., Is it an Agent, or Just a Program?: A Taxonomyfor Autonomous Agents. InIntelligent Agents III. Agent Theories, Architectures, andLanguages. Springer, 1996.

[FRÖ 97] FRÖHLICH P.,DE ALMEIDA MÓRA I., NEJDL W. and SCHROEDERM., “Diagnos-tic Agents for Distributed Systems”. InProceedings of ModelAge97, Sienna, Italy, Junary1997. Available at http://www.kbs.uni-hannover.de/paper/96/ma1.ps.

[GHA 94] GHALLAB M., “Past and future chronicles for supervision and planning”. InHATON J. P., Ed.,Proceedings of the 14th Int. Avignon Conference, p. 23–34, Paris, June1994. EC2 and AFIA.

[GHE 97] GHETIE I. G.,Networks and Systems Management. Platforms Analysis and Evalua-tion. Kluwer Academic Publishers, 1997.

[GIB 97] GIBNEY M. A. and JENNINGSN. R., “Market Based Multi-Agent Systems for ATMNetwork management”. InProc. 4th Communications Networks Symposium, Manchester,UK, 1997. Available from http://www.elec.qmw.ac.uk/dai/projects/agentCAC/.

[GOL 93] GOLDSZMIDT G., “Distributed System Management via Elastic Servers”. In IEEEFirst International Workshop on Systems Management, p. 31–35, Los Angeles, California,April 1993.


[GOL 95] GOLDSZMIDT G. and YEMINI Y., “Distributed Management by Delegation”. InThe 15th International Conference on Distributed Computing Systems. IEEE ComputerSociety, June 1995.

[GRE 96] GRESSLEYC.. “Network Management Resources - Reviews of Network Manage-ment Systems”, July 1996.Available at http://tampico.cso.uiuc.edu/˜gressley/netmgmt/reviews/.

[GRI 97] GRIMES G. and ALLEY B. P., “Intelligent Agents for Network Fault Diagnosis andTesting”. In Integrated Network Management V : Integrated Management ina virtualWorld, p. 232–244, San Diego, California, USA, May 1997. IFIP, Chapman & Hall.

[HJÁ 97] HJÁLMTÝSSON G. and JAIN A., “An Agent-based Approach to Service Manage-ment - Towards Service Independent Network Architecture”.In Integrated Network Ma-nagementV : integrated management in a virtual world, p. 715–729, San Diego, California,USA, May 1997. IFIP, Chapman & Hall.

[HUI 96] HUITEMA C.,Et Dieu créa l’Internet. Eyrolles, 1996.[IFI97] IFIP, Integrated Network Management V: integrated management ina virtual world,

San Diego, California, USA, May 1997. Chapman & Hall.[INT 97] FOR INTELLIGENT PHYSICAL AGENTSF.. “Application Design Test: Network Ma-

nagement and Provisioning”. http://drogo.cselt.it/fipa/spec, June 1997.[KAL 97] K ALYANASUNDARAM P., SETHI A., SHERWIN C. and ZHU D., “A spreadsheet-

based scripting environment for SNMP”. InFifth IFIP/IEEE International Symposium onIntegrated Network Management IM’97, vol. 5, p. 752–765, San-Diego, California, USA,May 1997. Chapman & Hall.

[KOO 95] KOOIJMAN R., “Divide and Conquer in Network Management Using Event-drivenNetwork Area Agents”. Technical Report, Technical Univerity of Delft„ The Netherlands,1995. available at http://netman.cit.buffalo.edu/Doc/Papers/koo9505.ps.

[LAB 93] L ABIDI S. and LEJOUADW., “De l’Intelligence Artificielle Distribuée aux systemesMulti-Agents”. Technical Report, INRIA Sophia Antipolis,INRIA Sophia Antipolis, 2004route des Lucioles, BP 93, 06902 Sophia-Antipolis Cedex,France, 1993.

[LUP 97] LUPU E. and SLOMAN M., “Towards A Role-Based Framework for Distributed Sys-tems Management”.Journal of Network and Systems Management, vol. 5, no 1, 1997.

[MAE 94] M AES P., “Agents that Reduce Work and Information Overload”.Communicationsof the ACM, vol. 37, no 7, p. 31–40, 1994.

[MAG 95] M AGEDANZ T., “On the Impacts of Intelligent Agents Concepts on FutureTe-lecommunication Environments”. InThird International Conference on Intelligence inBroadband Services and Networks, Crete, Greece, October 1995.

[MAG 96] M AGEDANZ T., ROTHERMEL K. and KRAUSE S., “Intelligent Agents: An Emer-ging Technology for Next Generation Telecommunications?”. In INFOCOM’ 96, p. 464–472, USA, March 24-28 1996. IEEE.

[MAR 96] M ARRIOT D. and SLOMAN M., “Implementation of a Management Agent for In-terpreting Obligation Policies”. InIEEE/IFIP 7th International Workshop on DistributedSystems and Operations Management - DSOM’96, L’Aquila, Italy, October 1996.

[MCC 94] MCCABE F. G. and CLARK K. L., “April: Agent Process Interaction Language”.Technical Report, Dept. of Computing, London, UK, November1994. available fromhttp://www-lp.doc.ic.ac.uk/˜klc/.

[MOU 96] MOUNTZIA M.-A.. “Intelligent Agents in Integrated Network and Systems Mana-gement”, 1996.

[MUL 96] M ULLER J. P.,The Design of Intelligent Agents - A Layered Approach. LNAIState-of-the-Art Survey. Springer, Berlin, Germany, 1996.

[NWA 96] NWANA H. S., “Software Agents: An Overview”. Knowledge Enginee-ring Review, vol. 11, no 3, p. 205–244, October/November 1996. Available athttp://www.cs.umbc.edu/agents/introduction/ao/.


[OLI 95] OLIVEIRA R. and LABETOULLE J., “Intelligent Agents : a way to reduce thegap between applications and networks”. In DECOTIGNIE J. D., Ed., Procee-dings of the First IEEE International Workshop on Factory Communications Sys-tems - WFCS’95, p. 81–90, Leysin, Switzerland, October 4-6 1995. Available athttp://www.eurecom.fr/˜oliveira/wfcs/wfcs.ps.gz.

[OLI 96] OLIVEIRA R., SIDOU D. and LABETOULLE J., “Customized network managementbased on applications requirements”. InProceedings of the First IEEE International Work-shop on Enterpr ise Networking - ENW ’96, Dallas, Texas, USA, June 27 1996.

[OMG 97] OMG, “Mobile Agent Facility Specification”. Technical Report, Crystaliz, Inc.,General Magic, Inc., GMD FOKUS, International Business Machine Corporation, June1997. OMG TC Document cf/xx-x-xx.

[PAR 95] PARKYN N.. “Architecture for Intelligent (Smart) Agents”, June 1995. available athttp://www.citr.uq.oz.au/.

[PRA 95] PRAS A., “Network Management Architecture”. PhD thesis, Universi-teit Twente, Department of Computer Science, February 1995. available fromhttp://wwwsnmp.cs.utwente.nl/˜pras/thesis.html.

[RIZ 95] RIZZO M. and UTTING I. A., “An Agent-based Model for the Provision of AdvancedTelecommunications Services”. InProceedings of TINA ’95, Melbourne, Australia, 1995.

[SAH 97] SAHAI A., BILLIART S. and MORIN C., “Astrolog: A Distributed and DynamicEnvironment for Network and System Management”. InProceedings of the 1st Euro-pean Information Infrastructure User Conference, Germany, February 1997. Available athttp://www.irisa.fr/solidor/doc/pub97.html.

[SKA 96] SKARMAEAS N. and CLARK K. L., “Process Oriented Programming for Agent-Based Network Management”. InECAI96 Workshop on Intelligent Agents for Telecom-munication Applications (IATA96), Budapest, Hungary, August 12 - 16 1996.

[SOF 97] SOFTWAREF., “FTP Software Agent Technology”. Technical Report, FTPSoftware,http://www.ftp.com/product/whitepapers/4agent.htm, 1997.

[SOM 96] SOMERS F., “HYBRID: Unifying Centralised and Distributed Management forLarge High-Speed Networks”. InNetworks Operation and Maintenance Symposium(NOMS96), Kyoto, 1996 1996. Available from http://www.broadcom.ie/˜fs.

[STA 96] STALLINGS W., SNMP, SNMPv2 and RMON, Practical Network Management.Addison-Wesley, USA, 1996.

[STE 95] STEVENSOND. W.. “Network Management: What it is, what it isn’t.”, 1995. avaia-lable at http://netman.cit.buffalo.edu/Doc/DStevenson.

[STE 96] STEENEKAMP P. and ROOS J., “Implementation of Distributed Systems Manage-ment Policies: A Framework for the Application of Intelligent Agent Technology”. In2ndInternational Workshop on Systems Management, Toronto, Ontario, Canada, June 1996.IEEE.

[SUZ 97] SUZUKI M., K IRIHA Y. and NAKAI S., “Delegation Agents: Design and Implemen-tation”. In Integrated Network Management V : integrated management ina virtual world,vol. 5, p. 742–751, San Diego, California, USA, May 1997. IFIP, Chapman & Hall.

[TER 92] TERPLAN K., Communication Networks Management. Prentice Hall, 1992.[TRO 97] TROMMER M. and KONOPKA R., “Distributed Network Management with Dyna-

mic Rule-Based Managing Agents”. InIntegrated Network Management V : integratedmanagement in a virtual world, p. 730–741, San Diego, California, USA, May 1997. IFIP,Chapman & Hall.

[WAG 96] WAGNER G., “Vivid Agents - How they Deliberate, How They React, How TheyAre Verified”. In DE VELDE W. V. and PERRAM J., Eds.,Agents Breaking Away, Proc. ofMAAMAW’96, 1996. Available from http://www.informatik.uni-leipzig.de/˜gwagner/.


[WEI 92] WEIHMAYER R. and TAN M.. “Modelling Cooperative Agents for Customer Net-work Control Using Planning and Agent-Oriented Programming”, 1992.

[WEL 93] WELLMAN M. P., “A Market Oriented Programming Environment and its Appli-cation to Distributed Multicommodity Flow Problems”.Journal of Artificial IntelligenceResearch, vol. 1, p. 1–23, 1993.

[WOO 95] WOOLDRIDGE M. and JENNINGS N. R., “Intelligent Agents: Theory and Practi-ce”. Knowledge Engineering Review, vol. 10, no 2, p. 115–152, 1995.

[YEM 91] Y EMINI Y., GOLDSZMIDT G. and YEMINI S., “Network Management by Dele-gation”. In The Second International Symposium on Integrated Network Management, p.95–107, Washington, DC, April 1991.

[ZNA 96] ZNATY S., LION M. and HUBAUX J.-P., “DEAL: A DElegated Agent Languagefor Developing Network Management Functions”. InFirst International Conference andExhibition on the Practical Application of Intelligent Agents and Multi-Agent Technology,1996.

intelligent agents in network management a state-of-the-art

Documents