11th ICCRTSCoalition Command and Control in the Networked Era

Agent Coordination Mechanisms for Multi-National Network Enabled Capabilities

(C2 Architecture, C2 Experimentation, C2 Modeling and Simulation)

P.P.A. Storms and T.J. Grant

P.P.A. StormsY’All B.V.

Grotestraat 1825141 HD WaalwijkThe Netherlands

phone: +31 (0)416340019email: patrick@yall.nl

T.J. Grant1

NLDAP.O. Box 900024800 PA Breda

The Netherlandsphone: +31 (0)765273261email: tj.grant@nlda.nl

Point of contact: T.J. Grant

1Also with Department of Computer Science, University of Pretoria, South Africa

Abstract

Modern advanced information technology enables military organisations to share information, suchthat decision making occurs at all levels within the chain of command. In a network centric approachto warfare, assets like sensors, shooters and C2 systems are interconnected in an infostructure, orinformation grid. By sharing information and combining capabilities, these assets work together toachieve enhanced capabilities. In the Netherlands, this is termed Network Enabled Capabilities (NEC).Under NEC, assets can dynamically form temporary “teams” to fulfill a specific task. To realize suchagile configurations of assets, the problem of managing tasks between assets has to be tackled.

In this work we argue that specialized software components, called intelligent agents, are suitablefor coordinating tasks between assets. NEC systems can be viewed as a type of multiagent systems(MAS) in which agents represent assets and connect them to the information grid. Now, coordinationmechanisms, as explored in the field of MAS design, are also applicable in a NEC setting.

The aim of the research reported in this paper is to identify which coordination strategies are suitedto NEC. We take the following approach. First, we give an overview and classification of agent basedcoordination mechanisms and their properties. We suggest a taxonomy of agent coordination strategies.Next, we identify the key requirements for coordination mechanisms in a NEC setting. Based onthese requirements we argue that explicit, centralized coordination strategies with low communicationoverhead in a cooperative environment are best suited as primary coordination strategy. Finally, wecompare our research with other initiatives, employing agent technology.

We recognize that there is no single best way to coordinate, and that for local clusters of agents othertypes of coordination strategies might be preferable. Therefore we suggest a hybrid NEC coordinationstrategy, composed of a global primary coordination strategy, with subordinate clusters of agents thatuse local coordination strategies. This can be a mechanism for handling multi-national coalitions. Inthe proposed architecture, agents are organized in a nested structure of clusters, or holons.

1

1 Introduction

1.1 Network Enabled Capabilities

Post-Cold War conflicts predominantly involve awide range of global and regional actors (majorpowers, international agencies, neighbouring states,terrorist groups and criminal networks) and mod-ern warfare has to cope with asymmetric threatsand networks of decentralised, loosely coordinated,fighting groups. Moreover missions are no longer asingle-nation effort, but are executed in Joint force,in conjunction with Allied and coalition partners.This imposes new requirements to militairy oper-ations such as increasing the tempo of operation,dealing with changing goals during mission execu-tion, being interoperable with own and coalitionforces, efficient deployment and coordination of as-sets, and sharing information with all elements toachieve a high level of shared situational awareness[1].

Modern rapidly advancing information technol-ogy enables military organisations to share infor-mation, such that decisionmaking occurs at all lev-els within the chain of command [2]. Technologicalchanges give organisations the opportunity to takefull advantage of all available information and tobring all available assets to bear in a rapid andflexible manner [3]. Thanks to the web structure,multiple redundant paths for information sharingare possible. This is exactly what drives the currenttransformation of hierarchical, platform centric or-ganisations to agile, network centric organisations.In the next subsection we will take a closer look atvisions on military networked organisations.

Network Centric Warfare (NCW) is a militaryconcept in which information superiority enablesoperations to generate increased combat power. Bynetworking sensors, decision makers and shooters,NCW aims at increasing shared awareness, speed ofcombat, tempo of operations, lethality, survivabil-ity and the level of self synchronisation. In essence,NCW translates information superiority into com-bat power by effectively linking knowledgeable en-tities in the battle space [4]. NCW is a formalUS networking concept and doctrine that seeks todevelop into a fully-fledged warfighting capability.The tenets of NCW are:

1. A robustly networked force improves informa-tion sharing.

2. Information sharing enhances the quality of in-formation and shared situational awareness.

3. Shared situational awareness [1] enables col-laboration and self-synchronization, and en-hances sustainability and speed of command.

4. These, in turn, dramatically increase missioneffectiveness.

By sharing information, NCW aims at achievinga heightened state of shared situational awarenessand knowledge among all elements. Consequentlyan increased effectiveness is reached with the samenumber of assets, enabling a webbed network ofmultiple military organisations to counter modernthreats.

The Dutch and UK Ministries of Defence use theterm Network Enabled Capability (NEC) to denoteevolving capability by bringing together sensors togather information, a command and control (C2)network to fuse, communicate and exploit the in-formation, and strike assets to act rapidly to de-liver the required effect. All available assets pooltheir information by networking, in order to achieveenhanced capabilities [5]. NEC shares the tenetsof NCW, but is more limited in scope. It is nota doctrine, but rather a conceptual and technicalframework for gradually implementing NCW the-ory to actual enhanced capabilities. Realization ofNEC is a process of change and evolution, startingwith automating and digitising current operational(decision) processes and integrating previously dis-persed systems. Throughout this paper we will usethe term NEC.

1.2 Assets connected in information

grids

NEC theory states that available assets can onlybe fully exploited if there is a high degree of de-composition and logical decoupling. For instance,if a sensor (e.g. tracking radar) is logically decou-pled from a weapon system (e.g. close-in weaponsystem) both the weapon and the sensor can becontrolled independently and can be shared withother assets and deployed more efficiently. Yet thesensor remains physically mounted on top of theweapon system. Now, one can think of a trackingradar supplying range information to complementangle-only tracks from an infra-red sensor. Also the

2

weapon system might acquire a target using datafrom other sensors than its own, e.g. in case ofa faulty sensor, to surprise the enemy, or if moreaccurate target data is already available.

Figure 1: Cooperation at platform level

Figure 2: Sharing information and control of re-sources in a NEC grid

Figures 1 and 2 illustrate the difference betweena platform centric and a network centric approach,respectively. In a platform centric approach eachplatform (in this case an aircraft carrier, a frigateand a submarine) is an indivisible unit in a mis-sion. Although the platforms can exchange tacti-cal information using datalinks, their capabilitiesare limited to the assets that are collocated on theplatform itself. The assets are typically sensors, C2nodes (such as the Command Information Center)and weapons. Each platform has its own decision

cycle (e.g. surveillance, target detection, threat as-sesment, sensor-weapon assignment, target aquisi-tion and interception) and mainly has to use itsown resources for execution.

In a NEC setting, the available assets are logi-cally decoupled from the platforms and organisedas nodes in an information grid. The grid enablesinformation sharing and control of assets betweenall nodes. Consequently assets can be used inde-pendently of the platform, yielding new capabilitiessuch as integrated fire control (IFC) [6]. IFC com-bines sensors, C2 nodes and weapons of differentplatforms to enable collaborative engagements. Anexample of an IFC capability is Engage On Remote,as illustrated in Figure 3: one platform tracks athreat and uploads fire control data to a firing unit,which is responsible for interceptor guidance. Us-ing the same assets, the effective range of weaponsis extended. Currently, US Navy developed IFCcapabilities for anti air warfare within the Cooper-ative Engagement Capability (CEC) programme.

THREATINTERCEPTOR

FIRING UNIT REMOTE UNIT

INTERCEPTORGUIDANCE

TRACKING

UPLOAD FIRE CONTROL DATATO FIRING UNIT

Figure 3: Integrated Fire Control Engage On Re-mote Capability

Another example of decoupling is the UnmannedAirborne Vehicle (UAV), a remotely controlled re-connaissance and combat airplane. Here the pilotis physically decoupled from the aircraft. Since thepilot is no longer required to be actually present atthe battlefield and the UAV is semi-autonomous,the pilot becomes an asset that can be shared.From a remote control center he is able to oper-ate multiple UAVs simultaneously. The UAV pilotcan be assigned to multiple missions, in differentorganisational structures.

3

1.3 Agents, Multi Agent Systems

and Agent Coordination

To realise NEC, existing assets need to be con-nected to the infostructure, and their information,services or capabilities need to be accessible forother assets. To achieve this a number of prob-lems have to be tackled. Amongst others, theseproblems concern interoperability, security and co-ordination of tasks between assets. In this workwe envision that specialized software components,called intelligent agents, are suitable to fulfill thistask.

Several different perspectives, property sets andclassifications for intelligent agents are describedin the literature [7, 8, 9, 10]. There is a consen-sus that agents are automated entities (machines,software processes) that have some degree of in-telligence and autonomy [11] in pursuing a set ofgoals. Agents are able to perceive their environ-ment and respond to changes in a timely fashionby acting upon their environment. Action can beboth reactive and proactive. Finally, agents shouldbe able to interact with other agents, humans andnon-agent systems to offer their services, take ac-tion on behalf of them (agency), and cooperate toachieve a set of goals. These goals can be commonor conflicting.

The term multiagent system (MAS) denotes anetwork of agents. Multiagent systems are typi-cally (large scale) distributed systems, comprisedof multiple individuals or services and engaged inmore than one task. Furthermore multiagent sys-tems have: goal-directed behaviour (where goalscan change), the ability to affect and be affectedby the environment, legal standing and the pres-ence of knowledge, culture, memories, history andcapabilities distinct from any single agent [12].

Figure 4: Agents connecting assets to the grid

Aspect NEC MAS

assets sensors, agents,shooters, humans,control systems, non-agent systemsmilitary units

network information grid agent platformsensor gridengagement grid

services information sharing, problem solvingsituation awareness,cooperative engagement

goal execute mission, perform tasks,achieve effect maximise utility

Table 1: Relation between NEC and MAS

NEC systems can be viewed as multiagent sys-tems in which agents represent assets and connectthem to the information grid. Figure 4 shows, ina simplified example, how naval assets are repre-sented by agents. The agents connect the assets tothe grid, and assets can only be accessed throughtheir corresponding agent. Thereby, all assets areshielded from each other’s specific technical and be-havioural characterists. Furthermore, the agentscan guard availability of the assets and access poli-cies. Finally, the agent can take care of dynami-cally forming temporary “teams” of assets to fulfilla specific task.

Table 1.3 further illustrates the relation betweenNEC and MAS. The primary assets in a MAS areagents which can represent other assets such as hu-mans or non-agent systems. An agent platformtakes care of transporting messages between agents,and provides additional networking services such asagent lifecycle control, addressing and lookup fa-cilities. In a NEC context the assets are sensors,shooters, control systems and military units, con-nected in a grid. A MAS provides problem solvingservices, in order to perform a set of tasks or to op-timise some utility function. The services and goalsstrongly depend on the application domain. In theNEC application domain, the services include in-formation sharing, situation awareness and cooper-ative engagements. The goals of the network are:executing a mission, or achieving a desired effect.

The increased gain in agility offered by NECcomes at a price. Dynamically organising assetsyields substantial coordination efforts, which arenot required in case of static, pre-defined organisa-

4

tional structures. To realise NEC systems we needto introduce coordination mechanisms that respectthe specific requirements of the military domain ingeneral, and of network centric operations in par-ticular. In the field of MAS design, coordinationhas been studied extensively [13, 14, 15, 16, 17].Theory and concepts stemming from this researchcan be applied to tackle NEC related coordinationproblems.

1.4 Problem statement and ap-

proach

As we pointed out in section 1, the gain in agilitycaused by a network centric approach to warfarecomes at the cost of increased effort for establish-ing coordination. We also argued that coordinationmechanisms as explored in the field of MAS designseem promising for a NEC setting, because of thesimilar characteristics of NEC and multiagent sys-tems.

The aim of this work is to identify coordinationstrategies suited to NEC. We take the following ap-proach. First, we will give an overview and clas-sification of agent based coordination mechanismsand their properties in section 2. Next, in section 3we zoom in on specific coordination issues in NEC,and identify the chief requirements to coordinationmechanisms in a NEC environment. We evaluatewhich types of agent based coordination methods,as identified in section 2, meet the NEC require-ments. Based on this result we propose a NECcoordination architecture. Section 4 discusses theresults, related work and future work.

2 Coordination in Multiagent

Systems

2.1 What is coordination?

Agent coordination is concerned with the controlof the activities and the regulation of object flowsfrom an operational perspective within a MAS.Here, object denotes an artifact that can be pro-duced, consumed or tranformed by agents. For in-stance, an object can be data, information, knowl-edge or a physical entity. Objects can have variousproperties, i.e. they can be discrete (e.g. messages)

or continuous (e.g. energy), divisable or indivis-able, sharable (e.g. can be accessed by multipleagents simultaneously) or exclusive, and static ordynamic (i.e. the properties of an object change intime). With activity we mean the act of consuming,transforming and distributing objects. Each suchsub-activity is called a job, i.e. consumer job, trans-formation job and distribution job respectively. Ac-tivities are part of primitive tasks, which are at thelowest level of a task decomposition hierarchy.

A common sense definition for coordination is:Coordination is the act of working together harmo-niously [13]. If we apply this definition to MAS,the act of “working together” implies that agentsperform activities and that interdependencies ex-ist between these activities, such as consumer /producer relations. The predicate “harmoniously”implies that either no conflicts exist, or conflictsare resolved. For establishing a conflict-free sys-tem some kind of management process is required[13, 17].

As pointed out by Corkill and Lander [15] coor-dination is an essential activity in multiagent sys-tems, since it permits agents to perform complexcomposite tasks and achieve (common) goals bymeans of interaction. Note that, considering thedefinitions of object, activity, job and task, the actof task decomposition is outside the scope of coordi-nation. We assume that in a MAS tasks are alreadydecomposed to a level of granularity that fits thecapabilities of individual agents, and the interde-pendencies between the sub-tasks (e.g. the orderof execution) are explicitly known. So, if agentscommunicate on coordination this communicationwill only involve objects, activities and jobs.

Agents can be organised in an authority struc-ture. For instance, in a hierarchical organisa-tion agents with managerial responsibilities dele-gate tasks to subordinate agents. Within an or-ganisation different control regimes (coordination)can exist. In general, an organisation sets generalnorms to agent behaviour, coordination patternsand authority. These norms are respected by themembers of the organisation for a potentially longterm. Coordination takes place within the “rules ofthe game” of an organisation, and is concerned withthe relatively short term management of specific ac-tivities. We must therefore always consider coordi-nation strategies in the context of an organisation[18]. For an overview of organisational paradigms

5

in multiagent systems, see Horling et al. [19].

2.2 Communication on coordination

In a MAS, part of the interaction between agentswill involve communication about coordination.For instance, agent A delegates an activity to agentB and requires a notification when B has finished.If A does not receive a notification within x sec-onds, A will assume that B has failed to completethe activity. Apart from a description of the activ-ity itself, A has to inform B about the procedure.This implies that agents need to be interoperableat the level of coordination.

Figure 5: Levels of interoperability

In a model for agent interoperability [17], that issimilar to the ISO-OSI Network Model [20], coordi-nation is positioned at the top layer. See Figure 5.The lower-level layers are concerned with seman-tic (i.e. the meaning of messages), syntactical (i.e.the structure of messages) and technical (i.e. thephysical distribution of messages) interoperability.

Communication on coordination can concern dif-ferent dependency relations between activities. Os-sowski [21] identified the following dependency re-lations between activities:

1. Producer-consumer dependencies. An agentmay produce one or more objects that are to beused by other agents. Sometimes transporta-tions (i.e. transactions within the flow of ob-jects) have to be performed to move objectsbetween agents. There are different ways tomanage this dependency. A producing agentmay either be reactive (i.e. wait for consumer

demand) or pro-active (i.e. produce objects inadvance). In both cases the producing agentcan either inform consumer agents about thepresence of an object (event notification), ordirectly distribute the objects.

2. Shared resource dependencies. Multiple agentsmay need mutually exclusive access to a singleobject. Resource allocation actions can man-age this dependency.

3. Simultaneity dependencies. Multiple agentsmay need to perform activities that cannot beperformed at the same time. Synchronisationactions restrict the periods in which activitiescan occur.

4. Task dependencies. A group of activities mayaccomplish tasks that jointly attain come over-all goal. Selection, delegation, interventionand aggregation actions must respectively en-sure that activities are selected in the right or-der, responsibility for performing activities isdelegated to available and competent agents,responsibility is taken back in case of failure,and results are aggregated to complete thetask.

These dependency relations can be seen as coor-dination constructs of a MAS. Each relation defineswhat agents talk about when communicating on co-ordination. How coordination is realized, is definedby a coordination strategy. These are discussed inthe following section.

2.3 Taxonomy of agent based coor-

dination strategies

Agent coordination strategies are designed to en-able a group of agents to work together in someway to complete a single task or well defined setof tasks (also called a problem-solving episode). Inthis section we give an overview of various typesof coordination strategies. Based on the classifi-cation criteria described in literature [14, 21, 18],we distinguish the following dimensions to classifycoordination strategies:

1. Implicit versus explicit coordination. See2.3.1.

2. Dynamic versus static coordination. See 2.3.2.

6

3. Coordination strategies for cooperative versuscompetitive agent systems. See 2.3.3.

4. Centralised versus decentralised coordination.See 2.3.4.

In addition to these dimensions, we will considerthe coordination strategy’s design metaphor as adiscriminating property. This is discussed in sec-tion 2.4.

2.3.1 Implicit and explicit coordination

For implicit (communication-less) coordinationstrategies there is no explicit inter-agent communi-cation related to coordination. Either agreementson coordination are shared by all agents in a MAS,or agents operate under local sensing and control.In the latter case system-level behaviour arises frominteraction between individual agents through theirenvironment. An example of such a mechanism isstigmergy [22, 23, 24, 25].

With explicit (communication-based) coordina-tion strategies, agents explicitly communicate in-formation related to coordination. Agents makeuse of a coordination strategy, which can be seenas a decision-making and communication patternamong a set of agents that perform activities to co-ordinate task execution [26]. Approaches to explicitcoordination are many and various. Among oth-ers we distinguish market-based, negotiation-basedand organisation-based approaches [27, 28, 29].

As observed by Jones et al. [30] the distinc-tion between implicit and explicit coordination isnot crisp. It is rather a continuous spectrum since(1) there will always be some coordination-relatedcommunication and (2) there will always be somekind of shared model on coordination.

2.3.2 Dynamic versus static coordination

Dynamic coordination strategies allow agents to al-ter their coordination strategy at runtime. Thiscan be done by either fine tuning the configurationof a specific coordination strategy, or replacing thecurrent coordination strategy by another.

With static coordination the coordination strat-egy for a MAS is determined and configured a pri-ori, e.g. at design time.

2.3.3 Coordination in cooperative and

competitive agent systems

A cooperative MAS is an association of agents thatjoin together to carry out an activity of mutual ben-efit. This sets specific requirements on the coordi-nation strategy, since agents need to share their re-sources and to orchestrate their activities such thatthe group can perform certain tasks better than asingle individual agent. The coordination strategyhas to support the agents to maximise some globalutility. Individual preferences or goals are of sec-ondary priority.

In a competitive MAS agents are self-interestedand will primarily pursue their individual goals.The overall performance of the group is of sec-ondary interest. The goal of a coordination strat-egy in a competitive environment is to “persuade”agents to cooperate in performing a task, by satis-fying individual goals or preferences. In a compet-itive MAS agents may have conflicting goals andmight not be willing to share all information.

Typical techniques used in cooperative agentsystems are blackboards, Contract Net (CNET)and distributed constraint satisfaction [31, 32, 33].Blackboard systems allow agents to contribute to acommon goal by sharing information and partial so-lutions with other agents. CNET is a well-knownmechanism for task allocation, based on biddingand contracting [34]. Techniques that stem fromthe field of distributed constraint satisfaction aresuitable for determining a task allocation, withoutrelying on a central contractor (as is the case withCNET).

Note that these techniques are only suitable ina cooperative environment. The use of a black-board architectures implies that agents are will-ing to share information. Using CNET and dis-tributed constraint satisfaction for task allocationrequires agents to be honest (i.e. they do not pro-vide false information on availability and suitabil-ity) and benevolent (i.e. an agent does not benefitfrom accepting a task).

In a competitive MAS agents will typically re-quire some kind of “reward” to perform a certaintask, will value the resources assigned to them, orare willing to “pay” other agents for providing aservice. Typical coordination strategies for com-petitive agent systems are:

• Market mechanisms. If in a MAS the main

7

relation to be managed between agents isconsumer-producer or producer-consumer (see2.2) then coordination can modeled as an ar-tificial economy in which agents trade objectsin exchange for some utility. The value of ob-jects is determined by economic rules of sup-ply and demand. An example of a well-knownmarket mechanism in agent systems is the auc-tion, such as the Dutch, English, first-price orVickrey auction [27, 35].

• Negotiation mechanisms. If a MAS mainly hasto deal with resolving conflicting goals thennegotiation mechanisms are suitable coordina-tion strategies. Whereas CNET is applicablein a cooperative setting, negotiation mecha-nisms aim at maximising an egalitarian socialwelfare function or Nash product seem moreapplicable in a competitive setting. The gen-eral idea is that agents negotiate on objects(tasks, resources) such that the utility of theleast-satisfied agent is maximised. For back-ground reading see [28].

The differences between these two strategies aresubtle. Typically, market mechanisms are mod-elled as auctions where third-party agents (auction-eers) mediate between producers and consumers.In case of negotiation, agents make (a sequence of)bi-lateral trades, without intervention of a third-party agent. Since both strategies involve the non-benevolent exchange of objects to establish an allo-cation of objects to agents, negotation can be seenas a market mechanism. Vice versa, an auction canbe seen as a form of mediated negotiation. For thesake of clarity, we will use the term market-basedcoordination to denote both strategies.

Note that for both cooperative and competitiveagent systems, overall goals are eventually reached.Even though in a competitive MAS agents mayhave conflicting goals, conflict can be resolved bynegotiation and trade. In this work we do not con-sider malicious agents, i.e. agents aimed at deliber-ately corrupting the functionality of a MAS. In thecontext of NEC, we assume that all agents (both co-operative and competitive) in a military coalition’sgrid are somehow willing to contribute to successfulexecution of a mission. Still, the grid itself shouldbe protected against possible intruding maliciousagents of the opposing force. Examples of maliciousagents are viruses, spyware and computer worms.

2.3.4 Centralised and de-centralised coor-

dination

In a MAS where coordination is centralised, a sep-arate set of (computational) assets can be distin-guished that is solely occupied with handling coor-dination. All the other agents in the MAS have nocapabilities to coordinate, other than informing thecentral coordination mechanism about their stateand obeying its instructions. Note that centralisedcoordination does not necessarily mean that coor-dination is performed by a single central system. Itcan also be a distributed system that is functionallyseparated from the MAS that is under control.

In a MAS with decentralised coordination eachagent has the capability to coordinate, as well astheir functional (problem solving) capabilities.

The distinction between centralised and decen-tralised coordination is not crisp. For instance, ifan agent in a MAS with decentralised coordinationdelegates a set of activities to a number of subordi-nate agents, then this agent will temporarily act asa centralised coordinator for a subset of the MAS.

2.4 Agent coordination design

metaphor

A metaphor is a comparison which imaginativelyidentifies one concept with another dissimilar con-cept, and transfers or ascribes to the first someof the qualities of the second. Metaphors are es-pecially powerful when used to help understand aconcept that is unfamiliar or unapproachable. Insoftware engineering, metaphors are an instrumentto conceptualise the structures and designs of infor-mation systems, and to communicate their mean-ing. Applied to this work, we can distinguish agentcoordination by their underlying design metaphor.The number of design metaphors that can be ap-plied to agent coordination is virtually unlimited.In this section we highlight a few commonly usedmetaphors.

2.4.1 Organisational metaphor

A commonly applied metaphor for realizing coor-dination in a MAS is the organisational metaphor,where multiagent systems are modelled as humanorganisations. An organisational structure consistsof stakeholders (i.e. individuals, groups, physical or

8

social systems) that coordinate and interact witheach other to achieve common goals. Hence organ-isations can be seen as a type of coordination [14].

Organization theorists have proposed patternssuch as the structure-in-5 [36], the matrix, the chainof value and the like to define organizational struc-tures and behaviours. Within a human organisa-tion individuals or units (groups of individuals) actaccording to a set of (social) rules and hold a posi-tion or role, that is associated with certain respon-sibilities and skills. For instance, in an organisa-tion one may distinguish units like planning, sales,marketing, procurement, product design, produc-tion, customer support, etc. Within these unitsone may find (operational) roles that are specific forthat unit (like planners, designers, assemblers) ormore generic roles that can also be found in otherunits (like management, administration). Agree-ments are made on how these units and positionsoperate, how they are controlled and to whom theyreport.

Organisational concepts can be mapped ontothe design of multiagent systems [37, 38], whereagents are arranged in organisational structuresaccording to function, skill, location, time, clientor location. Popular patterns used in the organ-isational metaphor for agent coordination are de-rived from Mintzberg’s [36] organisational struc-tures [17, 29, 39, 40]. Malone et al. [26] point outthat the goal of coordination is to manage inter-dependencies between positions and activities per-formed to achieve goals. A classification of de-pendencies is given by [21] (see also section 2.2).Mintzberg has mechanisms that can be applied tocoordinate these dependencies. These mechanismsare:

Direct Supervision This coordination mecha-nism achieves coordination by having one in-dividual take responsibility, which is taking alldecisions for the work of others, issuing in-structions to them and monitoring their ac-tions. This mechanism can be seen as a pat-tern for one central reasoning service (i.e. theManager) with several information providingprocesses (i.e. Operators). This form of coor-dination is suited when there is a clear distinc-tion between decision-making and operation.

Standardization of Work This mechanismachieves coordination by specifying the con-

tent of the technical activity. The content ofthe activity is specified in every step, fromgetting the input objects (consume job), whatto do with it (transform job) and to whom todistribute it (distribute job). In MAS designthis means a (hard coded) procedural programdictates the behavior of an agent, withoutany room for negotiation with other agents.Conflicts will be reported to the supervisor(e.g. Manager).

Standardization of Output Objects Thismechanism achieves coordination by onlyspecifying the result (i.e. the objects pro-duced) of the activities. In MAS design thismeans standardizing the agent’s interfaces(e.g. specifiying output objects) and exchangemechanisms, like languages and ontologies.

Standardization of Skills Here, coordination isachieved by only specifying what competencesare needed for the activity. In MAS designthe knowledge required for specific activitieshas to be specified and agents need to beequipped with knowledge about the compe-tences of other agents. By means of proto-cols, they can exchange knowledge and dis-cover competences of others.

Mutual Adjustment Coordination is achievedby a process of informal communication be-tween positions. This means that positionsare capable of solving coordination issues bythemselves. In MAS design this means thatagents have social abilities in the sense thatthey are capable of interacting and reasoningabout each other’s interfaces, knowledge andcompetences and activities to achieve, with lit-tle standardization or protocols.

Figure 6 gives a schematic representation ofthe coordination mechanisms (borrowed from [17]).The circles represent roles of agents. M standsfor Manager, O for Operator and R for request-ing agent. Authority structures are represented bystraight lines that connect positions. The curvedlines represent message flows. The figure represent-ing Standardisation is valid for Standardisation ofWork, Output of Objects and Skills. In the case ofMutual Adjustment there are no Manager or Op-erator roles. Here A denotes an agent capable ofsolving coordination issues.

9

M

O O

Manager

Operator Operator

t a s ka c t i v i t y o b j e c t a c t i v i t y o b j e c to u t p u to b j e c tR i n p u to b j e c t(a) Direct Supervision

M

O O

p r o c e d u r ep r o c e d u r ei n p u to b j e c t o b j e c to u t p u to b j e c tR

t a s k(b) Standardisation

R

A A

o b j e c to b j e c t o b j e c to b j e c to b j e c to b j e c t

(c) Mutual Adjustment

Figure 6: Three coordination mechanisms.

At the level of coordination the organisationalmetaphor allows engineers to design multiagent sys-tems as if they were human organisations. Thereby,engineers can make design decisions based on ideasand concepts from the field of organisation the-ory. Note that the metaphor allows for numer-ous types of coordination strategies. Based on theMintzberg’s structures, coordination by organisa-tion typically yields explicit coordination strategiessince agents directly communicate on coordination.

2.4.2 Market metaphor

Multiagent systems comprised of competitiveagents have to deal with non-benevolent, selfishagents that aim to satisfy private goals. The coordi-nation strategy exploits the competitive behaviourto optimise some global utility function, e.g. bytreating the MAS as a virtual market place of ne-gotiating agents. In this market buying agents mayrequest or place bids for a common set of objectssuch as services, information or access to resources.Seller agents, or third party auctioneer agents, areresponsible for processing bids and determining thewinner [19]. So, allocation of objects to agents iseither facilitated by a central auctioneer agent or a

by means of a sequence of (bilateral) local negotia-tions between buyers and sellers.

The arrangement of agents in buyers and sellers,competing for objects, creates a system that ex-cels in coordinating producer-consumer type rela-tionships. It closely mimics real-world market eco-nomics. Market-based coordination mechanismsborrow concepts from economics and trade, suchas the notion of auctions. Various auction-basedmechanisms exist, and can vary along a number ofdimensions:

• In private or sealed-bid auctions participantsdo not see competing bids, as opposed to publicauctions.

• An auction can either be single shot, or iter-ate over several rounds either as an ascending(English) or a descending (Dutch) auction.

• In a reverse auction, sellers bid rather thanbuyers.

• If either the buyer or the seller maintains afixed price, the auction is one-sided. If bothparties compromise, the auction is two-sided.

• In a combinatorial auction participants bid oncollections of objects, rather than single ob-jects. Here, the value of a combination of ob-jects is not necessarily the sum of individualobject values.

• Typically, the highest bidder wins the auction.In a Vickrey auction [35] the highest bidderwins but pays the second highest bid price.This promotes truthful bidding.

An alternative approach for establishing an allo-cation of objects in a competitive environment isby means of local negotiations, without the inter-vention of an auctioneer. Typically, this involvesa series of bi-lateral exchanges of objects. Eachparticipating agent assigns a value (utility) to thecurrent set of objects it owns, and interacts with apeer to see if an exchange (deal) can be made suchthat both parties benefit. An agent might offerside-payments to its peer, to compensate for possi-ble loss of utility. Ultimately, the goal is to reach anallocation of objects such that some global utilityfunction is optimised.

10

Current research focuses on establishing conver-gence and complexity properties for different kindsof negotiation, for different types of deals, utilityfunctions and data representations [28]. One im-portant result is that any sequence of mutually ben-eficial deals without side payments will convergeto a Pareto optimal2 allocation [41]. Although noguarantee can be given on the number of dealsneeded to establish this optimal solution, the over-all utility will increase after every deal. In time-critical applications, as found in the military do-main, a system can start with an initial sub-optimalallocation (that might be computed by a set of fastheuristics) and improve this result in the remainingcomputing time.

2.4.3 Biological metaphor

In nature, communities of living entities are foundthat coordinate their activities in a robust andadaptive way. For instance, many social insects de-posit and sense chemicals (pheromones) in a sharedphysical environment, that participates actively inthe system’s dynamics. The presence and char-acteristics of pheromone influences the behaviourof individual insects. Although the individual be-haviour follows simple rules, the community as awhole will display some kind of complex, intelli-gent behaviour. This phenomenon is known asswarm intelligence. The act of communicating viapheromone markers in the environment is calledstigmergy.

Computer scientists and software engineers areinspired to model coordination strategies for mul-tiagent systems using swarm mechanisms found innature. For computer systems Beni et al. [42] de-fine swarm intelligence as a property of systems ofnon-intelligent robots exhibiting collectively intel-ligent behaviour. Alternatively the term emergentintelligence or emergent behaviour is used, to in-dicate that intelligent behaviour arises from thecollective rather than the individual. In multia-gent systems simple (mobile) agents correspond tothe living entities in a swarm. The agents interactvia artificial equivalents of pheromones, i.e. digitalmarkers in the environment. Swarm-based systemshave the following qualities [43, 44]:

2An allocation is Pareto optimal if and only if it is notpossible to improve the individual utility of an agent withoutmaking any of the others worse off.

1. Simplicity. The logic for individual agents isfairly simple. The agents are easier to programand prove correct at the level of individual be-haviour. Also, they can run on extremely smallplatforms, such as “smart dust” micro chips[45, 46].

2. Robustness. Swarm-based systems favor largenumbers of entities that are continuously or-ganizing themselves. Therefore, the systemsperformance is robust against the loss of a fewindividuals. The simplicity and low expense ofeach individual agent means that such lossescan be tolerated economically.

3. Scalability. A swarm-based system is very scal-able from the coordination point of view. Sinceagents have some form of shared behaviouralmodel about how to react to each other’s be-haviour, markers (pheromone) in the environ-ment, or external conditions, the strategy re-quires no direct or extensive communicationbetween the agents. This yields efficient inter-agent communication.

In multiagent systems swarming has been ap-plied to various applications. The agents may ei-ther be physical entities in the real world (e.g. ve-hicles), objects that travel through a computer net-work (where the nodes of the network can be seenas “nests”), or software processes that represent astate in a computation (e.g. a set of image process-ing jobs). Some examples of applications are:

Region detection Bourjot et al. [47] used thecollective web building strategy of social spi-ders3 to develop image processing algorithmsfor detecting contours and regions.

Network optimisation A model of the foragingbehaviour of ants can be used to solve rout-ing problems [48] in networks (e.g. road trafficmanagement). Alternatively, this model canbe applied for computer network managementapplications [49].

UAV control Parunak et al. [50] suggest astigmergy-based system for controlling a setof UAVs. Each UAV must effectively cover alarge search space and revisit locations regu-larly, maximizing detection probability based

3Anelosimus eximius

11

on known characteristics of the target (e.g.,visibility angle), while not exhibiting any obvi-ous systematic search patterns that would per-mit mobile targets to execute simple avoidancestrategies. Alternatively, Hawthorne et al. [51]suggest a behaviour based swarm algorithm forunmanned vehicle control.

In a broader context, swarming has come intovogue in the military to describe a battlefield tacticthat involves decentralized, pulsed attacks [52, 53].

2.5 Taxonomy overview

Figure 7 shows a diagram of the coordination tax-onomy. Although the distinction between the val-ues of some of the dimensions are not crisp, wechoose to represent the first four dimensions ofour taxonomy in a tree structure. In this tree a“branch” represents the predominant value of a di-mension.

Note that in the taxonomy diagram not all com-binations of values are exhaustively enumerated.We argue that implicit coordination strategies areprimarily applicable in cooperative environments,since competitive environments suggest explicit in-teraction (e.g. negotiation) between agents aboutallocations of objects. An exception is reactive be-haviour, not driven by explicit rules, to an oppo-nent’s actions. Examples of this are the Cold Wararms race and dynamics in predator-prey popula-tions.

Furthermore, we argue that there cannot be acentralised coordination strategy without explicitcoordination. In a centralised setting, decisions oncoordination must somehow be communicated tothe agents under control. Although this could berealised with communication-less mechanisms suchas markers in the environment (stigmergy), suchmechanisms are only applicable for one-way unad-dressed messages, rather than the two-way com-munication that full coordination requires. Hybridforms are possible, in which there is implicit coordi-nation one way, and explicit coordination the otherway.

The fifth dimension, the coordination designmetaphor, cannot be placed in such a tree struc-ture, since it is impossible to enumerate all possi-ble metaphors a designer might imagine. Therefore,we position the three metaphors discussed in sec-

Figure 7: Coordination taxonomy

tion 2.4 as “shortcuts” in the tree structure. Co-ordination strategies that are based on a biologi-cal metaphor are usually implicit (e.g. using stig-mergy), decentralised and suitable for a cooperativeenvironment. The organisational metaphor typi-cally results in explicit coordination, since agentsdelegate work, give instructions on how to performan activity, request for services, report the status ofan activity, and so on. All agents obey the author-ity structures of the organisation, and will benevo-lently execute the tasks assigned to them. There-fore these coordination strategies are applicable incooperative multiagent systems. Finally, market-based coordination is a form of explicit coordina-tion because agents negotiate or bid on the allo-cation of objects. One can also imagine that thecoordination could use implicit mechanisms suchas stigmergy, e.g. when an agent bids by leaving amarker in the environment. In section 2.4.2 we ar-gued that market-based coordination is applicablein competitive multiagent systems.

3 Coordination in NEC

In this section we identify which coordinationstrategies are suitable for realising network enabledcapabilities. We take the following approach. Firstwe identify the coordination related requirementsof NEC, in section 3.1. Then, in section 3.2 wediscuss what types of coordination strategies meetthese requirements, expressed in terms of the tax-onomy defined in section 2.3.

12

3.1 Requirements

Before we identify requirements for coordination inNEC, we assume that an infostructure is available.This infostructure connects all available assets andallows them to share information. As discussed insection 1.3, we envision an infostructure in the formof a grid of interconnected agents, where agents ful-fill the role of mediator between the grid and themilitary assets (sensor, shooters, C2 nodes). Sinceassets can only be accessed through their corre-sponding agents, all assets are shielded from eachother’s specific technical and behavioural charac-terists. In this way, agents handle interoperabilityissues between assets. Also the agents coordinateand schedule activities, and guard the availabilityof the assets. The infostructure is a prerequisite foragile deployment of assets.

As a starting point for identifying coordinationrelated requirements for NEC, we take a look atthe NEC Core Themes as defined in the UK Out-line Concept for NEC [5]. The Core Themes can beseen as general requirements to NEC. The follow-ing three themes are directly related to coordina-tion, since they concern dynamic (re)configurationof assets, cooperation between assets, task assign-ment, team formation and managing interdepen-dencies between assets and activities.

1. Flexible Working. Enabling assets torapidly reconfigure to meet changing missionneeds, allowing them to work together withminimum disruption and confusion.

2. Agile Mission Groups. Enabling the dy-namic creation and configuration of MissionGroups that share awareness and that coor-dinate and employ a wide range of systems ora specific mission.

3. Synchronized Effects. Achieving over-whelming effects within and between MissionGroups by coordinating the most appropriateassets available in the battle space through dy-namic distributed planning and execution.

The fact that three NEC Core Themes involve co-ordination, proves the importance of coordinationwithin NEC. Together, they yield the most impor-tant requirement to coordination in NEC: flexibil-ity to reorganise and reconfigure assets in order tomeet changing mission needs.

The second requirement can be derived from thefact that network enabled capabilities typically in-volve many different types of assets. Each of theseassets has distinct characteristics, concerning forinstance quality of service (e.g. track accuracy ofa sensor), computational constraints (e.g. responsetime), physical constraints (e.g. maximum velocityof a vehicle) and communication constraints (e.g.transmitter bandwidth). This heterogeneity willincrease in joint and combined operations. More-over, NEC does not require replacement of existingassets by new, network enabled assets. Rather, ex-isting assets are incorporated in the network. Thismeans that even within a single-nation force therewill be a mixture of legacy and state-of-the art as-sets. Hence, a coordination strategy cannot assumeuniform behaviour of all assets. Being able to dealwith heterogeneity is the second requirement to co-ordination in NEC.

In the military domain we must address the issueof security. In a networked environment, maliciousagents might deliberately corrupt the functional-ity of a MAS. One can think of Red Force (enemy)agents intruding a Blue Force (friendly) grid, claim-ing vital resources (comparable to a denial of ser-vice attack) or corrupting agent interactions suchthat normal operations are disrupted. A coordina-tion strategy should be resilient to malicious agents.This security requirement should be considered atthe level of coordination. We assume that basic se-curity precautions (such as authentication, autho-risation and encryption) are taken irrespective ofthe coordination strategy.

The presence of an infostructure does not im-ply a high quality communication network. Thephysical communication infrastructure, underlyinga NEC grid, is a hybrid one. It can be a mixture ofsecure phone lines, satellite communications, dig-ital radio, and so on. Between nodes there willbe strong differences in network availability, band-width and latency. Since sufficient bandwidth orlow latency is not guaranteed, a coordination strat-egy should yield minimum communication over-head at the level of coordination interoperability(see figure 5).

The next requirement is concerned with robust-ness. In general this means that if parts of thenetworked system fail, the system can proceed withdegraded functionality. Specifically, if for some rea-son assets are not available (e.g. due to damage,

13

1. Flexibility: ability to readily reorganise and re-configure assets.

2. Heterogeneity: ability to coordinate heteroge-neous assets.

3. Security: Resilience to malicious agents.

4. Communication: resilient to various quality ofservice levels.

5. Robustness: graceful system degradation or fall-back in case of malfunctions.

6. Tempo: no negative effect on tempo of opera-tions.

7. Scalability: ability to handle increased systemcomplexity.

Table 2: Requirements to coordination in NEC

malfunction or failing communications), the coor-dination strategy should come up with an alterna-tive configuration of assets to meet mission needs.Moreover, if the coordination strategy fails to comeup with a suitable configuration at all, the net-worked system should, in the worst case, degradeto a platform centric force.

One of the merits of NEC is increased tempo ofoperations because of direct interaction between as-sets. Coordinating activities will require some com-putational effort. A requirement to the coordina-tion strategy is that it shall not negatively effectthe gain in tempo.

The final requirement considered in this workinvolves scalability. Since NEC allows assets tojoin or part the network, the coordination strat-egy should be able to handle an increased numberof assets without impacting performance.

Table 2 summarizes the requirements to coordi-nation in NEC. In the next section we will evaluatewhich coordination strategies are suitable for NEC,bearing in mind the requirements identified in thissection.

3.2 Evaluation

Considering the coordination taxonomy, as de-picted by figure 7, we must first determine whetherimplicit or explicit coordination mechanisms areapplicable, given the requirements for coordinationin NEC (table 2). Implicit coordination impliesthat either all agents have a shared understand-ing on how to manage activities, or agents interact

through their environment. In both cases, thereis very little communication between agents on thelevel of coordination. This satisfies the communica-tion requirement. However, there are some majordrawbacks to applying implicit coordination in aNEC context. First, implict coordination impliesa level of homogeinity in the behaviour of agents(e.g. “ant” agents in an ant-based system). A NECsystem should be open to agents of other (coali-tion) participants. These agents are likely not tohave the same characteristics, since they have beendeveloped independently. The problem of hetero-geneity can be resolved by defining shared mod-els of coordination. Then, each agent has to haveknowledge on the behaviour of its peers. However,if the coordination strategy changes, or agents withsignificantly different behaviour enter the grid thenthese models would have to be updated. This couldresult in a substantial maintenance effort. Second,systems in which agents interact through their en-vironment (e.g. by means of stigmergy) are sub-ject to errors or manipulation. Other agents, pos-sibly with malicious intents, can deliberately or ac-cidentally change or remove the marks in the en-vironment, thereby corrupting system behaviour.In short, implicit coordination is less suitable forthe open systems we envision, because the require-ments to heterogeinity and security cannot be met.

We will now look at explicit coordination strate-gies. However, we should only consider strategieswith a low communications overhead, such that thecommunication requirement is met and tempo ofoperations is not negatively influenced. Accordingto our coordination taxonomy, there are two op-tions: coordination strategies for cooperative andcompetitive agents. In case of competitive agentswe have to deal with non-benevolent, selfish agentsthat aim at satisfying private goals. We already es-tablished in section 2.4.2 that market-based coordi-nation strategies are suitable for competitive multi-agent systems. However, there are some drawbacksto a market-based approach [19]. The first concernsthe potential complexity of the system, requiringreasoning about the bidding process (counterspec-ulation) and determining the auction’s outcome.The latter can be particularly difficult in combi-natorial auctions (i.e. agents negotiate about mul-tiple objects), which is known to be a NP-completeproblem. If the number of agents that participatein a market increases, allocation of objects might

14

be become computationally infeasible. The seconddrawback concerns security. Auctions and nego-tiations are vulnerable to cheating and manipula-tion. For example, intruding malicious agents caneasily corrupt negotiations or auctions by startingfalse negotiations or by artificially increasing theprice at auctions. Finally, auctions and negotia-tions usually involve a lot of communication, sincebids and preferences have to be communicated be-tween buyers, sellers and auctioneers. In short,there is a risk to designing NEC as a competitiveMAS since, from a coordination perspective, it ishard to meet the scalability, security and communi-cation requirements. Furthermore, computationalcomplexity threatens tempo of operations.

Consequently, we prefer to design NEC as a MASof cooperative agents, possibly using an organi-sational metaphor. To respect the scalability re-quirement, a centralised approach seems promising.Agents that perform coordination tasks are sepa-rated from agents that represent assets, so bothgroups can be scaled independently. Note that acentralised coordination approach does not meanthat a single computational node is responsible forcoordination. Rather, we envision a grid consist-ing of two types of agents: agent that coordinateand agents that are coordinated. This separation ofconcern will decrease the complexity of the agentsthat represent the assets, since they will have vir-tually no capabilities to coordinate. Moreover, therobustness constraint is respected because there isno single point of failure.

In short, for NEC we primarily prefer explicit,centralised coordination strategies with low commu-nication overhead in a cooperative agent environ-ment. In order to meet the communication andtempo requirements, the agents should be arrangedin a flat organisation with short communicationpaths. Note that we have derived general proper-ties for NEC coordination. In practice, certain clus-ters of agents in a NEC grid might require differentcoordination strategies that better fit the specificcharacteristics. In the following section we proposea coordination architecture for NEC, that allowsfor sufficient flexibility.

3.3 Proposed NEC coordination ar-

chitecture

In this section we propose a coordination architec-ture for NEC, that is based on the considerationsdiscussed in section 3.2. This coordination archi-tecture puts the coordination strategy in an organ-isational contexct. We envision a hybrid NEC co-ordination strategy, composed of a global explicit,centralised coordination strategy with low commu-nication overhead in a grid of cooperate agents,with subordinate clusters of agents that use lo-cal coordination strategies. To realise this, thereshould exist an abstraction between the global co-ordination strategy and the local clusters of agents.The global strategy is leading and treats the sub-ordinate local structures as single functional units.At the local level, a limited set of agents and assetstake responsibility for local coordination, possiblyusing a strategy that differs from the global one.

A structure that fits this description is the hol-archy, or holonic organisation. A holarchy is a self-similar organisational structure comprised of multi-leveled, grouped organisations. Each grouping, orholon, has a character derived but distinct fromthe entities that are members of the group. At thesame time, this same holon contributes to the prop-erties of one or more holons above it. The nestingof holons within a holarchy is a hierarchy in whichsubordinate holons relinquish some of their auton-omy to the superordinate holon. Figure 8 shows agraphical representation of a holarchy. Black dotscorrespond to agents, the circles represent holonsand the arrows represent control flows.

The chief characteristic of a holarchy is that re-questing agents require little or no knowledge onhow a request is handled at the local level. Sub-ordinate holons have a high degree of autonomyin how they fulfill a request. In short, a hol-archy can be seen as a hierarchy that allows someamount of cross-tree interactions and local auton-omy. This characteristic closely relates to the NECconcept of power to the edge [54], where entitiesat the bottom of a hierarchy synchronise their ac-tivities. Also, there is a resemblance between ho-larchies and federations, especially when dealingwith flat holarchies that use agents as intermedi-ates between the holons. For a concise descriptionof holarchies, hierarchies, federations and other or-ganisational paradigms, see the work of Horling et

15

Figure 8: A holarchical organisation

al. [19].If we put NEC coordination in the context of a

holarchic organisational structure, an organisationarises in which the top-level holon is formed by aglobal grid structure of agents and assets. This gridis comprised of two types of agents: agents thatcoordinate and agents that are subject to coordi-nation. Agents of the first type are solely occupiedwith coordination tasks. Although distributed in agrid, these agents form the centralised coordinationcore of the system. Agents of the latter type havelittle or no coordination capabilities at the globallevel. These agents either represents assets (sen-sors, shooters, C2 nodes) or are subordinate holonsthat are treated as single functional units at theglobal level.

The subordinate holons themselves are multia-gent systems of coherent assets. With “coherent”we mean that the assets can be managed withthe same coordination strategy, and that from theglobal perspective the assets can be considered asa functional unit. For example, consider a set ofUAVs that cooperate in a reconnaissance task. Atthe global level the coordination strategy can treatthis set of UAVs as one logical sensor that providesinformation on enemy positions, buildings or ter-rain features, without having to bother about in-dividual UAVs. The coordination of activities be-tween individual UAVs is left to the subordinateholon. This holon can use a coordination strat-egy that best fits the specific characteristics of thetasks, the agents and the assets.

Figure 9: NEC holonic coordination archtecture

Figure 9 illustrates a holonic coordination archi-tecture for NEC. Agents are represented by circles.The colouring denotes the coordination capabilitiesof an agent. At the global level agents are organ-ised in a grid. Some nodes in the grid are holonsin which agents are organised and coordinated dif-ferently. Note that communication paths are short,since agents in the global holon take direct ordersfrom the coordinating agents. Between holons com-munication paths can be kept short by limiting thenumber of nestings within holons.

Coordination at the top-level holon can be re-alised with COMPASS, or similar technology. Co-ordinating agents in the grid must have someknowledge of the available agents and assets, theircapabilities, properties and interrelations. This canbe realised with a distributed knowledge base, fromwhich coordination plans are continuously inferred.If a new agent or holon is added to the grid, anew entry is added to this knowledge base. Besidesplanning and delegating tasks to agents, coordinat-ing agents should monitor the execution of tasksand intervene if necessary.

4 Discussion

In this work we recognised the need for tacklingcoordination related issues in NEC. As more andmore military assets (sensors, shooters, C2 sys-tems) are deployed in networked structures, suchthat their combined capabilities can be fully ex-ploited, the problem of coordination gets more

16

prominent. We observed that NEC systems canbe seen as a special kind of multiagent systems.Hence, theory and concepts concerning coordina-tion explored in the field of MAS design can alsobe applied to NEC systems.

We explored theory on MAS coordination andidentified five dimensions by which coordinationstrategies can be classified, yielding a coordina-tion taxonomy. Next we identified the require-ments that have to be met by a NEC coordina-tion strategy. These requirements concern flexi-bility, heterogeneity, security, communication, ro-bustness, tempo and scalability. Based on these re-quirements we argued that explicit, centralised co-ordination strategies, with limited communicationsoverhead, in a cooperative agent environment arebest suited for managing activities in a NEC grid.Other types of coordination might still be useful forspecific clusters of agents and assets, and they canbe incorporated in a hybrid holarichal coordinationarchitecture.

Although we identified the most promising typeof strategy for global NEC coordination, many de-sign decisions for actual implementations still haveto be made. The results of this work should be seenas a starting point in the design and implementa-tion process of coordination mechanisms for NEC.Furthermore, the analysis has been purely theoreti-cal. The proposed coordination architecture shouldbe further refined and its qualities should be eval-uated by means of simulations and prototypes.

Promising technologies for this future researchare COMPASS and Cougaar. COMPASS, devel-oped by defence industry (Thales Naval Nether-lands4), is a prototype of a multi-platform middle-ware system that is aimed at the dynamic configu-ration of capabilities in network centric systems.This prototype features mechanisms for explicit,centralised coordination with limited communica-tion overhead, as required for coordination at theglobal level in our proposed architecture.

Cougaar5 is an agent platform and frameworkdeveloped by DARPA6, for the development large-scale distributed agent-based applications. Aimedat defence applications, Cougaar has some power-ful features for reliable inter-agent message trans-

4http://www.thales-nederland.nl5Cognitive Agent Architecture, http://www.cougaar.org6Defense Advanced Research Projects Agency, http://

www.darpa.gov

port, task planning and service discovery. WithinCougaar, agents can be organised in logical groupsor communities, and can be addressed based oncommunity membership. Using Cougaar’s com-munity features, agents can easily be organised inholonic organisations. Both Cougaar and COM-PASS possess specific military-standard systemqualities such as robustness, security and fail-safety.

Defence is not the only domain in which the net-work centric paradigm is explored. The merits ofa network centric approach to operations is alsorecognised in the crisis response domain. Similarto the trend in the military domain, (Dutch) crisisresponse organizations tend to shift from hierarchi-cal, regionally organized structures to dynamic as-semblies of people and resources. Research projectsthat, among other things, focus on coordinationin agile networked crisis response organisations areICIS7 and COMBINED8, both supported by theDutch Ministry of Economic Affairs and hostedby the Decis Lab9. Because of the similarities intopics, further research activities should be com-bined to establish cross-fertilisation between crisisresponse research and defence research.

References

[1] M.R. Endsley. Theoretical Underpinning ofSituation Awareness. In M.R. Endsley andD.J. Garland, editors, Situation AwarenessAnalysis and Measurement. LEA, Mahwah,NJ, USA, 2000.

[2] N. Blatt. Trust and Influence in the Informa-tion Age: Operational Requirements for Net-work Centric Warfare. In Proceedings of the10th International Command and Control Re-search and Technology Symposium, McLean,Virginia, USA, June 2005.

[3] D.S. Alberts, J.J. Garstka, R.E. Hayes, andD.A. Signori. Understanding Information Age

7Interactive Collaborative Information Systems, http://www.icis.decis.nl

8Chaotic Open world Multi-agent Based IntelligentlyNEtworked Decision support Systems, http://combined.

decis.nl9Delft Cooperation on Intelligent Systems, http://www.

decis.nl

17

Warfare. DoD C4ISR Cooperative ResearchProgram, August 2001.

[4] D.S. Alberts, J.J. Garstka, and F.P. Stein.Network Centric Warfare: developing andleveraging information superiority. DoDC4ISR Cooperative Research Program, 2ndedition, 1999.

[5] Ministry of Defence. Outline Concept for theUK Network Enabled Capability (nec), May2003.

[6] B.W. Young. Future Integrated Fire Con-trol. In Proceedings of the 10th InternationalCommand and Control Research and Technol-ogy Symposium, McLean, Virginia, USA, June2005.

[7] S. Franklin and A. Graesser. Is it an agent, orjust a program? A taxonomy of AutonomousAgents. In Proceedings of the Third Interna-tional on Agent Theories, Architectures andLanguages, 1996.

[8] H. Nwana. Software Agents: An Overview.Knowledge Engineering Review, 11(3):205–244, 1996.

[9] J.M. Bradshaw, editor. Software Agents.AAAI Press, 1997.

[10] N. Jennings. On Agent-Based Software En-gineering. Artificial Intelligence, 177(2):277–296, 2000.

[11] M. Wooldridge. An Introduction to MultiagentSystems. John Wiley and Sons Ltd., Chich-ester, UK, 2002.

[12] A. Bond and L. Gasser. Readings in Dis-tributed Artificial Intelligence. Morgan Kauf-mann publishers Inc., San Mateo, CA, USA,1988.

[13] Thomas W. Malone and Kevin Crowston.What is coordination theory and how can ithelp design cooperative work systems? InCSCW ’90: Proceedings of the 1990 ACMconference on Computer-supported cooperativework, pages 357–370. ACM Press, 1990.

[14] N.R. Jennings. Coordination techniques fordistributed artificial intelligence. In Founda-tions of distributed artificial intelligence, pages187–210, New York, NY, USA, 1996. John Wi-ley & Sons, Inc.

[15] D. Corkill and S. Lander. Agent Organiza-tions. Object Magazine, 8(4), April 1998.

[16] A. Omicini, F. Zambonelli, M. Klusch, andR. Tolksdorf, editors. Coordination of Inter-net Agents: Models, Technologies, and Appli-cations. Springer, 2001.

[17] C.J. van Aart. Organizational Principles forMulti-Agent Architectures. Whitestein Seriesin Software Agent Technologies. Birkhauser,Basel, Switserland, 2005.

[18] S. Willmott. Coordination Structures forthe Intelligent Control of Dynamic DistributedSystems. PhD thesis, Ecole PolytechniqueFederale de Lausanne, 2002.

[19] V. Lesser B. Horling. A Survey of Multi-agent Organizational Paradigms. The Knowl-edge Engineering Review, 19(4):281–316, 2005.

[20] ISO Open Systems Interconnection ReferenceModel. http://www.iso.org.

[21] Sascha Ossowski. Co-Ordination in Artifi-cial Agent Societies: Social Structures and ItsImplications for Autonomous Problem-SolvingAgents. Springer-Verlag New York, Inc., Se-caucus, NJ, USA, 1999.

[22] R. Beckers, O. Holland, and J. Deneubourg.From local actions to global tasks: Stigmergyand collective robotics. In Proceedings of the4th International Workshop on the Synthesisand Simulation of Living Systems, pages 181–189, Cambridge, MA, USA, 1994.

[23] E. Bonabeau, M. Dorigo, and G. Theraulaz.Swarm Intelligence: From Natural to ArtificialSystems. Santa Fe Institute Studies in the Sci-ences of Complexity. Oxford University Press,New York, NY, USA, 1999.

[24] O. Holland and C. Melhuish. Stigmergy, self-organization, and sorting in collective robotics.Artificial Life, 5(2):173–202, 1999.

18

[25] Z. Mason. Programming with stigmergy: us-ing swarms in construction. In Artificial LifeVIII: Procedings of the Eighth InternationalConference on Artificial Life, pages 371–374,Sydney, Australia, 2002. MIT Press.

[26] T.W. Malone and K. Crowston. The interdis-ciplinary study of coordination. ACM Com-puting Surveys, 26(1):87–119, 1994.

[27] N.R. Jennings, P. Faratin, A.R. Lomuscio,S. Parsons, C. Sierra, and M. Wooldridge. Au-tomated Negotiation: Prospects, Methods andChallenges. International Journal of GroupDecision and Negotiation, 10(2), 2001.

[28] Y. Chevaleyre, P. E. Dunne, U. Endriss,J. Lang, M. Lemare, N. Maudet, J. Pad-get, S. Phelps, J. A. Rodruez-Aguilar, andP. Sousa. Issues in Multiagent Resource Allo-cation. Informatica, 2005. Accepted for pub-lication.

[29] P. Giorgini, M.Kolp, and J. Mylopoulos.Multi-Agent Architectures as OrganizationalStructures. Journal of Autonomous Agentsand Multi-Agent Systems, 2005. in press.

[30] C. Jones, D. Shell, M.J. Mataric, andB. Gerkey. Principled Approaches to the De-sign of Multi-Robot Systems. In Proceed-ings of the Workshop on Networked Robotics,IEEE/RSJ International Conference on Intel-ligent Robots and Systems (IROS-04), Sendai,Japan, 2004.

[31] M. Yokoo, E.H. Durfee, T. Ishida, andK. Kuwabara. The Distributed ConstraintSatisfaction Problem: Formalization and Al-gorithms. IEEE Transactions on Knowledgeand Data Engineering, 10(5):673–685, 1998.

[32] H. Jung, M. Tambe, and S. Kulkarni. Argu-mentation as distributed constraint satisfac-tion: applications and results. In AGENTS’01: Proceedings of the fifth international con-ference on Autonomous Agents, pages 324–331. ACM Press, 2001.

[33] T. Khedro and M.R. Genesereth. ModelingMultiagent Cooperation as Distributed Con-straint Satisfaction Problem Solving. In ECAI,pages 249–253, 1994.

[34] R. G. Smith. The contract net protocol:High-level communication and control in a dis-tributed problem solver. IEEE Transactionson Computers, pages 1104–1113, 1980.

[35] W. Vickrey. Counterspeculation, auctions,and competitive sealed tenders. Journal of Fi-nance, pages 8–37, 1961.

[36] H. Mintzberg. Structures in fives: designingeffective organizations. Prentice Hall, Engle-wood Cliffs, NJ, USA, 1993.

[37] M.S. Fox. An organizational view of dis-tributed systems. In Distributed Artificial In-telligence, pages 140–150, San Francisco, CA,USA, 1988. Morgan Kaufmann Publishers Inc.

[38] A. Fuxman, P. Giorgini, M. Kolp, and J. My-lopoulos. Information systems as social struc-tures. In Proceedings of the 2nd InternationalConference on Formal Ontologies for Informa-tion Systems, FOIS01, Ogunquit, USA, Octo-ber 2001.

[39] M. Kolp. Organizational Structures forMulti-Agent Architectures: Application toLego Mindstorms Mobile Robots. IAGWorking Paper 28/02, January 2002.http://www.cs.toronto.edu/ mkolp/pubs.html.

[40] T. T. Do, S. Faulkner, and M. Kolp. TheStructure-in-5 as an Agent Architectural Pat-tern. Working Paper IAG 38/02, May 2002.http://www.cs.toronto.edu/ mkolp/pubs.html.

[41] U. Endriss, N. Maudet, F. Sadri, and F. Toni.On optimal outcomes of negotiations overresources. In Proceedings of the 2nd In-ternational Joint Conference on AutonomousAgents and Multiagent Systems (AAMAS),pages 177–184. ACM Press, 2003.

[42] G. Beni and U. Wang. Swarm intelligencein cellular robotic systems. NATO AdvancedWorkshop on Robots and Biological Systems,Il Ciocco, Tuscany, Italy, 1989.

[43] H.D.V. Parunak and S.A. Brueckner. Engi-neering Swarming Systems. In F. Bergenti,M.P. Gleizes, and F. Zambonelli, editors,Methodologies and Software Engineering forAgent Systems, Amsterdam, The Netherlands,2004. Kluwer.

19

[44] J.A. Sauter, R. Matthews, H.D. Parunak,and S. Brueckner. Performance of digitalpheromones for swarming vehicle control. InProceedings of the 4rd International JointConference on Autonomous Agents and Multi-agent Systems (AAMAS 2005), pages 903–910.ACM Press, July 2005.

[45] K.S.J. Pister, J.M. Kahn, and B.E. Boser.Smart Dust: Wireless Networks of Millimeter-Scale Sensor Nodes. Highlight Article in1999 Electronics Research Laboratory Re-search Summary.

[46] J. M. Kahn, R. H. Katz, and K. S. J. Pis-ter. Next century challenges: mobile network-ing for Smart Dust. In MobiCom ’99: Pro-ceedings of the 5th annual ACM/IEEE inter-national conference on Mobile computing andnetworking, pages 271–278. ACM Press, 1999.

[47] C. Bourjot, V. Chevrier, and V. Thomas. Anew swarm mechanism based on social spi-ders colonies : from web weaving to region de-tection. Web Intelligence and Agent Systems,1:47–64, 2003.

[48] G. Di Caro and M. Dorigo. AntNet: A MobileAgents Approach to Adaptive Routing. Tech-nical Report 97-12, IRIDIA, Universit Libre deBruxelles, 1997.

[49] N. Foukia and S. Hassas. ManagingComputer Networks Security through Self-Organization: A Complex System Perspec-tives. In G. Di Marzo Serugendo, A. Karageor-gos, O.F.Rana, and F. Zambonelli, editors,The First International Workshop on Engi-neering Self-Organising Applications (ESOA).Springer Verlag, July 2004.

[50] H.V.D. Parunak and S. Brueckner. Swarm-ing Coordination of Multiple UAV’s for Col-laborative Sensing. In Proceedings of the Sec-ond AIAA ’Unmanned Unlimited’ Systems,Technologies, and Operations Conference, SanDiego, CA, USA, 2003.

[51] C. Hawthorne and D. Scheidt. Moving emer-gent behavior algorithms from simulation tohardware: Command and control of au-tonomous uxvs. In Proceedings of the 10th

International Command and Control Researchand Technology Symposium, McLean, Vir-ginia, USA, June 2005.

[52] John Arquilla and David Ronfeldt. Swarmingand the Future of Conflict. RAND NationalSecurity Research Division (NSRD), http:

//www.rand.org, Documented Briefings, no.DB-311-OSD, 2000.

[53] S.J.A. Edwards. Swarming on the Battlefield:Past, Present, and Future. RAND NationalSecurity Research Division (NSRD), http:

//www.rand.org, Documented Briefings, no.MR-1100-OSD, 2000.

[54] D.S. Alberts and R.E. Hayes. Power to theEdge. Information Age Transformation Series.DoD C4ISR Cooperative Research Program,2003.

20