A Group Selection Pattern for agent-based Virtual Organizations coordination in Grids Isaac Chaoa, Oscar Ardaizb, Ramon Sangüesaa a

Computer Architecture Department, Polytechnic University of Catalonia, Spain {ichao,sanguesa}@lsi.upc.edu

b

Department of Mathematics and Informatics, Public University of Navarra, Spain [email protected]

Abstract. A key challenge in Grid computing is the achievement of efficient and self-organized management of the Virtual Organizations composing the system. Grids are often very heterogeneous, incorporating high dynamicity and unpredictability. Introducing higher levels of adaptation and learning in the coordination protocols may help coping with complexity. We provide a solution based on a self-organized and emergent mechanism evolving congregations of policy-based resource management agents through a Group Selection process. We provide a formalization of the Group Selection pattern; we show how the mechanism fits in a Service Oriented Grid infrastructure and further evaluate by simulation its performance as an agent’s policy coordination mechanism in Virtual Organizations.

Keywords: Automatic Resource Allocation, Virtual Organizations Management, Group Selection pattern, Service Oriented Grids

1. Introduction Virtual organizations (VOs) are dynamic collaborative collections of individuals, enterprises, and information resources [Foster, 2001]. VOs are key enablers of the Grid Computing vision. VOs require very flexible sharing relationships. They are formed with the goal of performing resource sharing and coordinated problem-solving in dynamic, multi-institutional environments. VOs can be created to undertake their role for a very brief period of time, or exist for a longer term. They may be created on demand in dynamic, open and competitive environments. The following are examples of VOs: the application service providers, storage service providers, cycle providers, and consultants engaged by a car manufacturer to perform scenario evaluation during planning for a new factory; members of an industrial consortium bidding on a new aircraft; a crisis management team and the databases and simulation systems that they use to plan a response to an emergency situation; and members of a large, international, multiyear high energy physics collaboration. The more we want to allow for loosely coupled interactions, by conceding increased autonomy to the VO members, the more we lack in control over the VO emergent properties. The more we constraint VO participants freedom of choice, the more we prevent from VO formation between un-trusted potential new partners, inhibiting the full realization of the Grid vision in open systems (as precisely defined in [Sycara, 1998]). The Grid is intended to span across large and heterogeneous administrative domains, requiring for the Grid management system a high degree of scalability and flexibility. The bigger the scale, heterogeneity and dynamism of the target Grid, the more important becomes the inclusion of self-* properties. VOs have several key properties that distinguish them from traditional IT architectures: Autonomy of its members, which behave independently, constrained only by their contracts; heterogeneity of its members, which are independently designed and constructed, constrained only by the applicable interface descriptions; dynamism of the VO members which can join and leave with

minimal constraints, affecting the configuration of a VO at runtime; structure, with VOs having complex internal structures, reflected in the relationships among their members. Importantly, even in cases where the above properties are not required (such as within an enterprise where the members are controlled by one party), it is appropriate to architect a VO as if it had the above properties. Policy management has been considered in distributed systems for many years. The IETF Policy framework was designed for managing network resources [IETFWG, 2000], but has a structure that applies more generally. This architecture assumes a network element (router, hub, or switch) where resource allocation decisions are made. A policy enforcement point or PEP resides within the network element. A policy decision point or PDP may be within or outside the element; policy decisions are made here. A PDP may use additional mechanisms to decide on an action. In Grid Computing, middleware toolkits provide several low levels tools to aggregate resources in federated directories and management access permissions [Foster, 2005]. The problem with traditional policy management consists in forcing the Grid administrator to perform management using Grid toolkits and to manually interact with network administrators in each domain to guarantee that the underlying network is properly configured for the Grid operation. This leads to a situation where new Grid requirements imply manual coordination between the Grid and network administrators is. The support provided by Grid toolkits to solve this situation is very limited and, in most cases, not even exist [Wasson, 2003]. The evolution of Grid Computing to more dynamic and heterogeneous systems imposes serious scalability and manageability issues to current policy-based management toolkits. Addressing those issues will require a shift in engineering approach, from traditional policy management, to emergent, self-organizing policy management. Group Selection refers to a process of natural selection that favors characteristics in individuals that increase the fitness of the group the individuals belong relative to other groups. Group Selection implies that every member of the group depends on a group characteristic that is not isolated in a single individual [Wilson, 1975]. Partitioning the population in groups of interaction heavily impact the coordination. Such groups form isolated niches where the sub-populations are allowed to evolve behaviors independently of the rest of the populations. The existence of niches maintains a large diversity in an evolving population since the evolutionary paths in separated niches may develop in entirely different ways. In this paper we provide a coordination mechanism based on Group Selection which can be used standalone or plugged into existent coordination architecture in order to optimize its performance. The key idea is that biasing interaction between Grid nodes by arbitrary identifiers enables efficient grouping of agents. Further group’s evolution through Group Selection optimizes groups in performance. Moreover, this grouping evolution trough inter-group migration inherits from the mechanism many desirable properties: it is adaptive, self-organized, decentralized and highly scalable. Additionally, this solution can be used to automatically manage Grid VOs lifecycle in the Grid, hence providing a valid mechanism for VO formation, evolution and disbanding which proves also self-organized and decentralized. We test by simulation the performance of the mechanism in a agents policy alignment scenario in VOs. The rest of the paper is organized as follows. In section 2 we review related work. In section 3 we introduce the Group Selection pattern and its instantiation for policy-based VO management. In section 4 we provide experimental results extracted by simulation and evaluation. Section 5 concludes the paper.

2. Related Work Many projects are using VOs conceptually, but very few projects are addressing the management of VOs themselves. While the notion of a VO seems to be intuitive and natural [Camarinha, 2003], we still do not have clear definitions of what constitutes a VO or well-defined procedures for deciding when a new VO should be formed, who should be in that VO, what they should do, when the VO should be changed, and when the VO should ultimately be disbanded. In Conoise-G project [Patel, 2005] an agent system supporting robust and resilient VOs formation and operation is presented. Another project focusing on Trust issues is Tustcom [Trustcom, 2005] aiming to provide a trust and contract management framework enabling the definition and secure enactment of collaborative business processes within VOs that are formed on-demand, self-managing and evolve dynamically. In

both Conoise-G and Trustcom approaches to VO management, components for helping automated VO management are developed, but no specific self-organization mechanism is provided. Self-organization mechanisms incorporating emergence bring into VO management higher levels of flexibility and adaptability, providing much more generic models, applicable for a wider range of scenarios. Since a VO is only a temporary conglomerate that is established to quickly react to complex demands of the market, it is mandatory that every decision on the inter-enterprise level (horizontal integration), is propagated through and reflected in all levels down to the lowest level, the machine level (vertical integration). This has lead to the concept of a holonic enterprise [Ulieru, 2002]. A holon is an autonomous and cooperative building block of a system that has a unique identity, yet may be made up of sub-ordinate parts and in turn may be part of a larger whole. The concept of holons enables the construction of very complex systems that are nonetheless efficient in the use of resources, highly resilient to internal and external disturbances, and adaptable and flexible in the face of changes in the environment in which they exist. To the extent of our knowledge this is the only approach for VO management focusing on emergence and self-organization. As for policy-based resource management in VOs, most of the Grids deployed both in scientific research and industry work use convenient Grid middleware. The Globus Toolkit [Foster, 2005] has emerged as a “de facto standard”. Globus toolkit offers a set of low level services which can be used to publish, discover, monitor and meta-schedule resources on remote nodes in the Grid. However by no means it specifies mechanism to manage Grid VOs. The Globus Grid Security infrastructure (GSI) maps user identities to local user accounts, where access permissions are defined by the local system administration. This is shown to be too restrictive to support VOs in collaborative use of services The Globus Project has produced a Community Authorization Service (CAS) [Foster, 2003], which uses a push model to provide access permissions using X.509 extensions. The certificate extension contains attributes that detail a user’s permissions to Grid resources, including such low level details as to read/write file permissions. This micro management would be very difficult to maintain across a large Grid and almost impossible across organizations. Another push model system is the Virtual Organization Management System (VOMS) [Alfieri, 2003]. Similar to CAS, the user connects to the VOMS server and supplies the user with an extended X.509 certificate. In this case, VOMS extends the certificate with role and group attributes. The resource authenticating the certificate then needs to know the access policies for the roles and groups in order to make authorization decisions. CAS and VOMS are frameworks to managing and distributing attributes about authorization. Akenti [Thompson, 2003] is a policy engine providing a decision on a users’ request. It uses a pull model to obtain policies to certain permissions based on a user’s identity. Some of the systems described are able to work together, such as CAS and Akenti. However, all of these are focused on supporting large static communities and static resources. Alternative scenarios may contain a large number of users and requiring fine-grained access control policy for service instances that are shared between group members. The policy needs to specify user identities of the respective roles and must be dynamic. In this case, the access control policy for the group is an emergent property of the distributed policies for service instance access. The mechanisms described above do not provide the means to control the distributed policy across services and across organizations. Exploiting group structure in multiagent systems (MAS) has been proposed in previous research. Coalitions are formed by subsets of the population, and in general are goal oriented and short-lived. Coalitions have been studied in game theory community for decades, and can be composed of both cooperative and self-interested agents. Most of the coalition formation literature attempt to formalize optimal grouping mechanism for agent’s populations. Major limitations of these algorithms are a high computational complexity, and unrealistic assumptions regarding the availability of information [Shehory, 2004]. These issues prevents from practical usage of these mechanism in large scale scenarios such as Grids. An alternative group formation mechanisms proposed in MAS body of research is congregations. Congregations are subgroups in the agent population which have a defined purpose and organizational cost, though still releasing the full autonomy to agents. They are applied to electronic markets in [Brooks, 2002]. They are proved to serve as market optimizers. However we identify an important limitation in the congregation models proposed. Groups are static and agents can trade in just a specified number of subgroups. In contrary, Group Selection approaches enable for a dynamic view of the system, evolving the required number of subgroups depending on the changing agent’s requirements.

An important drawback in the group formation mechanism presented above is that the dynamical view of the system is not addressed. Normally the optimal groups are calculated at some computational costs, and entering in a new domain application requires a complete recalculation of groupings from scratch. An exception to this rule is the work in [Merida-Campos, 2004], where iterative formation of multiple coalitions is attempted in response to a dynamic task environment. In general, the proposed solutions address the calculation of optimal groups centrally, supposing complete system knowledge to the central coalition-maker, with few exceptions such as the mentioned work in [Ulieru, 2002]. This panorama contrast with realistic environment is nowadays large scale distributed systems, where small, decentralized components need to deal autonomously with coordinated decision making. Group Selection is a fully decentralized mechanism which focuses in the dynamic view of the groups, iteratively ruling its evolution towards more optimal configurations. It has been shown that Group Selection can lead to the spread of group beneficial characteristics in many different grouped settings of agent’s populations [Boyd, 2002]. This enables the application of Group Selection processes in any group-like structured agent population. VOs in computational Grids are a concrete case of such groupstructured MAS.

3. The Group Selection pattern

3.1. Pattern definition A software engineering design pattern is a general repeatable solution to a commonly occurring problem in software design. This is a description or template for how to solve a problem that can be used in many different situations. Patterns popularized after the publication of the classic book by Gamma et al. [Gamma, 1997]. Patterns in computer science have been used also coming from other disciplines. A relevant case is for bio-inspired computing, see Babaoglu et al. [Babaoglu, 2006]. They state in order to motivate the proposal of a family of design patterns coming from biology: “The motivation of the present work is that large-scale and dynamic distributed systems have strong similarities to some of the biological environments. This makes it possible to abstract away design patterns from biological systems and to apply them in distributed systems. In other words, we do not wish to extract design patterns from software engineering practice as it is normally done. Instead, we wish to extract design patterns from biology, and we argue that they can be applied fruitfully in distributed systems”. Another filed which has inspired several software design patterns is sociology [Edmonds, 2005]. Since Grids are naturally composed in VOs, a basic group unit already exists. The Group Selection process operates trough natural selection in several group-structured systems in nature: biological systems, evolving group-advantageous behaviors; humans societies, promoting high levels of cooperation [Bowles, 2004]; and economies, promoting the emergence of leading firms [Gowdy, 2003]. We want to build on these “good properties” of the mechanism to port the Group Selection process to an engineering pattern usable in large scale distributed systems amenable to group structure, such as Grids. Building on the experience gained by Babaouglu et al., we provide an algorithmic approach to the pattern (Figure 1) which can be instantiated in different “flavours” by simple variation of Interaction and Migration rules. The proposed synchronous algorithmic realization does not prevent for application in a realistic, asynchronous environment, since there is not any synchronization step required to update agent’s strategies and group membership.

APPLY INTERACTION RULE

Bootstrap agents in groups LOOP a number of rounds LOOP each group

STAY IN MY GROUP

No

LOOP each agent in the group (operation phase)

MIGRATION CONDITION?

Apply Interaction rule with another agent/agents from the group Collect Payoff

Yes

ENDLOOP APPLY MIGRATION RULE

ENDLOOP LOOP each agent in the population (evolution phase)

STAY WITH MY CURRENT STRATEGY (static strategies /policies)

No

CHANGE STRATEGY /POLICY?

Yes

CHANGE MY CURRENT STRATEGY (dynamic strategies /policies)

Select partner agents in the population Apply Migration rule ENDLOOP ENDLOOP

(a)

(b)

Figure 1: Group Selection pattern: (a) Individual agent flowchart; (b) Algorithmic representation for structured population of agents

First the agent’s population is bootstrapped (randomly or any pre-configured arrangement) into a number of VOs. Two phases are then executed: Policy Coordination Interaction rule: Two basic modalities of interaction are possible. Bilaterally for each pair of agents (depending on their policies) with payoffs obtained individually by each agent; or collective interaction inside the group, with a payoff shared equally between agents composing the group; also more elaborated payoff sharing scheme are possible (might be a combination of the other two). Policy Coordination Migration Rule: Several learning methods are available. Agents compare with external agents from other groups and migrate to groups hosting outperforming agents (copying or not the other agent strategy); agents inspect environment gathering relevant information on suitable groups and then decide a target groups based on some internal rules; agents inspect internally their own performance in the last interactions and then decide if explore randomly new groups or stay in the current group; others more elaborated. Regardless of how utilities are derived or migration is performed; the important thing to keep “inside the pattern” is that those two phases must be present. Varying interaction and migration rules we get different instantiations of the Group Selection pattern, producing different coordination mechanisms.

3.2. Group Selection pattern deployed as a VO management service In a service infrastructure, the Group Selection pattern needs to be deployed as a support service for the existent VO coordination services in a Service Oriented Grid (SOG). Figure 2 describes the main steps in the interaction trough the access point. When a client issues a request for service or a workflow, the application determines which Grid services are required to fulfill it. These Grid services represent either software services (e.g. a data processing algorithm) or computational resources. The application invokes the Group Selection service, which in turns transfers the request to the VO access point which parses the corresponding request. The access point further triggers the logging of the new Agent in the VO, and monitors the policy-based resource sharing activities within the VO.

GRID

Client

Request Core VO Management Services

Application Service / Workflow request (WS-Agreement)

Group Selection Service

1- Interaction Phase

VO Access Point (WS)

VO i of size N

2- Migration Phase → inter-VOs re-arrangement Policy Based Resource Man Agent 1

Policy Based Resource Man Agent 2

Policy Based Resource Man Agent 3

Service Provider 1 Policy Based Resource Man Agent 4

Service Provider n

Group Selection Service in a SOG Infrastructure Figure 2. Group Selection Service deployed in a SOG

The two phases of the group selection pattern are executed in turns: First the interaction phase, where the resources are allocated following the agent’s policies in the VO. Once this phase is completed, after utility calculations a migration phase takes places which re-arrange agents in VOs and update policies. This process, dynamically adapts VO memberships of the agents in order to maximize coordination. Depending on the scenario, coordination will be achieved by different agent’s groupings and adaptations. The pattern could operate as a policy-management mechanism implemented in realistic VOs with the following identifications: The PEP resides on each VO, which is able (trough a coordinator or any other more decentralized mechanism) to enforce the commitments of agents to the emergent VO policies (this should not undermine agent’s autonomy since only autonomously emerged policies are to be enforced). PDP´s are not explicitly represented in any member, and instead are distributed in all the agents composing the VO, since the decision of the selected VO policy is an emergent property of the co-evolution process of agents in each VO.

4. Optimizing Policy-based VO management trough Group Selection 4.1 Policy-based resource management pattern instantiation In our VO model, each agent (representing an organization) has a policy A = A(p) , from a set of M policies. A VO consists of a set of agents. VO ={A1(p1), A2(p2), A3(p3)….}. The VO defines the scope of agent operation; Policy based resource sharing utility gets maximized by coordinating the policies of N agents forming each VO. The objective is achieving policy coordination in each of the VOs in the system, forming clusters of compatible policies. The compatibly depends on the specific scenario and is measured differently depending on the scenario. For the experiments here we have implemented the simplest of the collective interactions, corresponding to a VO policy alignment scenario. The payoff is calculated on the alignment level over the whole VO and payoffs are shared collectively. As for the migration phase, the agents compare their performance against their own past performance (internal learning). Migration to a group implies the copying of the policy of one random agent in this target group. This maps VOs configurations of large pools of resources optimizing their performance by acting together using a similar policy. The metric we employ to measure the alignment degree is the Shannon Entropy Index. N

I=

∑p i =1

i

⋅ log p i

log 2 ⋅ M In this equation pi stands for one of the policies in a set of size M. Minimizing the index is equivalent to maximizing homogeneity in VOs. We reverse this measure to have a performance scale from 0 to 1: This gives 1 when all policies in the group are aligned (minimal entropy) and 0 when all present policy in the group is represented equally (maximal entropy). The goal is to minimize diversity (entropy) within each VO, achieving the highest policy alignment possible inside each VO. 4.2. Experimental results and evaluation The experiments are conducted in an open source, generic agent-based Grid simulator specifically built for developing agent coordination mechanism on top of Grids [AgentGridSim, 2007]. The models explained in this paper are included as scenarios for the Grid simulator and its source code can be inspected. The experiments conducted here are fully repeatable by downloading the simulator in the provided URL. The total number of agents is N=100, distributed originally in 5 groups of 20 agents each. The set of different policies is M=10. Two parameters rule the dynamics of the mechanism: First, the migration probability: The probability at which the agents the apply migration rule, the rate at which the agent is testing the existence of better congregations that its current. Higher migration probabilities tend to reduce the number of groups since better performing groups get crowded at a higher rate, and less performing groups get extinct quicker The other parameter, the mutation probability, rules the extent to what the agent decides to explore a brand new group, starting a group on its own and waiting for others to join. It is important to notice the importance of this parameter, since setting this to 0 would normally provoke a quick convergence to one single group. Mutation needs to be enabled to introduce variability. In these experiments, mutation probability is fixed to a small value of 0.01. We see from Figure 3 that reaching a high alignment (low entropy) up 0.8 is possible before 1000 rounds and maintained afterwards for a migration probability of 0.3. Figure 4 shows that the number of VOs oscillates in this case between 10 and 20. If we use a higher migration rate of 0.7, this implies a higher ration migration/mutation and consequently less number of groups in the system in average. Increasing the migration rate achieved a worse performance as seen from Figure 3. As we can see from figure 4, larger number of small groups (with migration probability of 0.3) generates better coordination. .

migration=0.3 migration=0.7

1930.0

1810.0

1690.0

1570.0

1450.0

1330.0

1210.0

1090.0

970.0

850.0

730.0

610.0

490.0

370.0

250.0

130.0

1 0,9 0,8 0,7 0,6 0,5 0,4 0,3 0,2 0,1 0 10.0

Utility [0,1]

Collective Coordination: AverageResourceSharingUtility (inverse Shannon Entropy Index)

sim clock tick

Figure 3. Resource Sharing Utility, collective interaction for two migration/mutation rates

migration=0.3

Collective Coordination: Number of groups

migration=0.7

25

# groups

20 15 10 5 1990.0

1880.0

1770.0

1660.0

1550.0

1440.0

1330.0

1220.0

1110.0

1000.0

890.0

780.0

670.0

560.0

450.0

340.0

230.0

120.0

10.0

0

sim clock tick

Figure 4. # Group present, collective interaction for two migration/mutation rates

We can study more in deep what is the shape of these small groups. We consider the run leading to the best scoring performance in policy alignment (migration probability 0.3), and we plot histograms showing the distribution of agents in groups for two different timestamps: tick 250, when coordination is still increasing, and tick 1500, when coordination close to 0.8 is firmly stabilized. We see from Figure 5 that by tick 250, two groups of 10 agents and other two groups of 9 agents still exist in the Grid. By tick 1500, no group is bigger than 7 agents. This means having larger number of VOs of small number of agents each VO. In each of these small VOs, self-organized alignment towards a common policy is easier. The mechanism can automatically regroup agent’s upon any perturbation, generated internally in any VO, or triggered by external influence (e.g. by the arrival of new agents into the Grid).

by tick 250(growing alignment)

Groups Sizes distribution for experiment migration prob= 0.3

by tick 1500 (stable alignment)

9 8 # of groups

7 6 5 4 3 2 1 0 1

2

3

4

5

7

9

10

Size of the group

Figure 5. # Group Sizes distribution for the experiment with migration probability =0.3

5. Conclusions In large scale Grids, system dynamicity and uncertainty are high; automatic, decentralized and selforganized control becomes a requirement. Our proposal is that a simple, rather powerful coordination mechanism based on Group Selection can be used to self-organize a set of agents in VOs to coordinate more effectively their resource sharing polices. Additionally to its effectiveness coordinating activities in groups, the mechanism is modular enough to be incorporated as generic coordination mechanisms into loosely coupled Service Oriented Architectures (SOAs). Group Selection does not impose any constraint on system size or computational requirements, hence enabling for high scalability in both physical and organizational dimensions. Clustering in small groups is a tendency largely observed in all kind of human organizations [Levine, 2005]. In cooperation building scenarios, it has been shown that smaller group sizes ease cooperation in both social-networks based cooperation and group selection based evolution of cooperation [Nowak, 2006]. Our results for the evaluation of Group Selection in Grid coordination scenarios suggest that the same conclusion applies in fully cooperative domains: “Small and dynamic groups of agents evolved trough Group Selection optimize better fully cooperative coordination scenarios”. In our Group Selection pattern, the migration to mutation rate determines the average number of dynamic groups in the systems. A rate tuned to evolve dynamic and small VOs achieves the best optimization in a policy alignment scenario. Future work includes extending the number of coordination scenarios those where diversity or complementary policies are required, such as for example a VO requiring more complex workflow compositions. Also, considering Grid users and services belonging to many VOs simultaneously, hence trading in different market segments at the same time is an scenario closer to the expect shape of realistic Grids. Deploying the mechanism in real VO prototypes can provide further validation of the pattern.

References [AgentGridSim, 2007] https://sourceforge.net/projects/agentGridrepast [Alfieri, 2003] Alfieri R., Cecchini R., Ciaschini V., dell’Agnello L., Frohner A., Gianoli A., L˝orentey K., and Spataro F., 2003. “VOMS, an authorization system for virtual organizations”, DaTaGrid. [Babaoglu, 2006] O. Babaoglu, G. Canright, A. Deutsch, G. Di Caro, F. Ducatelle, L. Gambardella, N. Ganguly, M. Jelasity, R. Montemanni, A. Montresor and T. Urnes. Design Patterns from Biology for Distributed

Computing. In ACM Transactions on Autonomous and Adaptive Systems, vol. 1, no. 1, 26--66, September 2006. [Boyd, 2002] Boyd, R. and Richerson, P. (2002) “Group Beneficial Norms Can Spread Rapidly in a Structured Population.” Journal of Theoretical Biology 215. 287–296 [Bowles, 2004]Bowles S., Gintis H. (2004). The Evolution of Strong Reciprocity, Theoretical Population Biology 65, 2004, 17-28 [Brooks, 2002] Chistopher H. Brooks and Edmund H. Durfee. Congregating and market formation. In Proceedings of the 1st International Joint Conference on Autonomous Agents and MultiAgent Systems, pages 96-103, 2002. [Camarinha, 2003] Camarinha-Matos, LM and Afsarmanesh, H. (2003) A Roadmap for Strategic Research on. Virtual Organisations, in Proceedings of PRO-VE 2003, 33-46, Kluwer [Chao, 2004] Isaac Chao,Ramon Sangüesa and Oscar Ardaiz, Design, Implementation and Evaluation of a Resource Management Multiagent.System for a Multimedia Processing Grid, Workshop on Grid Computing and Its Application to Data Analysis (GADA) On the Move to Meaningful Internet Systems 2004. [Edmonds, 2005] Edmonds, B., Gilbert, N., Gustafson, S., Hales, D. and Krasnogor, N. (eds.) (2005) Socially Inspired Computing. Proceedings of the Joint Symposium on Socially Inspired Computing, University of Hertfordshire, Hatfield, UK 12 - 15 April 2005, Published by AISB [Foster, 2001] I. Foster, C. Kesselman, S. Tuecke. The Anatomy of the Grid: Enabling Scalable Virtual Organizations. International J. Supercomputer Applications, 15(3), 2001 [Foster, 2003] Foster, I., et al. The Community Authorization Service: Status and future. in CHEP 03. 2003. La Jolla, California. [Foster. 2005] Globus Toolkit Version 4: Software for Service-Oriented Systems. I. Foster. IFIP International Conference on Network and Parallel Computing, Springer Verlag LNCS 3779, pp 2-13, 2005 [Gamma, 1997] Gamma, Erich; Richard Helm, Ralph Johnson, and John Vlissides (1997). Design Patterns CD. ISBN 0-201-63498-8.

[Gowdy, 2003] Jonh Gowdy and Irmi Seidl. Economic Man and Selfish Genes: The Relevance of Group Selection to Economic Policy,” Journal of Socio-Economics 33(3), 2004, 343-358. [IETFWG, 2000] R. Yavatkar, D. Pendarakis, and R. Guerin. A framework for policy-based admission control. IETF WG – RFC 2753, Jan. 2000. [Levine, 2005] Levine, S.S., & Kurzban, R. (2006). Explaining Clustering in Social Networks: Towards an Evolutionary Theory of Cascading Benefits. Managerial and Decision Economics, 27(2-3), 173-187. [Merida-Campos, 2004] Merida-Campos, C. and Willmott, S. (2004) Modelling Coalition Formation over Time for Iterative Coalition Games. In: The Second European Workshop on Multi-Agent Systems, Barcelona, Spain. [Nowak, 2006] Nowak, M. (2006) Five Rules for the Evolution of Cooperation. Science 314, 1560 [Patel, 2005] J. Patel, L. Teacy, M. Luck, N. R. Jennings, S. Chalmers, N. Oren, T. J. Norman, A. Preece, P. M. D. Gray, P. J. Stockreisser, G. Shercliff, J. Shao, W. A. Gray, N. J. Fiddian, and S. Thompson Agent-based virtual organisations for the Grid, in Proceedings of the 1st Int. Workshop on Smart Grid Technologies, Utrecht, Netherlands, July 2005 [Shehory, 2004] Coalition Formation: Towards Feasible Solutions. Fundamenta Informaticae, Vol. 63, No. 2 - 3. (January 2004), pp. 107-124 [Sycara, 1998] Multiagent Systems K. Sycara, AI Magazine, v.10, No. 2, 1998, pp. 79 93 [Thompson, 2003] Thompson, M., A. Essiari, and S. Mudumbai, Certificate-based Authorization Policy in a PKI Environment. ACM Transactions on Information and System Security, 2003. 6(4): p. 566 - 588. [Trustcom, 2005] http://www.eu-trustcom.com/ [Ulieru, 2002] Mihaela Ulieru, Robert Brennan and Scott Walker, “The Holonic Enterprise – A Model for Internet-Enabled Global Supply Chain and Workflow Management”, International Journal of Integrated Manufacturing Systems, No 13/8, 2002, ISSN 0957-6061. [Wasson, 2003] G. Wasson, and M. Humphrey, “Toward Explicit Policy Management for Virtual Organizations”, 4th International IEEE Workshop on Policies for Distributed Systems and Networks (POLICY 03), pp. 173-182 [Wilson, 1975] Wilson, D.S. (1975). A theory of Group Selection. Proc. Nat. Acad. Sci. USA 72: 143-146.