P2P Agent Platform: Implementation and Testing - CiteSeerX

P2P Agent Platform: Implementation and Testing Vladimir Gorodetsky, Oleg Karsaev, Vladimir Samoylov, Sergey Serebryakov St. Petersburg Institute for Informatics and Automation, 39, 14-th Liniya, St. Petersburg, 199178, Russia {gor, ok, samovl, sergey_s}@iias.spb.su

Abstract. Peer-to-Peer (P2P) computing, a novel paradigm for distributed information technology, is currently receiving ever increasing interest of both academia and industry. Recent efforts intended to integrate multi-agent and P2P architectures is one of such very promising new opportunities. Indeed, multiagent system (MAS) may be thought of as a multitude of autonomous entities, and, therefore, structuring the agents as a P2P network of nodes may result in architecture providing new dimension for design of open MAS composed of highly transient population of agents. The paper presents implementation of P2P Agent platform providing transparent interaction of distributed P2P agents. The developed P2P platform implements the basic mandatory components assumed by the functional architecture proposed by FIPA Nomadic Agents Working Group. This implementation is supported by search mechanism of underlying P2P infrastructure. The platform verification is done via prototyping P2P ground object recognition MAS in which the agents situated on top of the distributed instances of the P2P Agent platform provide classification services. Keywords: agents, P2P Agent platform, open multi-agent system, services.

1 Introduction and Motivation Peer-to-Peer (P2P) computing, a recently developed network architecture for distributed systems, is currently receiving ever increasing attention of both academia and industry. It was originally proposed "for the sharing of computer resources (content, storage, CPU cycles) by direct exchange, rather than requiring the intermediation or support of a centralized server or authority" [2], while providing the network nodes with identical roles, when any node may act both as client and server [18]1. P2P computing has become a driving force for many new ideas and opportunities in design and implementation of modern large scale applications composed of highly autonomous entities. Recent efforts intended to integrate multiagent and P2P architectures is one of such very promising new opportunities. Indeed, in many cases multi-agent system (MAS) may be thought of as a set of autonomous entities, and, therefore, structuring the agents of MAS as a P2P network 1

Client–server is an attitude about transactions but not about computers or other devices [2].

Vladimir Gorodetsky, Oleg Karsaev, Vladimir Samoylov, Sergey Serebryakov

of nodes may result in architecture providing new dimension for design of open MAS composed of highly transient population of agents. Such architecture should potentially meet basic requirements to many modern, often critical, applications. An example of such applications of greatest concerns is remote sensing and information fusion for monitoring of large scale geospatial regions intended for detection and prediction of natural anomalies (tsunamis, hurricanes, earthquakes, etc.) [20]. Rapidly developing ad-hoc service-oriented networks of mobile devices, large scale embedded systems of various purposes (e.g. Smart Home applications), anti-terrorist systems, etc. exemplify other applications where P2P agent systems may be highly effective. Current research shows that practical realization of the idea of open P2P MAS may become feasible if it is founded on seamless integration of the MAS technology and P2P networking mediated by a middleware (gateway) intended to provide transparency of interaction of agents which know nothing about particular implementation of P2P networking, on the one hand, and, at the same time, know very few or nothing about existence and particular functionalities of most of or even all agents of MAS. In other words, "transparency" of agent interaction means that the agents have not to care how P2P networking is provided, while having the impression of direct interaction with the whole network of agents while requesting from it, but not from particular agents, some services using semantic queries and/or, in its turn, providing the services it possesses in reply to semantic requests from the network. Such agent-to-agent communication can be achieved by making use of adequate FIPA Message Transport Service (MTS) [8]. For P2P ad-hoc networks of agents, FIPA MTS should be realized over P2P transport protocols, for example, a JXTA transport, a Bluetooth OBEX or over some other P2P transport provider. While using an analogy with P2P Content Distribution technologies and systems [2], to provide transparent agent-to-agent interaction, it is necessary to develop an overlay network of software entities set up on top of a P2P provider. The distributed entities of such overlay network have to be thought of as distributed services of the P2P provider enabling discovery of agents and their functionalities on remote devices in serverless environment [10]. Such service is below called P2P Agent Platform. The development of P2P Agent Platform recently has become the task of efforts of Nomadic Agent Working Group (NA WG6) of FIPA established in the end of 2005 [10] which, in its first document [10], states that "No generic P2P Nomadic Agents exists anywhere in the world, and FIPA will make a breakthrough by providing the first complete specifications and fostering implementations. The goal of this WG is to provide a specification for P2P Nomadic Agents as well as a reference implementation capable to run on small nomadic devices." The draft of what is called below P2P agent platform functional architecture was issued in the early 2006. Unfortunately, till now no implementation of such platform has been published. This paper presents the developed by the authors P2P agent platform implementing basic mandatory components of the P2P agent platform functional architecture [9]. This implementation has to be supported by presence and search mechanisms of underlying P2P infrastructure which, in our implementation, is supported by the developed P2P provider. The developed platform is carefully tested based on several case studies, in particular, P2P agent-based intrusion detection described in [11], P2P intelligent sensor network for ElNino Ocean Current Prediction [15] described in [12] and collective ground object detection system

P2P Agent Platform: Implementation and Testing

performed by airborne surveillance system which is composed of several flying objects (e.g. Unmanned Aerial Vehicles, UAV) provided by observation sensors of the same type. The last system is composed of many agents situated on board of different UAVs, at that each agent is trained to detect objects of a particular class if it is observed under particular view angle (either from the front, or from the back, or from the left, or from the right) The agents of the this open system are considered as providers of particular classification services which may be made available to other agents of the system on request. This case study is described in the end of the paper. The rest of the paper is organized as follows. Section 2 briefly outlines the related works. Section 3 sketches NA WG functional architecture [9]. Section 4 presents the details of the implemented P2P agent platform and compares it with one of NA WG. Section 5 outlines for service specification format which is compatible with FIPA service specification format. Section 6 outlines the implemented case study, collective ground object detection system used for testing the develop P2P agent platform as well as for explanation of how the latter operates. Conclusion summarizes the main paper results and outlines perspectives.

2 Related Works The problem of to merging together multi agent and peer-to-peer systems recently became of ever increasing research. In [13], the authors describe multi agent system intended for distributed composition of semantic web services, where agent and service discovery is facilitated in P2P fashion. The proposed MAS is based on structured P2P Chord network implemented as application layer set up on top of P2P network. Paper [22] presents agent-based collaborative framework for mobile P2P applications. The developed system, FRAGme2004, consists of 3 layers: (1) bottom layer providing peer communication, (2) intermediate one comprising agent system managing network resources and (3) application layer representing domain–specific aspects. Agents coupled to peers are considered as resource managing entities. In [21] authors present MAS (agent–community based P2P information retrieval system) where agents communicate in P2P manner. In [4] agents of MAS are organized in hybrid P2P network and every agent can freely connect to or disconnect from the system. Peers are represented as mobile devices with the application agents running on these devices. In [23, 24, 1], the authors designed peers themselves as software agents which, at the same time, integrates application aspects. The common property of the reviewed works is that they do not detach P2P networking layer from peer agent layer and integrate agent system and P2P provider in ad-hoc fashion which is application–dependent and thus non-reusable. The FIPA idea of using yellow pages presenting information needed for Agentsto-Peers coupling is described in [16]. It considers several interacting MAS when each MAS is set up at single peer. In [3] JXTA infrastructure is used as a communication environment. The entity implementing particular service that is published using JXTA is called agent. Thus, this paper proposes distributed multiagent system of services accessible through JXTA infrastructure.

Vladimir Gorodetsky, Oleg Karsaev, Vladimir Samoylov, Sergey Serebryakov

In contrast, the paper [19] takes a step ahead while intending for providing reusability of the software integrating MAS and P2P networking facility via implementation functionalities which are called in [9] agent platform. The proposed idea is to use a special software entity, peer agent, to provide P2P communication of the FIPA compliant application agents using JXTA infrastructure. Thus, although P2P MAS are currently the subjects of many research, they mainly proposed ad-hoc solutions, where P2P agent interaction is designed in application–or problem–oriented style. The advanced solution proposed in [19] is definitely an important step to development of a standard solution. The same basic idea but implemented in more application–abstracted fashion is the subject of the present paper. It strictly follows the intention to completely separate application agent layer and P2P networking layer via using a middleware tending to be standard. This middleware is developed as overlay network of P2P agent platform instances set up top of a P2P provider. P2P agent platform instance is considered as P2P consumer, which may register at some local peer, thus allowing agent to communicate with other agents deployed on other instances of agent platforms registered at remote peers. P2P agent platform itself implements functionalities supporting for semantically rich agent and service search in P2P manner whereas the application agents set up on top of agent platform are responsible only for domainspecific functionality. Let us note, that application agents can also be thought of as overlay network set up on top of agent platform instances.

3 Nomadic Agent and P2P Agent Platform by NA WG Functional architecture of Nomadic agent containing P2P agent platform1 proposed in [9] is depicted in Fig.1. It comprises three layers. The bottom layer corresponding to P2P provider (hereinafter "peer") contains two sub layers, P2P core and P2P services. The former is intended for support for a peer presence mechanism, which has to signal if a network peer is currently connected or not to the network. It also contains the peer contact list managed by peer agents creating/deleting the peer records through corresponding interface if this is assumed by the policy. Other, optional, peer component may provide extra services (in addition to the services of P2P agent platform), e.g., file sharing, installation of direct channel for audio-video streaming (behind agent platform), etc. These services may also be accessible by peer agents. The intermediate layer is exactly what is called P2P agent platform (Agent-to-Peer platform). It may be thought of as a service provided by the peer to its agents. The platform is organized in two sub layers. The upper one, Agent Platform Services contains three mandatory components: Agent and Service discovery, Agent Platform Discovery and Agent–Agent communication mechanisms. According to [9], Agent and Service discovery component is composed of Agent Management System (AMS) and Directory Facilitator (DF). AMS is in charge of White Pages (WP) service (it specifies the list of agents coupled to the platform) and

1

In [6] it is called as "Agent platform" although it is other than Agent platform of FIPA [3]. To avoid ambiguity, hereinafter the term P2P agent platform is used for the NA WG platform [6].

P2P Agent Platform: Implementation and Testing

the agent life cycle maintenance. DF provides Yellow Pages (YP) service to agents of peer. YP service is accessible to agents coupled to other agent platforms of the P2P network. The total collection of YP-services of the P2P agent platforms situated on network peers constitutes what is called distributed YP. Agent Platform Discovery is in charge of finding other P2P Agent platform instances and connecting to them. The third component of the sub layer in Fig. 1. Functional Architecture of P2P Nomadic Agents (http://www.fipa.org/subgroups/P2PNA- question is Agent–Agent WG-docs/P2PNA-Spec-Draft0.12.doc) communication mechanism to be specified in the FIPA ACL [5]. It is intended to provide the transparent connection of agents through Message transport service. The Agent-to-Peer coupling system is in charge of linking agents to the P2P system and vice-versa down to the application agent and peer levels thus providing interoperability. The top layer of the functional architecture (Fig.1) corresponds to FIPA agent services. The next section describes developed implementation of the P2P agent platform and P2P provider and their comparison with the architecture shown in Fig. 1.

4 P2P Agent Platform and P2P Provider Developed and implemented architecture that includes P2P agent platform and P2P provider preserving the basic ideas of NA WG (Fig.1) is shown in Fig. 2. It is structured in the same 3-layer architecture assumed NA WG . At the bottom layer, the peer providing P2P services to consumers1 of these services is situated. In the current version, it provides communication between peers as well as contact list management. At the intermediate layer, the P2P Agent platform providing specific functionalities to agents set up on the platform is situated. At the current stage of development, P2P Agent platform supports agent and service discovery, agent–agent communication and agent–peer coupling. Agent and service discovery performs distributed P2P search of agents’ services and agents’ locators (addresses). Agent– agent communication service provides P2P MAS with message passing. Agent–peer coupling determines management policies between peer and agent platform. At the current version single policy (static) is used. At the top layer, generic agent services provided with basic capabilities of service discovery (matching) and negotiation is situated. 1

The consumers of services provided by peer may be thought of as its clients.

Vladimir Gorodetsky, Oleg Karsaev, Vladimir Samoylov, Sergey Serebryakov

While comparing the functional architecture of Fig. 2 Negotiation with one proposed in [9] (Fig. 1), the following mapping between them can be noticed. Agent and Agent-Agent At P2P provider layer, "Peer to service communication P2P peer communication" Agent component presented in Fig. 2 is a component that partially Agent/peer coupling (static policy) implements the “Communication /Network” functionality assumed by Fig. Peer to Peer P2P Contact communicatio 1. This component, together Provider list with "Agent-Agent Fig. 2. Functional architecture of the developed P2P communication" component agent platform and P2P provider (Fig. 2) supports transparent communication of peer agents with agents of other peers. Let us note that the last component is not explicitly shown in Fig. 1. At the intermediate layer, "Agent and Service Discovery" and "Agent/Peer coupling" components in Fig. 2 correspond to the same components in Fig. 1. At the top layer, "Service discovery, matching" and "Negotiation" clearly match those ones defined by NA WG (Fig. 1). It can be seen that basic mandatory components of NA WG architecture [9] (except “Agent platform discovery” and "Intelligence" components) are implemented in the developed and implemented version of P2P agent platform. Let us describe the operation scenarios of the developed P2P agent platform components, peer and application agents set up on top of the platform. In general, standard software like JXTA, WiFi OBEX or other P2P provider may be used. In the implemented case, P2P provider developed by the authors is used. The Agent platform can be though of as a client or consumer of the P2P provider services. Interaction between P2P provider and consumer comes to interaction through standard interfaces supporting access of P2P provider to consumer and vise versa. In order to use P2P service, consumer has to register at the local peer that is a node of P2P network. To register, Application agents consumer has to specify its own type and identifier, which must be unique Agent platform instance amongst all P2P consumers of the Agent-peer White and Agent-agent network. If particular application does coupling yellow pages communication no longer need P2P provider service it has to deregister. Also peer provides some functionality to manage its own Peer, node of P2P network contact list via adding and deleting Peer-to-peer records, etc. Agent platform can Contact List communication suspend and resume its own presence at peer, for instance, during temporal Fig. 3. Interaction between agents, agent unavailability. If some applications on platform and peer at local host Generic Agent Services

Service discovery, matching

P2P Agent Platform: Implementation and Testing

concrete device need to use P2P transport (e.g., to send message to a P2P Agent out-of-process-peer in-process-peer platform) a P2P providers has to be installed Particular IPC on that device. mechanism P2PCore Interaction between agents, Agent platform and peer at the device is shown in Peer Fig. 3. The peer, a node of the P2P network, Peer is P2P provider of the device. Any application needed to connect to the P2P Fig. 4. Peer functional architecture network as peer client must identify peer running on the device and register. In our case, peer and P2P consumers are "weakly coupled" in the sense that the latter is an independent application, which, in some scenarios, may work without P2P services. In contrast, agents and Agent platform are "tightly coupled", i.e. P2P agents cannot operate without use of the agent platform: the latter fully manages the agent life cycle, i.e. loads, creates, suspends, resumes and destroys it. Let us note that agency and P2P integration can be considered from two points of view. At the bottom layer, Agent platform uses P2P to communicate with other P2P agent platform instances that may be situated at remote devices, i.e. P2P provides standard communication service linking instances of Agent platform into distributed agent platform. On the other hand, P2P agent platform can be considered as driving force P2P MAS. For instance, the search of agents and services happens in P2P manner. This is possible because agents of White and Yellow pages create coalition with their neighbors located on neighbor instances of agent platform. This integration of agency and P2P gives promising new opportunities to P2P multi-agent systems. P2P provider comprises peer factory, peer and P2P core library (Fig. 4). Peer factory library is used by P2P consumers and serves as gating interface to peer. Peer can be created as in-process-peer (application level peer) and out-of-process-peer (system level peer). In-process-peer is not shared among consumers, because it belongs to application created it. Out-of-process peer created at system level is shared among consumers. Peer factory library implements one of IPC mechanism to communicate with this type of peer, thus hiding from consumer all communicating routine details. An application that wishes to get access to P2P through interface provided by peer factory library creates in-process or out-of-process peer, gets its interface, and communicates with the peer via this interface. Agent platform uses P2P search algorithms to discover agents and their services using Yellow and White page services of the Agent platform instances. Conjointly, the latter implement what is called distributed Yellow and White pages services. In the developed P2P Agent platform, White and Yellow pages services are implemented as Yellow and White page agents. They are the mandatory components of any Agent platform instance. White and Yellow page agents create association with appropriate agents at peers that are neighbors of local peer thus allowing for using distributed P2P search methods to discover agents and their services within the whole network. If a particular application agent needs to look for some service it interacts with the local Yellow page agent through interface provided by P2P Agent platform instance. If specific service is found locally then Yellow page agent sends back the Consumer

Peer Factory

Vladimir Gorodetsky, Oleg Karsaev, Vladimir Samoylov, Sergey Serebryakov

result; otherwise it initiates distributed search through other White and Yellow page agents situated on other instances of the Agent platform.

5 Services and Messages Formats In this development, specification format of agent services is based on FIPA documents [7] that is the draft of service specification format issued in 2003. Specification of agent service–associated components comprises descriptions of the services themselves, messages intended for service search ("Service search message") and messages expected in reply on service search message ("Search reply message"). Due to lack of the paper space, the main attention is below paid to the developed service specification format which is represented in Fig. 5. Any service is described by its name, type, set of service-name properties and set of protocols. Service name is the unique name amongst other services registered in distributed yellow pages of the Agent platform. For instance, as a service-type name of service, globally unique identifier (GUID) may be used. The type of the service corresponds to the conceptual service description, referring to its functionality, for … instance, “classifier” or “scheduler”. Service properties present service at more detailed level and are specified by … notions. Protocols define the way the service can be invoked. Any protocol is specified by its name, input and … output notions. Notion is presented by its name and set of attributes. In the current implementation, the notions used in service specifications must be “simple”, i.e. they must not inherit protocol-name other ontology notions and must of integer, real, string or Boolean type. Service search message contains the only field – scene "search query". Search query is specified in terms of agent ontology and represents constrains on the services under … search. These constrains are defined in terms of notions’ attributes. The following types of constraints may be used within query: ‘=’, ’=’, ’!=’. Access to the object attributes is specified with the following string: notion name.attribute-name. … The service name and type are specified by the tags service-name and service-type. For instance, let us show ... the example of search query:

Fig.5. Template for service specifications

Service-type == 'Classification' AND Cost.Cost ‘3’ AND Experience.Assistances >= ‘1’ AND Experience.AvAssistExp > ‘0.6’.

P2P Agent Platform: Implementation and Testing

One can see that search query consists of sub-queries separated by keyword AND. Service invocation request and service invocation reply messages are defined by protocols, which were specified within service description section.

Fig. 6. P2P agent-based service oriented classification network: network topology and distribution of agents over peers

6 P2P Service Oriented MAS for Ground Object Recognition: A Case Study The case study described below is half-artificial. It considers collective ground object detection task in which particular solvers interact on P2P basis. It is assumed that an airborne surveillance system is composed of several flying objects (e.g. Unmanned Aerial Vehicles, UAV) provided by observation equipment, sensors, of the same type. The flying objects are interpreted as the peers of a P2P network intended for on-line detection of ground objects. Information perceived by the sensor of individual flying object is processed by several agents set up on top of the peer (flying object) P2P agent platform instance. Each agent processing sensor information is trained to detect ground objects of definite class if the latter is observed under fixed diapason of view angels, in particular, either from the front, or from the back, or from the right, or from the left. The flying objects observe different ground area that may be overlapping. Therefore, different sensors may perceive the same ground objects but, possibly, under different view angles. The instances of P2P Agent platforms installed on top of each peer together with the P2P services of the peers support transparent communication of the object detection agents. Each agent of peer is interpreted as a provider of classification service differing from other P2P network services in class of objects they are capable to detect and object observation angles. In the implemented case study, 4 peers (called "BaseStation" A, B, C, and D) are introduced. Fig. 6 illustrates the P2P classification network topology and the number

Vladimir Gorodetsky, Oleg Karsaev, Vladimir Samoylov, Sergey Serebryakov

of agents set up on each peer. In the case study, the P2P ground object detection network consists of 22 agents. Among them, 21 are the agent classifiers and one more agent provides visualization thus performing the role of user interface. The P2P agent collaboration scenario is as follows. The agent, having identified, on the scene, an object of its "own" class, sends message to other agents using Yellow and White pages services of P2P agent Fig. 7. User interface of the software prototype platform. This message contains identifier of the detected object class and coordinates of the rectangular on the ground within which the above object is detected [Fig. 7]. In this message, the agent "asks" the network agents whether they detect an object in the specified rectangle. After receiving "in reply" messages, it combines decisions using majority voting algorithm, Fig. 8. Agent ontology and sends the result, i.e. identified object, if any, and its location, to the visualization agent. Let us note that the number of peers in P2P network and their connectivity, the number of agents situated on each peer and the service allocation over the agents were generated automatically using software tool developed by the authors [11]. The agents interact in terms of ontology in which four basic notions are defined. First of them is the object class notion, describing which object class the classifier was trained to detect. The second notion represents the object to be detected. The third one represents the fact weather requested object was detected in specific region. The last notion corresponds to the analyzed scene. Ontology notions and their attributes are shown in Fig. 8. The agents’ services are understood as their capabilities to detect the objects of the fixed classes if they are observed under given diapason of the view angles. Fig. 9 demonstrates the list of agents registered at local white pages. Every agent is described by its name and state. At the bottom part of the figure, other White pages agents situated on remote instances of the agent platform are registered. These instances are the neighbors of the former. The neighbors together Fig. 9. List of agents registered at local white pages create search coalition. The found agents are also cached by White

P2P Agent Platform: Implementation and Testing

Fig. 10. Agent services, registered at local yellow pages

pages agents. In Fig. 9 this section is empty. Fig. 10 demonstrates the list of agent services registered at local Yellow pages. Every service is described by its provider, service name and type. At the bottom one can also see agents that are coalitionists with local Yellow pages agent. Coalitions are used to perform conjointly distributed P2P search of services.

Acknowledgement. This research is funded by the Information Technology and Computer Systems branch of the Russian Academy of Sciences, Project 2.4.

Conclusion The main paper contribution is development, implementation and testing of the P2P agent platform intended for integration of P2P technologies and agent systems. Using P2P services via standard interfaces, applications may have available all functionalities and possibilities of P2P systems such as communication channels between peer devices, distributed search of peers and their services over network and so on. On the other hand, agent–based systems can be thought of as a distributed community of autonomous entities capable to asynchronous computing, interacting with each other, striving to achieve their goals, operating in cooperating and/or competing domain, where single entity is likely to fail. Thus, standardized P2P infrastructure used as communicating environment for instances of agent platform allows for creating distributed systems situated on different kind of devices, i.e. computers, cellular phones, laptops, pocket computers and so on, and having different communication channels between them (tcp/ip, soap, smtp, wifi, bluetooth and so on) while supporting highly transient population of agents. On the other hand, integrating P2P ideas into agent platform allows for creating multi–agent systems that possesses emergent behavior, such as ability to efficiently operating in large scale domains and to supporting highly transient population of agents. Developed software prototype consists of P2P provider (implemented as peer providing P2P communication with other peers ), P2P agent platform (implemented as application that is able to manage agents and realizing distributed Yellow and White pages) and generic agent services implementing generic interaction with P2P agent platform. The simulation of the software prototype of the distributed ground object recognition system proved that ideas of the P2P service oriented multi–agent systems are highly promising for design and implementation of modern open distributed large scale applications.

Vladimir Gorodetsky, Oleg Karsaev, Vladimir Samoylov, Sergey Serebryakov

References 1. Abu-Draz, S., Shakshuki, E.: Peer to peer multi agent system for online trading. LNCS 2671, Springer (2003) 611-613 2. Androutsellis-Theotokis, S. and Spinellis, D.: A Survey of Peer-to-Peer Content Distribution Technologies, ACM Computing Surveys, Vol. 36, No. 4 (2004) 335–371. 3. Bertolini, D., Busetta, P., Nori, M., Perini, A.: Peer-to-peer multi-agent systems technology for knowledge management applications. An agent-oriented analysis. WOA (2002) 1-6. 4. Carrillo-Ramos, A., Gensel, J., Villanova-Oliver, M., Martin, H.: A peer ubiquitous multi agent framework for providing nomadic users with adapted information, LNCS 4118, Springer (2006) 159-172. 5. FIPA Agent Communication Language Specification, http://www.fipa.org/specs/fipa00061/ 6. FIPA Agent Discovery Service Specification, http://www.fipa.org/specs/fipa00095/. 7. FIPA Agent Management Specification, http://www.fipa.org/specs/fipa00023. 8. FIPA Agent Message Transport Service Specification, http://www.fipa.org/specs/fipa00067/. 9. FIPA P2P NA WG6. Functional Architecture Specification Draft 0.12, http://www.fipa.org/ subgroups/ P2PNA-WG-docs/P2PNA-Spec-Draft0.12.doc. 10. FIPA P2P Nomadic Agents Working Group (P2PNA WG6), http://www.fipa.org/subgroups/ P2PNA-WG.html. 11. Gorodetsky, V., Karsaev, O., Samoilov, V., Serebryakov, S.: Agent-based Service-Oriented Intelligent P2P Networks for Distributed Classification. International Conference "Hybrid Information Technologies" (ICHIT-2006). Korea, IEEE Computer Press, (2006) 224–233. 12. Gorodetsky, V., Karsaev, O., Samoilov, V., Serebryakov, S.: Agent-based Service-Oriented P2P Information Fusion. Submitted to the International Conference Fusion 2007, Quebec, Canada, July (2007). 13. Kungas, P., Matskin, M.: Semantic web service composition through a P2P-based multiagent environment. LNCS 4118, Springer (2006) 106-119. 14. Mine, T., Kogo, A., Amamiya, M.: ACP2P: agent-community-based peer-to-peer information retrieval – an evaluation. LNCS 4118, Springer (2006) 145-158. 15. Pacific Marine Environmental Laboratory. http://www.pmel.noaa.gov/tao/ 16. Panti, M., Penserini, L., Spalazzi, L., Tacconi, S.: A multi-agent system based on the P2P model to information integration. Proposal to Agentcities Task Force (2002). 17. Patkos, T., Plexousakis, D.: A semantic marketplace of negotiating agents. LNCS 4118, Springer (2006) 94-105. 18. Peer-to-Peer Research Group, http://www.irtf.org/charter?gtype=rg&group=p2prg. 19. Purvis, M., Nowostawski, M., Cranefield, S., Oliveira, M.: Multi agent interaction technology for peer-to-peer computing environment. LNCS 2872, Springer (2004) 150-161. 20. Tate, A.: The "Helpful Environment": Geographically Dispersed Intelligent Agents That Collaborate, IEEE Intelligent Systems, May/June 2006 (2006), 57-61. 21. UPNP Forum, www.upnp.org. 12. Wang, M., Wolf, H., Purvis, M., Purvis, M.: An agent-based collaborative framework for mobile peer-to-peer applications. LNCS 4118, Springer (2006) 132-144. 23. Yu, B., Singh, M. P.: Incentive mechanism for agent based peer-to-peer systems. LNCS 2872, Springer (2004) 77-88. 24. Zhang, H.Z., Lesser, V.: Multi-agent based peer-to-peer information retrieval systems with concurrent search sessions. ACM (2006) 305-312.