Hybrid Coordination Models for Handling Information Exchange among Internet Agents

Information exchange in multi-agent systems raises a multiplicity of problems, which the Internet further emphasises – heterogeneity and dynamics of information among the main ones. Knowledge scientists and engineers have proposed several solutions (mediators, information brokers, ontologies) which, roughly speaking, mainly focus on information representation, retrieval and interpretation. Meanwhile, software engineers and computer scientists have concentrated on the general aspects of inter-agent interaction, by defining models and languages for the coordination of multi-agent systems, which focus instead on the acts of producing, accessing and consuming information, rather than on information per se. In this paper, we first discuss the benefits and limitations of classical coordination approaches to the problem of inter-agent information exchange on the Internet. Then, we survey a new class of coordination models, the hybrid ones, and show how they may help in handling information exchange in the context of Internet-based multi-agent systems. In particular, we take into account Law-governed Linda and TuCSoN. 1 Handling Information Exchange in a MAS Multi-agent systems are rapidly becoming a fundamental paradigm for the engineering of complex software systems in the Internet era [1, 2, 3]. Agents, agent societies and environments are likely to be the basic abstractions around which intelligent systems of tomorrow will be built [4]. In this context, the issue of handling information exchange among agents is a particularly relevant one. Agents are usually supposed to have a partial knowledge of their surrounding environment and limited capabilities to affect its state, to perceive its modifications, and to understand and foresee its possible evolution. So, they typically depend on interaction with both other agents and the environment to fulfil their goals: each agent supplies knowledge to other agents, and relies upon other agents as well as on either reactive (like knowledge bases) or active (like sensors) information sources. Correspondingly, agents in a multiagent system may be seen as both knowledge consumers and knowledge producers, and handled as that. When Internet agents are involved, the heterogeneity, dynamics, and unpredictability of the agent environment make the information exchange issue a quite complex one. Typically, agents have to deal with incomplete information unpredictably coming from heterogeneous sources in different formats. In the AI field, this has led knowledge scientists and engineers to propose several solutions, from information infrastructures (mediators, ontology services) to middle agents (like information brokers). Roughly speaking, all these approaches mainly focus on the representation, retrieval and interpretation of information, with the main goal of enabling agents to effectively exchange knowledge despite differences in syntax, semantics, and ontology. In the Software Engineering and Programming Language fields, computer scientists and engineers have instead concentrated on the general aspects of inter-agent interaction, by defining models and languages for the coordination of multi-component systems [5, 6, 7, 8, 9, 10]. To put it simply, coordination models and languages focus on the acts of producing, accessing and consuming information, rather than on the information exchanged among components. Basically, a coordination model provides for a general framework for the management of agent interactions, which includes (but is not limited to) information exchange issues. So, in the remainder of this paper, we first discuss the benefits of coordination-based approaches to the problem of inter-agent information exchange on the Internet (Section 2). Then, in Section 3 we introduce a new class of coordination models, the hybrid ones, and try to suggest how they may help in handling information exchange in the context of Internet-based multi-agent systems. 2 Data-driven Coordination The classical AI view of coordination [11] typically intermixes the individual (agent) viewpoint with the global perception of a multi-agent system. Speaking very roughly, “coordination” is often seen there first in terms of the efforts of each agent to coordinate its activities with other agents, then in terms of the algorithms which can be adopted so as to ensure that collectivities of coordinating agents actually reach their goals, both as individuals and as groups. In this acceptation, this view of coordination promotes an intraagent approach to coordination in the context of multi-agent systems, and is sometimes referred as subjective coordination [12]. Dually, literature on coordination models and languages that originates from the Parallel and Distributed Systems field [13] promotes a view of coordination which focuses on the models and mechanisms which govern inter-component interactions, rather than on intra-component issues. In the context of multi-agent systems, this view has been denoted as objective coordination [12]: in this acceptation, a coordination model is a conceptual framework for shaping the space of agent interaction, including both agent-to-agent and agent-to-environment interactions [14]. By clearly keeping intra-agent issues separated from inter-agent ones, this view promotes the separation of concerns between computation aspects and coordination ones [15]. Basically, this is what makes it possible to exploit coordination models as the sources for the abstractions required for the engineering of inter-agent aspects (organisations, agent societies) in multi-agent systems [4, 16]. Broadly speaking, the main benefit of coordination models is that they can work in principle as the sources of general-purpose abstractions and mechanisms which can be used in general for handling inter-agent issues, and in particular for managing information exchange problems, like heterogeneity and incompleteness of information, as shown in [17] in the context of the coordination of Internet-based mobile information agents. As a consequence, one may think of choosing a coordination model, and exploiting its coordination abstractions (tuple spaces [13], manifolds [18], . . . ) to face both information-related issues (like supplying an ontology service) and general coordination problems (like synchronisation of agent activities). Coordination models are traditionally divided in two classes: control-driven and data-driven ones [19]. Both classes supply a general framework for managing interactions in multi-agent systems: however, data-driven models straightforwardly address the problem of handling information exchange, since there coordination is basically expressed in terms of producing, accessing and consuming information. Even more, in the case of tuple-based coordination models, like Linda [13] and its many successors and extensions [20], agent synchronisation is based on availability of information, represented in terms of tuples in tuple spaces, and accessed in an associative way, which helps when dealing with incomplete knowledge [17]. This makes data-driven coordination models the most natural candidates for handling both coordination and, in particular, information exchange issues in a multi-agent system. 3 Hybrid Coordination Models However, data-driven coordination models typically lack the flexibility and control required by complex multicomponent applications, which are instead typical of control-driven models. Features like the full observability of communication events, the ability to react selectively to communication events and to implement the coordination rules by manipulating the interaction space are instead required to handle the complexity of interaction in typical Internet-based systems. For instance, in the context of interagent information exchange, it is likely to have the need to intercept a query from an agent, and to perform a sequence of operations before returning an answer – like verifying agent’s identity and permissions, getting the information content from the context, manipulating it, redirecting the query, producing the answer – all transparently to the agent itself. This has lead to the emergence of a new class of coordination models, the hybrid ones, aimed at combining the information-oriented interpretation of coordination typical of data-driven models with the expressive power featured by control-driven models. In a sense, these models look data-driven from the agent viewpoint, but are control-driven from the viewpoint of multi-agent system engineers. TuCSoN Among the main representatives of this class, the TuCSoN model exploits tuple centres for the coordination of Internet agents [21]. A tuple centre is a tuple space enhanced with the notion of behaviour specification, which allows the behaviour of a tuple centre in response to communication events to be defined according to the coordination requirements. More precisely, a tuple centre behaviour specification makes it possible to associate any communication event to a (possibly empty) set of computational activities called reactions. So, each reaction can in principle access and modify the current state of the communication space (by adding or removing tuples) and access all the information related to the triggering communication event (such as the performing agent, the operation required, the tuple involved, etc.), which is thus made completely observable. Since it has exactly the same interface, a tuple centre is perceived by agents as a standard tuple space, that is, as a plain information container. However, a tuple centre behaviour can be specified so as to encapsulate the coordination rules governing agent interaction, by decoupling the agent’s view of the tuple centre (perceived as a standard tuple space) from the actual state of a tuple centre, and relating them so as to embed the coordination laws governing the multi-agent system. So, on the one side, looking like standard tuple spaces to agents, tuple centres basically endorse a data-driven approach to coordination, and promote an information-oriented design which well suits intelligent multi-agent systems [20]. On the other hand, tuple centres also embed a control-driven mechanism like the reaction one, which provide the TuCSoN tuple-based coordination model with the expressive power of control-driven models. In particular, the behaviour of a TuCSoN tuple centre in response to communication events (like agent requests for information) could be defined so as to handle inter-agent information exchange issues. For instance, as shown in [17], the capability of programming the behaviour of the communication abstraction makes it possible to choose a model for the representation of information independently of the peculiar representation adopted by each agent. Then, whenever an agent accesses information in the tuple-based communication space, a tuple centre could react so as to bridge the gap between the global representation chosen and the one of agent involved. Also in [17], further examples are given in the field of information retrieval that show how TuCSoN tuple centres could be exploited to provide for information integrity, and to deal with dynamic and unpredictable information sources. Law-governed Linda Like TuCSoN, Law-governed Linda (henceforth LGL, [22]) exploits a notion of programmable coordination media [23]. As in TuCSoN, LGL coordination media are programmable through a logicbased language: however, while TuCSoN leaves the semantics of the coordination primitives untouched, enabling further effects to be added incrementally through reactions, LGL makes it possible to completely redefine the effect of a communication primitive. However, since the expressive power of the two languages is essentially the same, the main difference between the two models is where coordination rules are stored: while tuple centres incorporate reactions independently of agent distribution over the network, thus working as global coordination abstractions, LGL associates proxies locally to each agent, and makes it possible to program them so as to intercept communication operations and possibly change their semantics. In principle, this makes the tuple centre approach well-suited for expressing global coordination policies, while LGL easily accounts for security and efficiency issues. When managing inter-agent information exchange, this approach would made very easy to define local interaction rules specifically for each of the agents involved. For instance, local LGL prescriptions may be used to take the information emitted by an agent with an out and to translate it in a form that can be understood by other agents, in the same way as in TuCSoN. Or, in case the agent is not allowed to write such information in the shared communication space, the same operation could be prohibited – differently from TuCSoN, where the same operation would have been allowed, then its effects modified so as to make it ineffective. 4 Conclusions and further work In this introductory paper, we have discussed a new class of coordination models, the hybrid ones, and suggested that they may be effectively exploited not only in the general case of the coordination of Internet-based multi-agent systems, but also in the particular case of the management of the problems related to inter-agent information exchange. By combining the benefits of information-oriented coordination with the power of control-driven models, hybrid models can be shown to be expressive enough to help engineers in facing problems like heterogeneity, dynamics, incompleteness, safety, integrity, unpredictability of information. As a result, we argue that suitable infrastructures for the engineering of Internet-based multi-agent systems are to be provided, which supply hybrid coordination services to be exploited in the management of information, too. However, this approach is not meant to prevent more specific abstractions or mechanisms to be used (like ontology services, or mediators). Instead, the idea is that hybrid coordination media should be specialised so as to encapsulate efficient solutions tailored to specific problems, like the ones related to the information exchange issue: for instance, an ontology service may be in principle built upon a TuCSoN tuple centre, but more specialised mechanisms should be added in order to have an effective and efficient solution. An approach of this kind is currently under exploration in TuCSoN regarding security issues in multi-agent systems [24]. Even more, the use of meta-level coordinating agents (or middle agents), also typical in the DAI field, is not really to be taken as an alternative, but instead as a complement to the use of (objective) coordination models. In particular, whenever deliberative capabilities are required in the coordination of a multi-agent system, a middle agent may be the most effective solution. However, this could be easily combined with the use of an effective hybrid coordination models, setting the coordinating agent free from the burden of all the coordination policies which can be charged upon the agent infrastructure. For instance, a translator agent would be surely needed to handle agent communication in different natural languages. At the same time, a coordination infrastructure could be effectively exploited to set the middle agent free from the problems of dealing with all the different formats in which sentences may be represented.