Activity Theory as a Framework for MAS Coordination

Approaches to the coordination of multiagent systems (MAS) have been recently classified as subjective – typically coming from the distributed artificial intelligence (DAI) –, and objective – coming from the community of Coordination Models and Languages. Subjective and objective approaches have a very different impact on the engineering of social aspects of MAS, in particular with respect to the ability of specifying and enacting social laws to achieve global coherent behaviours. In this work, we provide a conceptual framework – influenced by the research on Activity Theory – where both subjective and objective coordination play an essential role, each providing effective means for the same coordination/cooperative problems at different abstraction and operational levels: co-construction/co-operation level for subjective coordination, and coordination level for objective coordination. In particular, the work shows the benefits of supporting dynamic transitions between such levels, alternating co-operation stages – in which agents reason about coordination and cooperatively forge coordination artifacts (laws, constraints, norms) – and co-ordination stages – where the artifacts, embodied in proper coordination media, are exploited, so as to enact automated, consistent and prescriptive coordination. 1 Objective vs Subjective Coordination in MAS Interaction and coordination are interdisciplinary issues, and therefore it is not surprising that their study and development in multiagent systems (MAS) is supported by approaches coming from heterogeneous contexts (refer to [20, 13, 36, 27, 28] for comprehensive surveys). The major contribution to coordination in MAS comes from the distributed artificial intelligence (DAI) field. Generally, DAI approaches explicitly deal with subjective coordination [32], considering coordination as “the process by which an agent reasons about its local action and the (anticipated) actions of others to try and ensure the community acts in a coherent manner” [13]. Consequently, the coordination of the overall system is determined by both the mental attitudes (beliefs, goals) and the role of individuals, and by the subjective perception of the inter-dependencies among the members of a society [24]. Well-known examples of subjective coordination techniques in MAS are multi-agent planning [10], where agents build plans and commit to behave in accordance with them, and negotiation [4, 20], where agents communicate so as to reach a mutually accepted agreement on some matter. Typically, subjective coordination is exploited in MAS composed by intelligent (such as BDI) agents, provided with an high-level agent communication language (ACL), such as KQML or FIPA, whose formal semantics can become the means to realise flexible coordination [3]. Other substantial contributions to MAS coordination are rooted in concurrent (parallel and distributed) systems. Starting from the need to explicitly separate coordination from computation issues in system design and development [12], several coordination models and languages have been developed [29], and applied to MAS [28]. Generally, these approaches explicitly deal with objective coordination [32], promoting the separation between the individual perception of coordination and the global coordination issues, enabling the modelling and shaping of the interaction space independently of the interacting entities [24]. This kind of coordination is called objective because it prescinds from the subjective view of the coordinated agents. It is also called uncoupled [36], since coordination is no longer coupled with the computational issues of the coordinated entities. According to [32], objective coordination is mainly concerned with inter-agent issues such as (i) the description of how the MAS environment is organised, and (ii) the management of interactions between both agents and their environment, and agent themselves. Objective coordination models are heavily based of the concept of mediated interaction: agents are provided with specific abstractions that enable their actions (typically communications), and mediate the emerging interactions, caused by the dependencies inside the agent ensemble. In the classification generally adopted by the coordination community [6], this abstraction role is assumed by the coordination medium, which rules agent interactions by applying coordination laws, which represent social laws and system constraints. Tuple centres [23] and the related TuCSoN coordination infrastructure [25] are an example of a coordination approach promoting objective coordination. The variety and complexity of the interaction scenarios that characterise multiagent systems and agent societies require approaches with the properties of both subjective and objective coordination. For instance, distributed workflow management systems (WfMS) and office automation environments require the automatism and prescriptiveness that are easily provided by objective approaches; instead, unstructured environments (from the coordination point of view) and market-oriented competitive scenarios mostly rely on the reasoning and interaction capabilities of individual agents. Moreover, relevant coordination scenarios, such as business process coordination and computer supported cooperative work (CSCW), require models providing the means for balancing subjectivity and objectivity, mediating between application-centric and humancentric tasks, unstructured process and highly structure process structure [8]. In order to face such a complexity, we consider useful from both a scientific and an engineering point of view to provide a conceptual framework exploiting both the subjective and the objective coordination approaches: this is important to ground models and infrastructures aiming at supporting both approaches in MAS coordination. The interdisciplinary nature of interaction and coordination makes it possible to find relevant contributions in disciplines and theories outside computer science: social theories about organisation and coordination in human societies can be very effective, given the complexity of the interaction space involved in some MAS scenarios. An example is the theory of coordination developed by the Centre of Coordination Study (CCS) at MIT: from the analysis of heterogeneous contexts, from economy to computer science, coordination emerges as the process of managing dependencies among activities [17]. Accordingly, engineering coordination inside systems means first identifying the dependencies among the individual parts, then defining strategies to manage the identified dependencies (or interactions, when dependencies are among agent actions), and finally enacting the strategies. The concept of dependency is a very important element for both subjective and objective approaches: in the first case of subjective approaches, dependency appears as the basic relationship on which the construction of the whole agent social life is founded [5, 7], while objective coordination approaches are explicitly focused on the management of dependencies/interactions. The coordination theory approach is clearly bottom-up (regulatory): dependencies are the starting point, then coordination activities are designed and developed on top of dependency analysis. However, the social issues that we aiming at engineering in MAS also account for a top-down (constructive) approach to coordination: the starting point here are the social goals and the properties that must be provided by the aggregation of agents as a system/society, then dependencies among the individual parts are determined/induced accordingly. While the theory of coordination provides a good conceptual background for the bottom-up approach, we found Activity Theory [37] very effective to frame the role of objective and subjective coordination (and their relationship) in the top-down case. Initially developed to study dynamics in collective human work activity, Activity Theory has been recently introduced in computer science disciplines, such as CSCW and HCI (Human Computer Interaction). Both activity and dependency theories refer (either directly or indirectly) to the notion of mediated interaction: interaction media are studied, catalogued and used to manage dependencies in the theory of coordination, and media are forged as coordination artifacts in Activity Theory. The need to define abstractions and “social structures” mediating agent interaction is clearly visible also in the evolution path of some subjective coordination approaches in MAS: the notion of social agency [34] and social laws [33], the STEAM teamwork module in the TeamCore coordination architecture [35], and stigmergy coordination [30] are notable examples. In the last case, in particular, agent interactions are mediated and ruled by the environment, with its own laws and constraints, and coordination appears as an emergent phenomenon. Given these premises, in this seminal work we provide a conceptual framework derived from Activity Theory where both subjective and objective approaches are exploited in the same coordination context, but at different conceptual and operational levels; the framework gives insights into the dynamics between the approaches/levels, showing the fundamental role of mediated interaction (exemplified by the coordination medium/artifact abstractions) for that purpose. Three hierarchical levels for analysing every social activity in MAS are identified: co-construction, co-operation and co-ordination. Accordingly, we show how subjective and objective approaches are both fundamental to support such levels: in particular subjective approaches for co-construction and co-operation, and objective ones for co-ordination. In this way, both high level cooperation protocols among intelligent agents – typically found in subjective approaches – and scripted coordination driven by laws embedded in coordination media – as typically found in objective coordination – turn out to be necessary to build MAS with autonomous agents behaving efficiently and coherently in a social/systemic context. Moreover, the dynamics between co-operation and co-ordination is discussed, providing the notion of coordination reflection and reification. The remaining part of the work is organised as follows: Section 2 provides the basic elements of Activity Theory useful for this work. Section 3 comes back to coordination in MAS, and provides a conceptual framework – derived from previous section – in which co-ordination/co-operation activities are analysed, and the relationship with subjective and objective coordination discussed. Finally, Section 4 provides conclusions and future works.