Zooming Multi-Agent Systems

Complex systems call for a hierarchical description. Analogously, the engineering of non-trivial MASs (multiagent systems) requires principles and mechanisms for a multi-layered description, which could be used by MAS designers to provide different level of abstractions over MASs. In this paper, we first advocate the need for zooming mechanisms, promoting a coherent and consistent multi-layered view of agent systems. After surveying the best-known AOSE methodologies, we focus on the scaling mechanisms of the OPM process-oriented methodology. Then, by adopting SODA as our reference, we show how an AOSE methodology can be enhanced with simple yet expressive zooming mechanisms. Finally, we present a simple case study where the enhanced agent-oriented methodology (SODA+zoom) is exploited and put to the test. 1 Zooming as a Principle in the Design of MASs As advocated in [1], MASs (multiagent systems), once developed up to their full potential, can be generally seen as representing a class of complex artificial systems, wide and meaningful enough to legitimate, in principle, the application to MASs of the general principles and laws governing complex systems. While modelling complex systems and understanding their behaviour and dynamics is the most relevant concern in many areas, such as economics, biology, or social sciences, the complexity of construction is of paramount interest when dealing with software systems—MASs in particular. Drawing results from heterogeneous scientific areas, and bringing them to the MAS field, is then particularly meaningful and promising when principles and ideas that are known to model and describe complex systems in general are taken and shown to be applicable and useful to build MASs—in other terms, become ideas and principles for agentoriented engineering processes and methodologies. 1.1 Hierarchies in Complex Systems According of the theory of hierarchies [2], all biological systems are amenable to be represented as organised on different layers, ranging from genes and cells up to organisms, species and clades. Each level is essential to the general understanding of the system’s wholeness, and is autonomous with its own laws, patterns and behaviour. At the same time, no level can be understood in isolation independently of all the other levels, and the system as a whole can be understood only through the understanding and representation of all its levels. When generally ascribed to complex system, this sort of “hierarchy principle” might also be seen as a defining one: that is, a complex system is a system requiring layers—independent but strongly correlated ones—in order to fully understand and reproduce its dynamics and behaviour. When brought to MASs, in particular, this first suggests that MAS models, abstractions, patterns and technologies can be suitably categorised and compared using a layered description, as shown in [1]. More simply and directly, when applied to the engineering of MASs, the hierarchy principle suggests that agent-oriented processes and methods should support some forms of MAS layering, allowing engineers to design and develop MAS along different levels of abstractions—a number of independent, but strictly related, MAS layers. Accordingly, one should expect that existing methodologies actually do support abstractions and processes for MAS layering. Quite interestingly, however, current AOSE methodologies offer very little (if any) support for hierarchical representation of MASs. So, in the following subsection we first survey the main AOSE methodologies to look for some support for layered representation of MAS, then we advocate the need of a simple layering mechanism (called here zooming) to be applied to any meaningful agent abstraction at any stage of the MAS engineering process. 1.2 Zooming in AOSE Methodologies Many methodologies exist in the literature aimed at the engineering of artificial systems in terms of MASs. Some example are GAIA [3], MaSE [4], Tropos [5], MESSAGE [6], Prometheus [7]. Although none of those methodologies provides MAS engineers with an explicit layering mechanism, some of them exhibit some implicit mechanisms that make it possible in some sense to analyse the system at different levels of detail. At the best of our knowledge, the most cited AOSE methodology, GAIA, does not introduce any mechanism providing for MAS layering. In MaSE, instead, two models allow MASs to be represented at different levels of abstraction: the creating-agent-classes model should provide a high-level vision of the MAS agents and of their main conversations; instead, the assembling-agent-classes model “zooms” on the inner agent structure, and provides for a number of predefined components, which may also have sub-architectures (with further subcomponents) of their own. Tropos promotes a form of refinement across different stages of the MAS analysis process, such as when the actor and dependency models built in the early requirements phase are extended during the late requirements phase by adding the system-to-be as another actor, along with its inter-dependencies with social actors. Also MESSAGE use a refinement model in the analysis phase: the level 0 model gives an overall view of the system, its environment, and its global functionality; next level 1 defines the structure and the behaviour of entities such as organisation, agents, tasks, goals, domain entities; further levels (2, 3, . . . ) might be defined for pointing out specific aspects of the system dealing with functional requirements, as well as non-functional requirements such as performance, distribution, fault tolerance, security. In Prometheus, a progressive refinement process is used which starts by describing agents internals in terms of capabilities. The internal structure of each capability is then given, optionally using or introducing further capabilities, which are refined in turn until all capabilities have been defined: capabilities nesting is allowed, thus allowing for arbitrarily many layers, in order to achieve an understandable complexity at each level. The above forms of layering, however, are quite limited. First of all, they enforce only a top-down, mono-directional form of zooming—so, refinement allowed, abstraction not allowed. Then, they have ony a pre-fixed scope and structure, which limit in principle their flexibility and possibly their ability to fit the many different MAS application scenarios. Mechanisms for zooming are then not explicit, and no ontological support is currently provided by any of the available AOSE methodologies to the best of our knowledge.