Designing Multi-Agent Systems around an Extensible Communication Abstraction

What is relevant for the effectiveness of a multi-agent system is the model of the interaction between agents, rather than their peculiar internal model. The design of a single agent architecture should then concentrate on agent observable behaviour and on its interface towards the outside. A multi-agent architecture should then be designed around the choice of a suitable coordination model, accounting for all the aspects of agent interaction. Accordingly, the effective design of a multi-agent architecture should focus on the role and properties of the coordination media (the communication abstractions) within the coordination model, instead of the coordinable entities (the agents).

The main aim of this paper is to show how a multi-agent system may benefit by a coordination model whose flexibility and expressive power lays in the extensibility of the behaviour of the coordination medium. Extensibility can result from the embodiment of computational properties typically in charge of the agents into the communication abstraction.

As an example, we show how a shared communication device à la Linda works as the core of a flexible coordination architecture in the ACLT coordination model. ACLT tuple centres are logic tuple spaces, enhanced to be reactive to communication events. Tuple centres are programmable through reactions by means of ReSpecT, a specification language based on logic tuples. Moreover, reactions have the semantics of asynchronous, mutually-independent atomic transactions. By defining different observable behaviours for ACLT tuple centres through reaction programming, a multi-agent architecture can then exploit a number of different agent coordination policies without affecting the single agent behaviour.