Résumé
This work investigates the use of organizational abstractions in multi-agent systems. The thesis shows how adopting a social-based point of view on these systems, instead of a classic individual-based assumption, may be an answer to increasing needs in heterogeneity, adaptation and control in large-scale agent systems. The main proposition is a generic model of agent organizations, based on the concepts of agent, group and role. We show how a structural and functional decomposition can federate multi-agent system. We focus on agents with highly heterogeneous internal architectures, possibly antagonist goals and with different interaction models. The model is then formally expressed in a pi-calculus variant. We illustrate its operational semantics through an overview of meta-level design. Implementation aspects are also discussed through the presentation of MadKit, a generic agent infrastructure toolkit, based on our organization model and designed to host multiple architectures. Empirical results validate our contribution, through architecture-specific evaluations and real-world applications.