Résumé
This habilitation thesis synthesizes research works which are mainly related to the field of Multi-Agent Based Simulation (MABS). MABS is a general framework for modeling and experimenting with systems in which the dynamics emerges from local interactions among individuals (autonomous agents). Examples of use range from the study of natural systems (e.g. ant colonies, crowds or traffic jams) to the engineering of artificial ones (e.g., collective robotics, distributed artificial intelligence-based softwares). To this end, MABS modeling represents the behavior of individuals, their environment and interactions, so that global dynamics can be computed and studied from the bottom up. In this context, we have been investigating research on the theory and practice of MABS from two different perspectives : (1) the design of generic abstractions dedicated to the modeling of multi-agent dynamics (e.g., the IRM4S model) and (2) the engineering of MABS (MaDKit and TurtleKit platforms). Besides, we have been experimenting with MABS in different application domains such as image processing, video games, and collective robotics. Contrary to approaches that put the emphasis on the agent behaviors, all these works have been done by considering the environment of the agents as a first order abstraction. In this thesis, we first reflect upon the research we have conducted according to this perspective. Next, we show how we actually use this perspective to propose an original approach for using General-Purpose processing on Graphics Processing Units (GPGPU) within MABS, and then present the research perspectives related to our positioning.