Multi-agent simulations are interesting tools to reproduce and explain complex phenomena. These phenomena are often characterized by an emerging behaviour at the population level that is induced by the interaction at the individual level. When such simulations involves a high number of agents and interactions, it becomes intractable to execute it on a single computer. This thesis is focused on two aspects to enable the simulation of large scale multi-agent systems: the first one is concerned with design choices that can be used to distribute the simulator on a computer network, while the second one explores the relaxation of synchronisation constraints in this distributed context. Concerning the distribution challenge, we propose the use of two strategies, one relying on a repartition based on agents properties while the other is based on a division of the environment. These two approaches have been implemented and validated through two applications involving different emergent properties. The second aspect is based on the hypothesis that in such large scale systems, if some entities fails, it should not affect the observed phenomena that is characterized at the macroscopic scale. The type of failure that we have introduced is linked to the relaxation of synchronization constraints by allowing computers to progress independently in a given time window. We have studied this relaxation and its impacts on performances and on the emerging behaviour of two different applications exhibiting different population dynamics. This work has led to the design and implementation of a distributed simulator for large scale multi-agent systems that has been validated on a commodity computer network and on the Grid'5000 infrastructure. The largest simulation that has been tested has involved 50 millions agents running on 50 computers.
Mme. Zahia GUESSOUM, Maître de Conférences HDR, Université de Reims Champagne-Ardenne (Rapporteur). M. Laurent VERCOUTER, Professeur, INSA Rouen (Rapporteur). M. Gregory BONNET, Maître de Conférences, Université de Caen Basse-Normadie (Examinateur). Mme Sophie TISON, Professeur, Université Lille 1 (Examinateur). M. Philippe MATHIEU, Professeur, Université Lille 1 (Directeur). M. Yann SECQ, Maître de Conférences, Université Lille 1 (Co-directeur).
Thèse de l'équipe SMAC soutenue le 3 décembre 2014