This thesis takes place in the field of Multi-Agent Systems (MAS) and Distributed Artificial Intelligence (DAI). We particularly focus on the issue of the organization in "open systems". Thus our research aims at defining social behavior models, making the agents organize and adapt themselves to their environment. Our approach implies (among other methodological principles) the use of a "virtuous circle" that proposes to use metaphors from other scientific fields to design models in computer science. In the first part, we seek metaphors of behaviors that are close to our needs in animal societies (especially primate ones). We define agent models that are able to produce, in simulation, one of the social features observed in primates, i.e. the ability to recognize dominance relations. We then transpose those models into a different application, the collective building of a lexicon, in order to get indications on the collective dynamics involved in the models. In the second part, we present an experiment of "Open Collective Robotics" in which a group of robots has to adapt to an human-inhabited environment (MICRobES project). We show that a simple transposition is not possible any more under these conditions, and that the embodiment of the robots has to be taken into account. Thus, we propose new design principles for agent behaviors involving natural selection ("Ethogenetics") and present the results we have obtained with a framework implementing these concepts (ATNoSFERES). We therefore generalize the initial principles by conciliating the specific features of both a multi-agent approach and evolutionary algorithms through ethological concepts.
