Définition d'un cadre méthodologique pour la conception de modèles multi-agents adaptée à la gestion des ressources renouvelables

Abstract

The potentialities of the MAS should not hide the difficulties the modelisator can encounter. More precisely, the question of models validation is regularly mentioned. On the contrary of mathematical models, it is impossible to prove the characteristics of a MAS. Moreover, the probabilities of revealing errors or bugs are not negligible. One can then legitimately wonder about the reliability of the simulators. However, the main problems dont come from errors of coding but rather from the management of the interactions or activations of the agents. Often underestimated, a rough management may produce artifacts and one can give by mistake the wrong properties to a model. Some techniques limit the emergence of this kind of bias by reinforcing for example the autonomy of the agents. Indeed, by protecting their data integrity, these architectures prevent the direct interactions between agents. Thus, they necessitate a treatment deviated towards a higher nature entity (environment or universe) in charge of solving interactions, conflicts and parallel treatments. Nevertheless, these techniques are constraining and difficult to implement. Single- handedly, they bring hardly satisfactory answers to prevent from bias appearance. It is rather important to be aware of the hot areas that are likely to generate artifacts and to use standard procedures by adapting them to the modeled field. This does not mean that it is necessary to sophisticate the interactions between the agents or to impose a particular system of time management. On the contrary, it is necessary to handle these vulnerable points by proposing simple and controllable models and by considering that activation and interaction are at the heart of a model. The modeling of the socio-ecosystems is not just a data-processing speciality, but request a confirmed know-how and the modelisator must take a critical look at its own tools. It is necessary to know by advance the vulnerable points of the MAS but it is also necessary to improve the robustness of its results by showing that the simulator exhibits relatively stable behaviors. The independent replication of a simulator reinforces its reliability. Moreover, to check and reproduce experimentation (whatever type) appear as the rule of any rigorous scientific method. But the difficulties of replication are affected by miss readability of the MAS. It is thus essential to describe its model in a clear and non-ambiguous way. After the implementation of a simulator, a work of re-presentation must be carried out to design reorganized diagrams that put the light on the essential points of the model and authorize discussions and criticisms. We defend the idea of an exposed modeling that must assume its choices without imposing its points of view. It is urgent to give up the naive vision which consists in thinking that a model is objective. The idea of neutral model as a simple copy of reality in silico would imply that the modelisator has no presupposed on the studied system and that the world is reflected in its thought like an image in a mirror. However, a model is inevitably a subjective representation. It is thus necessary to clarify its choices and to present them in the most readable way so that they can be understood, shared or criticized. Due to the lack of reliability of the MAS and also to the naive approaches of modeling, it is preferable, in the actual position of the MAS, to choose simple models rather than a descriptive approach which would seek to introduce into the model the maximum of information considered as given by nature.

Cite this document ^(BETA)