Innovative approach

Modelization: a key to comprehension

By definition, the behavior of a complex system can not be intuitively inferred. But when confronted to a complex system or process, engineers will nonetheless develop some insightful expertise. Actually, they will create their own representation, or model, of the system they are confronted to. Even if this model is not clearly defined, it is almost always possible to explicit the inner mechanisms and their feedbacks. For instance, a biologist is unable to describe how ants can build geometrically complex nests, but he will enunciate the basic behavior rules of individual ants. A physicist cannot guess the evolution of climate over several decades and deduce the production of a solar plant over the same period of time, but he will have a good knowledge about every single phenomenon and the various feedbacks among them. Theoretical physics has been trying to manage complex systems for more than thirty years, in various application domains. The main idea has been to make use of universality classes: a given complex system must belong to one of the classes to which a solution is known. By not taking into account the specificities of the system, this approach is both a success and a failure. A success because the model is then easier to manipulate. But also a failure because all the valuable expertise is neglected. Specific features that make the initial model unique are not preserved. The first step of the methodological chain for complexity engineering that has been developped by Meso-Star is to enunciate and formalize the model from the knowledge that can be found on the field. This expertise is fully preserved throughout all subsequent steps.

Sampling as a tool to explore complexity levels

The next question is: "How do we take into account the full complexity of the model?" Grasping this complexity means that the full phenomenology of the system has to be preserved. Which raises the following contradiction: a comprehensive approach is clearly not tractable, while the most significant aspects may be missed when using a simplified model. Our approach is based on a statistical exploration of the model defined by the experts: rather than simplifying the model in order to compute the trajectory of every possible variable over the whole definition domain, we use statistical algorithms together with relevant sampling schemes in order to focus on a few number of significant variables. This philosophy induced a deep methodological redesign in the field of computer graphics simulation during the last few years. This greatly improved the quality of the production while also granting a lot more freedom to artists and designers. Beyond mere computational considerations, using a statistical formulation provides new ways of thinking the system as a whole, while preserving every relevant phenomenon and feedback.