This paper is about the understanding and control of complex systems modeled by a multi agent system. We propose an approach to experimentally approximate the behavior of the system as a function of its parameters. The originality of our proposition is the ability to dynamically generate data to train the model and therefore reducing the number of executions needed to build the approximation while maximizing its accuracy. We detail our proposition and provide an example of how to put it into practice. Some experiments are presented and the obtained results suggest its viability.