With the emergence of parallel computational infrastructures at low cost, reducing simulation time becomes again an issue of the research community in modeling and simulation. In this context, our previous papers presented a method to reduce the simulation time in a parallel DEVS context. This approach reduces simulation time without reaching the maximum gain. The partitioning method used does not take into account the dynamic of models. To address this problem, we propose in this paper an approach to weight the model graph to take into account this dynamic when partitioning. This paper presents the weighting graph process by learning of the dynamic of models states using Hidden Markov Models. The purpose of this article is to determine the quality of this weighting method compared to a simulation approach.