Nowadays, Recurrent Artificial Neural Networks (RANN) are the most appropriate tool to face pattern recognition or forecast problems in complex domains or with a temporal component. However, the use of RANN has some problems, due to their slow training and to the fact that convergence is difficult to reach. The utilization of Genetic Algorithms (GA) in the development of ANN is a very active area of investigation. The works that are being carried out at present tend, more and more, to the development of systems which realize tasks of design, optimization and training, in parallel. In this paper we propose a distribute GA architecture which establishes a difference between the design, the optimization of the training set and the training process. In this system, the design tasks and the optimization of the training set are performed in a parallel way, by using a net of computers. Each design process has associated a training process as an evaluation function. Every design GA interchanges solutions in such a way that they help one each other towards the best solution working in a cooperative way during the simulation.