In the planning phase of cyclically interlinked production lines neural networks are used to learn and forecast the characteristics of these systems. To increase the results of learning and to accelerate the training this paper presents a method, based on genetic algorithms, that reduces the attributes to describe the behaviour of the production lines.