Nowadays, recurrent ANN are the most appropriate tool to facing pattern recognition or forecast problems in complex cdomains or with a temporal component. The use of feedfordward ANN in these cases means to force this type of network into a task for wich it has not been designed. However, the use of ANN poses some problems, due to their slow training and to the fact that convergence is difficult to reach. For these reasons, the use of ANN is not very common. It has been proposed to substitute training algorithms based on gradient descent for others such as GA in order to solve this problem.
We have implemented a system of this type wich trains RANN by using GA. This training is carried out by adjusting the weights among the connections lf the ANN´s neurons and by selecting the activation functions for each neuron in the network. The system has been proved with success at predicting at short term real-time series.
This paper shows the research carried out to find the adequate values for GA function parameters, so that RANN training is done in a faster and more accurate way. We also comment on the experience of working with individuals formed by more than one data array, after codifying the weights and network activation functions independently.