Evolutionary Design Of Neural Networks: Application To River Flow Prediction

Time series prediction has been successfully usedd too support decision-making in several real worldd applicationn areas. Different techniques can be deployedd for predicting time series, such as Box-Jenkins models andd Artificial Neural Networks (ANNs). The former approach, althoughh widespread, is only capable of constructingg linear models. In contrast, ANNs are able to model non-linear functi ons. However the choice of an appropriate ANN's architecture is not always trivial. In this paper, we use Genetic Algorithms (GAs) to optimize the choice of ANNs' architecture for river flow prediction problems. Too improve the GA's performance, we propose initializing i ts firstpopulationwithwell-succeededdarchitectures previously usedd in similar problems. For that, we are developingg a case base in which each case associates a particular river flow prediction problem to an architect ure usedd to predict that river's flow. After the execution of the GA, the resultingg ANN architecture is ready for use, and cann also be associatedd to the input problem to produce a new case. We expect that the selection of architectures from the case base will become better as more cases are inserted. The initial prototype was testedd against a ra ndom search algorithm andd the results obtainedd were very encouraging.