Optimizing Neural Networks for Time Series Prediction

In this paper we investigate the effective design of an appropriate neural network model for time series prediction based on an evolutionary approach. In particular, the Breeder Genetic Algorithms are considered to face contemporaneously the optimization of (i) the design of a neural network architecture and (ii) the choice of the best learning method. The effectiveness of the approach proposed is evaluated on a standard benchmark for prediction models, the Mackey--Glass series. 1. Introduction The main motivation for time series research is to provide a prediction when a mathematical model of a phenomenon is either unknown or incomplete. A time series consists of measurements or observations of the previous outcomes of a phenomenon that are made sequentially over time. If these consecutive observations are dependent on each other then it is possible to attempt a prediction. Clearly it is supposed that the process is somehow predictable. The time series prediction problems are usually..