Time Series Analysis Using GA Optimized Neural Networks

Time series has been one of the most important data used in system analysis. However, the underlying relationship usually conceals itself deeply in large data sets and is difficult to identify using conventional tools. To deal with the inherent complexities of real world systems, this paper presents a hybrid evolutionary- neural modeling approach to model the time series and extrapolate them to the future to make prediction. In this method, a back-propagation neural network is trained to mapping the underlying relationship. To improve the training efficiency, a genetic algorithm is employed to optimize the input series of the model as well as the network topology; the genetic algorithm is also used to search the global optimal initial weights for the local gradient-descent training algorithm. The genetic-algorithm optimized neural learning algorithm is applied to a landslide dynamic system and the results show the great performance of the proposed hybrid approach, both in learning and generalization.