An experimental study with a Hybrid method for tuning neural network for time series prediction

This paper presents an study of a new hybrid method based on the greedy randomized adaptive search procedure(GRASP) and evolutionary strategies(ES) concepts for tuning the structure and parameters of an artificial neural network (ANN). It consists of an ANN trained and adjusted by this new method, which searches for the minimum number of (and their specific) relevant time lags for a correct time series representation, the parameters configuration and the weights of the ANN until the learning performance in terms of fitness value is good enough, which found, for an optimal or sub-optimal forecasting model. An experimental analysis is presented with the proposed method using three relevant time series, and its results are discussed according to five well-known performance measures, showing the effectiveness and robustness of the proposed method.