Optimizing member selection for Neural Network ensembles using Genetic Algorithms

Artificial Neural Network (ANN) is a widely used technique in forecasting applications. An ensemble of ANNs can produce more accurate forecasts than a single ANN. The performance of the ensemble depends on its' member ANN. Member selection for an ensemble is a complicated task that need balancing conflicting conditions. This paper presents a method to optimize the selection of members for an ANN ensemble using Genetic Algorithms approach. To develop the models daily weather data are used. Rainfall data for Colombo, Sri Lanka are used to develop and test the models and rainfall data for Katugastota, Sri Lanka are used to validate the models. The results obtained are compared with two widely used member selection methods Bagging and Boosting. The ensemble model (ENN-GA) performed better than Bagging and Boosting methods and managed to produce forecasts with RMSE 7.30 for Colombo and RMSE 6.21 for Katugastota.