Evolving General Regression Neural Networks for Learning from Noisy Datasets

In this work, a novel algorithm for evolving general regression neural networks is suggested for regression tasks with noisy data. The proposed approach divides the training dataset into slices and then applies the evolving algorithm to create and delete networks to improve the training accuracy. The suggested algorithm performance is verified using a variety of benchmarking datasets where random noise is added to every dataset. Also, different portions of the datasets are utilized for the training, namely, 70% and 50% were used to measure the consistency of the algorithm and its generalization abilities when the training data shrinks. The algorithm results demonstrate its ability to perform well especially when the dataset size is large. The algorithm not only results in reasonable performance in noisy regression tasks but also helps to reduce the network size significantly for many datasets. It is also noted that the training and testing errors are close to each other for many of the tested datasets which ensure the ability of the algorithm not only in the learning stage but also in the generalization stage.