Neural Networks Trained by Randomized Algorithms
Qin Qin, Qingguo Wang, Shuzhi Shuzhi Sam Ge, Chao Yu · Transactions on Machine Learning and Artificial Intelligence · 2014
In this paper, a new model framework is proposed where a group of neural networks are trained with randomized algorithms. By incorporating randomization of random forests into a training algorithm of a neural network, the repeated running of such a revised training algorithm yields multiple independent neural networks. This group of multiple models jointly may outperform individual models. Simulation studies are conducted on various examples including practical ones such as stock markets and show that the proposed model group overall performs better than single neural network and a random forest. When there is significant noise in the data set, the performance of the former drops relatively less than the latter. In particular, the former produces much lower deviation of the performance and higher mean performance compared with the latter. Therefore, the proposed method has strong ability to classify the noisy data and perform robustly.