Neural Network Regression Based on Falsity Input

In general, only the truth input is used to train neural network. This paper applies both truth and falsity input, which is the complement of the truth input, to train neural network to solve regression problems. Four neural networks are created. The first two networks are trained using the truth input to predict the truth and falsity outputs based on the truth and falsity targets, respectively. The last two are trained using the falsity input to predict the truth and falsity outputs as well. In order to add more diversity, ensemble of neural networks is applied. Each component in the ensemble contains four types of neural networks created based on our proposed techniques. Aggregation techniques are proposed to provide more accuracy results. Three classical benchmark data sets from the UCI machine learning repository are used in our experiments. These data sets are housing, concrete compressive strength, and computer hardware. It is found that the four proposed networks improve the prediction performance when compared to backpropagation neural network and complementary neural networks. Keywords— Feedforward Backpropagation Neural Network, Ensemble Neural Networks, Complementary Neural Networks, Regression Problems, Truth Neural Network, Falsity Neural Network