A learning strategy for developing neural networks using repetitive observations

Abstract

Neural networks can model system behaviors by learning past system observations. As system observations are usually collected by human judgments, physical experiments or sensor measures, they can be inherently imprecise and inconsistent over time. System behaviors can be learned more completely from repetitive observations. However, repetitive observations can be very different due to system or measurement uncertainty. If abnormal observations are used for developing neural networks, spurious behaviors can be learnt and the neural networks are likely to generate spurious prediction. If abnormal observations are excluded, important system behaviors can partially be ignored. In this paper, a novel strategy is proposed to develop neural networks by learning repetitive observations. Numerous neural networks are developed individually based on either abnormal or normal observations. The predictions generated based on the individual neural networks are integrated to a single prediction. Analytical proof indicates that the overall observation uncertainty involved on the proposed learning strategy is less than the uncertainty involved on the general ones. As less uncertainty is involved, more effective learning can be performed on the proposed strategy. Two case studies are conducted in order to evaluate the effectiveness of the proposed learning strategy, where the two case studies are involved data collection from either sensor measures or human evaluations. Numerical results indicate that the proposed strategy can generate better neural networks which have higher fitting capability to captured observations and higher generalization capability to uncaptured samples.