INFORMATION FUSION VIA A HIERARCHICAL NEURAL NETWORK MODEL

摘要

This paper considers information fusion at the intermediate and raw data levels in order to improve the quality of decision making using currently available information. Information fusion becomes critical when records have been collected over certain periods of time at multiple locations with multiple systems. This is because they are often temporally and spatially correlated, and each system has its own bias and variance. To address these problems, a hierarchical neural network model is presented. The proposed model can learn both spatial and temporal dependence from data, and accommodate different information from multiple data sets. At the lower level in the proposed system, neural networks called local experts are specialized to learn spatial, temporal, or combined signals. At the upper level, another neural network called global expert is built. The global expert takes as inputs the estimates of local experts and searches the space of hypotheses to capture possible non-linear relationships among local experts. The proposed hierarchical model is calibrated on snow water equivalent (SWE) data that has been collected over last 89 years at multiple locations in the western United States. Experimental results confirm that the hierarchical model with information fusion at the intermediate and raw data levels shows superior performance to models without fusioned information. Further, the presented model is the best in terms of predictive accuracy, while it is as reliable and robust.