Data Segmentation and Genetic Algorithms for Sparse Data Division in Nome Placer Gold Grade Estimation Using Neural Network and Geostatistics

Abstract - Ore reserve estimation, based on sparse drill hole data, was conducted for a placer gold property in Nome, Alaska. A problem with sparse data is that random subdivision of the data into modelling and evaluation subsets (as is commonly done) becomes a problem, as random selection may result in biased/skewed subsets. Therefore, a technique that combined data segmentation with genetic algorithms (GA) was applied to divide the samples into three equivalent subsets: training, validation and testing. Data segmentation was done on the basis of the distribution of gold values. Neural network and a variety of kriging techniques were used to estimate gold grades. A multi-layer feed forward neural network along with "early/quick stop" training was used for neural network modelling. A comparative evaluation of kriging and neural network methods was then performed. The results revealed that neural network was generally superior to the kriging techniques for gold grade estimation in the Nome deposit.