Combination Neural Networks
Steven Whites, Six Sigma Black Belt, Kennecott Utah Copper, Rio Tinto
JMP analysis has shown how different ores affect our metal concentration process. This analysis may shed light on other business problems, such as economic demand curves where theory and observed results appear to contradict each other. The metal concentration process has an underlying relationship similar to the economic demand curve, where there is a negative relationship between price and quantity. However, in the case of metal concentration processes, a scatterplot of daily results often shows an unexpected, positive relationship due to a shift in ore characteristics. By analogy, a shift in the demand curve due to changing customer preferences may show a similarly unexpected, positive price vs. quantity relationship. This presentation will describe how theoretical curves can be found within these seemingly contradictory actual results. Analysis begins with a complex data set of several hundred potential metallurgical and operational terms. The Partition platform in JMP is used to reduce this to a more manageable list of candidates. The Multivariate platform and principal components analysis reduces this list further. Straightforward JSL scripts are then used to test combinations of the remaining terms to find an optimal neural model. A bubble chart is used to display analysis results, where X = the median value of r-sq (test), Y = the standard deviation of r-sq (test), and bubble color = number of terms for that particular model. Multiple runs (limited by computer time) are made with different random training/validation/test subsets, making use of the repeated holdout method for error rate estimation. Optimum model selection is then simplified with a visual display of number of terms, model effectiveness and model stability.