Cross Validation - XVALID

The cross-validation process XVALID is designed to assist in the selection of parameters for grade estimation, using the cross-validation method. The input data file is the sample data which will later be used for estimating grades into a block model. For kriging it allows different model variograms to be tested and compared. For inverse power of distance it allows different powers to be compared. The input to XVALID is consistent with the input required for the grade estimation process ESTIMA. In fact the three input parameter files are identical for the two processes. The cross-validation method works by removing each point in turn from the data file and estimating its value from the remaining data. In this way a table of actual and estimated values is created. A detailed statistical analysis is then carried out comparing the actuals and estimates. One or more of the estimation parameters can then be changed and the process rerun to see whether the new parameters improve the results of the statistical analysis. The method is therefore iterative, requiring several runs to establish the best set of parameters.

Cross Validation Optimization

Cross-validation is an iterative process – you make your first estimate, look at the stats, change one of the parameters, and try again to see if you improve the stats. XVOPT does all this for you. You need to define an initial or base set of parameters as before, but you also specify a minimum and maximum for each parameter. Then you define a step size for discrete increments between the maximum and minimum. XVOPT then just loops round for all possible combination of parameters and calculates the statistics. So what you also need to do is to define a sort of objective function. You specify weights for each statistic defining their relative importance, and penalty for the distance of the actual statistic from it’s target value. defining their relative importance, and a penalty for the distance of the actual statistic from its target value. The process then calculates the penalty for each run, weights them and finds the one with the lowest penalty points.