I use the "Geostatistical Wizard" in ArcMap 10 to do my interpolations. For example, when I do IDW and I want to decide for the optimal p power or the optimal number of neighbors that will contribute with weights to my interpolation I try several runs with several combinations (f.i. for 5: min number of neighbors and 10: max number of neighbors), I try all different p powers from 1 and above, and every time I check the RMSE and ME provided. I write down on an excel table all my results and decide for the optimal p or number of min - max neighbors based on which combination gives me the smallest RMSE and ME. What I don't understand and I ask you to help me is, why when I press the optimal button for the p power, the p that is computed usually gives an RMSE and ME larger than what I have computed when I try the different p powers manually?
cross validation uses a set of point for the validation that is different from the set o points used in the model. Therefore the RMSE with an independant set of point is larger than the RMSE on the points used to build the model.