I am using Kriging for optimization where local searches are often carried out. However, because of these local searches my sample design becomes poor and my Kriging prediction as well as variance estimates are poor. I read that this is because of ill conditioning of correlation matrix. Choosing a small subset of data is not preferred for me because then you are just losing information. Question: Are there any other alternatives (including subset selection given that I am able to utilize most of my data)? I am using a nugget term and that doesn't seem to help much.