- First determine approximately how many data points are available on input. How are these points arranged? Are they evenly distributed over the area of interest or are they in clusters? Is there any noise or errors in the x,y or z values? Once you have determined these facts you can make the best choice possible. If you don't know anything about the dataset, assume the points have no noise, they are badly clustered and there are lots of them; more than a few hundred.
- The second line of reasoning to employ is to examine what these points represent and what you want or hope to get out of them. If you know that the underlying function or structure that the points are taken from is rough or smooth this can help specify the parameters needed. If you don't know, assume the function is rough. As rough as the observed points can define it.
- The last criteria is just how much computer time do you want to spend on this job? This may seem arbitrary, but if you know nothing else about the data or the results you desire, this is all that you have left.