Nonparametric regression, such as smoothing spline, is widely used in many scientific disciplines as a valuable data-analyzing method. The use of a smoother requires the choice of a smoothing parameter which, by balancing fidelity and roughness, controls how much smoothing is done. Two popular selection criteria to choose the smoothing parameter are $C_p$ and generalized maximum likelihood (GML). Each, however, has its own problem. For $C_p$ the problem is its high variability, while for GML, the problem is its potentially big bias. In this talk we propose a new selection procedure: the extended exponential (EE) criterion, which combines the strengths of $C_p$ and GML, yet avoids their weaknesses in that the EE\ criterion has (a) small variability and (b) small bias. In addition to these, it also has (c) small tendency toward under and oversmoothing. All three criteria turn out to have simple geometric interpretations, which plays a pivotal role in our finite-sample, non-asymptotic theoretical analysis. The EE criterion is also shown to be more robust against non-normality. Some large sample results will be presented and compared with their finite sample counterparts.