Linear modeling by best subset selection

The linear models of highest rI (for the learning set) containing n = 1. 5 descriptors are the following  

Random experiments were performed to select best combinations of n = 1. 5 out 

Despite the large descriptor pool, mhrR values here are somewhat lower than in the previous example, illustrating the (positive) effect of 80 instead of 50 observations to be fitted. For our best models, the difference between and mhrR is between 

2 and 30 standard deviations, so the original models fit the data far better than the random models, and it is extremely unlikely that our models are based on chance correlations. 

For most models. Rig is higher than R g, as expected. Further, the best models with respect to Rig do not, as a rule, exhibit the best values of R g. These data are visualized 

---