I have some question about the best subset selection in statistical learning.

There are 9 independent variables X_i and 1 dependent variable. I need to use these 9 independent variables to predict dependent variable Y. I have splitted the data to two parts. 75% data will be my personal training data, and 25 % data will be my personal testing data. I need to find a multiple linear model which has the lowest test mean squared error. I have found that all 9 independent variables should be added to the model. Now I need to consider the interactions and X_i². What should I do now? Do I need to repeat the subset selection for all possible interactions and X_i² ?