Observations: just based on the R-Squared values, it seems that Lasso and Elastic Net were able to find combinations of predictors fitting the data slightly better than stepwise. Of course, R-Squared alone only depicts a partial picture about the quality of a model and, regardless of the approach chosen for variable selection, additional quality metrics should be assessed. Finally, this is such a small dataset that I would be hesitant to place too much confidence in the ability of any model to learn meaningful patterns