How would you choose regularization strength without using the test set?
Instruction: A standardized ridge-regression pipeline gives validation MSE 12, 8, and 11 for alpha 0, 1, and 100. Training MSE is 2, 5, and 10 respectively. All scores use the same training/validation split. Select a candidate and describe final evaluation; the test set has not been opened.
Context: Use training and validation evidence to select model complexity while preserving an independent final evaluation.
Updated
Official answer available
Read the opening below, then unlock the full answer and practical guidance.
Choose alpha=1 provisionally: it has the lowest supplied validation MSE, 8. Alpha=0 fits training very closely but has a large generalization gap; alpha=100 increases both errors, suggesting too much shrinkage in this comparison. This small experiment does not establish that 1 is universally optimal.
For a stronger selection, compare candidate...