Validation Curves
A plot of training and validation error against one hyperparameter with data fixed, whose validation curve traces a U marking over- and underfitting.
Training error falls, or flattens, monotonically as the swept hyperparameter gives the model more freedom. Validation error traces a U instead: high where the model is too constrained to capture real signal, high again where it has enough freedom to fit training noise, lowest where its capacity matches what the data supports.
Only the validation curve's minimum is worth keeping. If it sits at the edge of the swept range, the range was too narrow, not the answer.