📊 Evaluating Hyperparameter Tuning Results

About this simulation

Finding a good hyperparameter combination is only half the job — you also have to judge whether the result is actually trustworthy. This scene renders a validation-loss landscape over two hyperparameters (learning rate and regularization strength). The purple surface is the validation loss the model would score on unseen data at each combination; the marker ball sits at your current choice. Floating bars compare that point's training loss against its validation loss — the gap between them is the classic sign of overfitting.

How it works

• Learning rate & regularization — move the marker across the loss surface; too little regularization with an aggressive learning rate widens the train/validation gap (overfitting), while too much regularization pushes both losses up together (underfitting).
• Overfit gap — validation loss minus train loss. Near zero is healthy; a large gap means the model memorized training data instead of generalizing.
• K-fold count & dataset size — orange dots scattered around the marker are individual cross-validation fold scores. More folds and more data shrink their spread, narrowing the reported confidence interval — this is why a single train/test split can be misleading.
• 🎯 Jump to best config — snaps the marker to the true minimum of the underlying landscape, so you can see what "good" looks like.

Controls

• Learning rate — step size used during optimization (log-scaled).
• Regularization strength — penalty applied to model complexity.
• Dataset size — more samples reduce estimate variance across folds.
• K-fold count — number of cross-validation splits (2–10). Drag to rotate the scene, scroll to zoom.

Did you know?

A model can look excellent on a single validation split purely by chance. Reporting a mean ± standard deviation across k folds — rather than one lucky (or unlucky) split — is one of the most reliable ways to tell whether a tuning result will hold up on new data.