Hyperparameters are configuration values set *before* training begins and are not updated by the optimization process itself — learning rate, batch size, number of layers, regularization strength, and number of training epochs are canonical examples. Hyperparameter tuning is the systematic search for the combination of these values that yields the best model performance on a validation set, using strategies such as grid search, random search, or more sample-efficient approaches like Bayesian optimization and population-based training.

This is explicitly distinct from option A, which describes the *training* process itself — weights and biases are trainable parameters, updated automatically via backpropagation and gradient descent, not selected through hyperparameter search. Option B describes algorithm selection, a higher-level modeling decision that may precede hyperparameter tuning but is not what tuning itself accomplishes (you tune hyperparameters *within* a chosen algorithm/architecture). Option C describes data engineering work that happens upstream of model training entirely, unrelated to parameter search.

In practice, hyperparameter tuning requires careful experimental design to avoid overfitting to the validation set — techniques like k-fold cross-validation, held-out test sets, and tracking tools (e.g., experiment trackers logging each trial's configuration and resulting metric) are standard practice, connecting this topic directly to the Experimentation domain's broader emphasis on rigorous, reproducible model evaluation.