hyperparameter

A hyperparameter is a configuration value external to a machine learning model that must be set before training and is not learned from the training data by standard optimization. Hyperparameters differ from model parameters, such as weights and biases, which are learned during training. Hyperparameters influence the learning process, including the capacity of the model, the speed of convergence, and the strength of regularization.

Examples include the learning rate used by optimization algorithms, regularization strength, the number of layers and units in a neural network, the choice of optimizer, batch size, dropout rate, and the number of trees in ensemble methods. In decision trees or boosting, the maximum depth or the learning rate acts as hyperparameters. Some algorithms also expose controls such as early stopping criteria or data augmentation settings.

Hyperparameter tuning aims to identify values that maximize performance on a validation dataset. Common strategies include grid search, which exhaustively tests combinations; random search, which samples combinations randomly; and Bayesian optimization, which builds a probabilistic model to guide exploration. Other approaches include Hyperband and population-based training, as well as nested cross-validation to assess tuning stability.

The choice of hyperparameters can substantially affect model performance and should be reported to support reproducibility. Tuning incurs computational cost and can lead to overfitting to the validation set if performed excessively. In practice, practitioners start with reasonable defaults and use systematic search constrained by available resources.