hiperparàmetres

Hiperparàmetres are configuration settings that influence the behavior of an algorithm but are not directly learned from data during training. Unlike model parameters, which are updated implicitly by optimization procedures, hyperparameters are specified before training and remain fixed throughout the process. Common examples include learning rate, number of hidden layers in a neural network, tree depth in decision trees, batch size, regularization coefficients, and the number of clusters in clustering algorithms.

The concept of hyperparameters emerged with early machine learning methods in the 1980s and has become central as models grow in complexity. They serve to balance bias and variance, control overfitting, and guide the search space of learning algorithms. Their selection often determines the practical performance and computational cost of a model.

Typical hyperparameter selection strategies encompass manual tuning, grid search, random search, Bayesian optimization, gradient‑based approaches, and hyperband. Each method trades off thoroughness against computational expense. In many modern systems, hyperparameter optimisation is automated, often treating the optimisation problem itself as a machine learning task.

Successfully tuned hyperparameters can yield significant improvements in accuracy, convergence speed, and generalisation capabilities. Conversely, poorly chosen values may result in underfitting, overfitting, or unnecessarily long training times. Researchers and practitioners continually develop novel optimisation techniques, reflecting the importance of hyperparameters in achieving reliable predictive performance.