parametergrid

A parameter grid, often referred to as a param_grid, is a structured specification of hyperparameter values used to explore model configurations during hyperparameter optimization. It represents a grid of possible values for one or more hyperparameters, typically forming the Cartesian product of the provided value sets.

In practice, grids are expressed as dictionaries or mappings from parameter names to lists of values. When the parameters are part of a pipeline, names may be qualified to indicate the stage, such as 'clf__C' or 'preprocess__n_features'.

Grid search algorithms generate every combination from the grid and evaluate each by fitting the model and assessing performance, usually via cross-validation. The process yields a best-performing configuration and the corresponding performance estimate. The approach is simple and thorough but can be computationally expensive as dimensionality grows.

Limitations include combinatorial explosion and potential overfitting to cross-validation metrics. Practical use often employs smaller grids or uses alternative strategies such as randomized search, Bayesian optimization, or successive halving. Grid search remains popular for its interpretability and reproducibility, particularly when combined with well-chosen priors on hyperparameters.

Example: For a support vector machine, a parameter grid might specify C values of 0.1, 1, and 10 and kernels of 'linear' and 'rbf', yielding six configurations to evaluate.