| Strategy | Trials | Total fits | Estimated time | Coverage |
|---|
Formula used
Grid search evaluates the Cartesian product of active parameter values. Conditional parameters only multiply compatible branches. Exclusion rules then remove invalid configurations.
How to use
- Select a preset or create custom parameter rows.
- Enter comma-separated values or generate numeric ranges.
- Mark fixed values and configure conditional dependencies.
- Add exclusion rules for impossible parameter combinations.
- Set folds, repeats, datasets, seeds, and model count.
- Enter timing, worker, memory, storage, and cost assumptions.
- Calculate, inspect warnings, preview combinations, and export results.
Worked example
A grid with three learning rates, four depths, and two optimizers contains twenty-four combinations. Five-fold validation produces one hundred twenty model fits. Four efficient workers reduce elapsed time significantly.
| Parameter | Values | Count |
|---|---|---|
| learning_rate | 0.001, 0.01, 0.1 | 3 |
| max_depth | 3, 6, 9, 12 | 4 |
| optimizer | adam, sgd | 2 |
| Total combinations | 24 | |
Frequently asked questions
What is a grid search combination?
It is one unique selection containing one value from every active hyperparameter.
Why do combinations grow quickly?
Every additional value multiplies the entire existing Cartesian search space.
Do fixed parameters increase the count?
No. They remain in exported configurations but add a multiplier of one.
How are conditional parameters counted?
They apply only when their parent parameter matches an allowed condition value.
How do exclusion rules work?
Any generated configuration matching every condition in a rule is removed.
Why are total fits larger than combinations?
Cross-validation, repeats, datasets, seeds, and models multiply training operations.
Does doubling workers halve runtime?
Only ideally. Worker efficiency accounts for scheduling, data, and communication overhead.
When should random search be preferred?
It is useful when the full grid is large or many dimensions matter unevenly.
What does Bayesian search change?
It chooses promising trials sequentially instead of exhaustively evaluating every combination.
Can the preview show millions of combinations?
The calculator limits browser generation while preserving exact analytical totals when possible.