ATLASResearch
methods
Quantitative/ Model selection

Hyperparameter search

Search configurations under a fair evaluation budget

Hyperparameter search chooses settings not fitted directly by the learner, such as regularisation strength, depth or learning rate. Grid search enumerates a fixed lattice; random search samples a defined distribution; Bayesian optimisation uses previous trials to guide future trials. All require a validation criterion and budget, while final assessment needs untouched test data or outer resampling.

WHEN IT FITS

Use it when plausible configurations materially affect performance and you can afford a transparent search budget, particularly when comparing algorithms whose default settings receive unequal tuning effort.

Strengths

  • Replaces undocumented manual adjustment with a reproducible search
  • Random or adaptive search can explore influential parameters efficiently

Limitations

  • Large searches can overfit validation noise
  • Unequal compute budgets can make comparisons misleading

Know the boundary

The best validation trial is not guaranteed to be the best generalising model.

USED ACROSS
Computer science