Run unmodified training scripts, capturing each run result as a unique experiment.
Automate trials using grid search, random search, and Bayesian optimization.
Compare and analyze runs to understand and improve models.