問題文
A team has budget for 40 training runs to tune five hyperparameters. A grid with three values per parameter would need 243 runs. What is the practical approach within the budget?
選択肢
- Reduce the grid to two values per parameter, giving 32 runs, because a complete grid is the only design that lets the team attribute the effect of each parameter separately and staying complete is worth the loss of resolution in each range.
- Run all 243 configurations for one tenth of the epochs each.
- Sample configurations at random from the ranges, which explores each parameter at many distinct values within the same number of runs.
- Tune the parameters one at a time, fixing the others at their defaults.