Which parameters matter

Before optimizing a dozen parameters, it can pay to find out which of them matter. sensitivity() evaluates the objective at points chosen to measure each parameter's effect, through the same workers, timeout and failure handling as optimize(), and reports them most important first:

from hone import choose_optimizer, sensitivity

report = sensitivity(run_simulation, search_space, max_evaluations=500, workers=8, seed=0)
print(report)
reduced = report.reduced(search_space)  # the unimportant parameters fixed
result = choose_optimizer(reduced, max_evaluations=200).optimize(run_simulation, workers=8)

How it measures depends on the budget:

Method	Evaluations	Reports
Morris screening	About 10 × (parameters + 1), at least 2 × (parameters + 1)	mu_star, the mean size of each parameter's effect: the change in the objective across its whole range. mu is the mean with its sign, and sigma large against mu_star means the effect depends on the other parameters
Sobol indices	64 × (parameters + 2) or more, in powers of 2	total, the share of the objective's variance the parameter accounts for with its interactions, and first, the share it accounts for alone. A total near 0 means the parameter can be fixed with little effect

Without method=, Sobol is used when at least 64 × (parameters + 2) evaluations fit in max_evaluations, and Morris otherwise. Each index comes with a 95% bootstrap interval. report.ranking lists the names, most important first, report["fillet"] gives one parameter, and report.best_params is the best point evaluated (maximize=True for the highest).

report.reduced(search_space) returns the search space with the parameters report.unimportant() names fixed: those whose whole interval is below 0.05 for Sobol, or a tenth of the largest mu_star for Morris. threshold= changes that, and keep=4 keeps the four most important instead. A fixed parameter is still passed to the objective, at its value in the best point evaluated unless at={...} gives one. A parameter can barely change the objective and still have a best value, so fixing it can cost a little; fix one you know the best value of with at=.

Every kind of parameter is measured on the scale the optimizers search it on: a log range in its logarithm, an integer range rounded, and a choice parameter among its choices. A conditional parameter is passed only while switched on, so its indices include the points where it has no effect. Search spaces with constraints aren't supported yet.

examples/heat_sink.py finds which of a heat sink's eight settings drive a chip's temperature, fixes the four that don't, and on a budget of 60 evaluations finds a cooler design than a search over all eight.