Time Budget (AutoML)
Ask an AutoML system to find you the best model and the honest first question back is "how long have I got?". The time budget — the wall-clock allowance the search may spend — is the single setting that shapes the answer most.
Two budgets, not one
A workable configuration sets more than one clock:
- a total budget for the whole search: auto-sklearn's time_left_for_this_task, FLAML's time_budget, H2O's max_runtime_secs;
- a per-trial limit that kills one pathologically slow configuration before it swallows the run — auto-sklearn's per_run_time_limit;
- a reserve for refitting and ensembling after the search, which naive settings forget and then run out of time to use.
Spending it well
Modern schedulers treat the budget as something to allocate rather than merely obey. Successive halving and ASHA start many configurations cheaply and promote only the promising ones; BOHB adds Bayesian proposals on top. A good system is also anytime: interrupt it at any moment and it returns the best model found so far, its quality curve climbing steeply early and flattening later.
Why it is reported, never assumed
Doubling the budget can change which library wins, so AutoML benchmarks fix it explicitly — one hour and four hours per dataset are the common settings — and compare only under that constraint. Treat it as a first-class parameter of the experiment: too short and you are measuring a warm-up, too long and you are paying cloud hours for third-decimal gains.
A time budget is the wall-clock allowance given to an AutoML search, normally split between a total limit for the run and a per-trial limit that stops one slow configuration consuming everything. It determines the outcome directly, since the search explores only as far as the clock permits — which is why budget-aware schedulers such as ASHA exist and why benchmarks fix the budget before comparing tools. A reported result without its budget means little.
Imagine sending a friend into an enormous car showroom and saying "find me the best car". Give them ten minutes and they will grab something decent near the door. Give them a whole afternoon and they will test-drive properly and come back with a better pick.
A time budget is that same instruction given to AutoML software: how long it may spend trying out different models before it has to hand one over. Good tools use the time cleverly, abandoning weak candidates early and always keeping their best find ready in case you stop them. It also explains why two people testing the same tool can disagree about it — they gave it different amounts of time.