Secretary-Problem Sampling Rule¶
Heuristic policy — instantiates Sequential Stopping Boundary Design
Splits a no-recall sequence into a learn-only sampling phase and a commit phase, then takes the first later option that beats everything seen so far.
A Secretary-Problem Sampling Rule solves a very specific version of the stopping problem: options arrive one at a time, each must be accepted or permanently rejected on the spot, and you cannot go back to a passed-over option later. Its defining move is to divide the sequence into two phases and set no numeric value threshold at all. In the first phase you only sample — you interview, view, or test a fixed fraction of the options and reject every one, using them purely to learn the standard the field can offer. In the second phase you commit to the first option that beats the best one seen during sampling. The "threshold" is not a value; it is the best-so-far record established in the sampling window, and the boundary is positional — defined by where you are in the sequence, not by how good the current option is in absolute terms.
Example¶
Someone relocating has a tight two weeks to find an apartment in a hot rental market, where good listings are taken within hours — see it, decide now, or lose it, with no calling back yesterday's place. They expect to view about 20 units in the window. The sampling rule tells them: for roughly the first seven viewings (about 37% of the pool), commit to nothing — just look, and remember the best one. The best of those first seven becomes the benchmark. From viewing eight onward, take the first apartment that is better than that benchmark.
Their fallback is a month-to-month sublet they can fall back on if nothing clears — the walk-away that keeps them from panic-signing. Unit twelve turns out to be the first that beats the sampling-phase best, so they sign it that afternoon rather than gambling that units thirteen through twenty hold something better. They will never know whether unit nineteen was a gem; the whole point of the rule is that, given no-recall and a hard deadline, "sample ~37%, then take the next record-beater" gives them a strong shot at the best available without paralysis — and if nothing beats the benchmark, they land on the sublet rather than the last tolerable unit.
How it works¶
- Fix the pool and the split. Estimate how many options the window holds and set the sampling fraction — classically about the first 37% — as a pure learning budget.
- Sample without committing. View, interview, or test each option in the sampling phase and reject all of them, recording only the best-so-far as the benchmark.
- Switch to commit mode. Past the split, accept the first option that exceeds the sampling-phase benchmark; the positional boundary, not an absolute score, triggers the stop.
- Fall back if none clears. If no post-split option beats the benchmark, the sequence ends on the pre-named fallback rather than on whatever arrived last.
Tuning parameters¶
- Sampling fraction — how much of the pool is spent learning before committing. The classic value maximizes the chance of the single best pick; a smaller fraction commits sooner and settles for good-enough, a larger one risks sampling past the winner.
- Objective — whether the rule chases only the very best or expected quality. Optimizing for best-or-nothing uses the classic split; softening toward expected rank shortens the sampling phase and accepts strong-but-not-top options.
- Pool-size estimate — the assumed number of options. Overestimating stretches the sampling phase and can burn the whole window learning; underestimating commits too early.
- Recall leniency — whether any late return to a passed option is possible. Even a little recall breaks the rule's premise and argues for a value-threshold mechanism instead.
When it helps, and when it misleads¶
Its strength is that it needs no prior and no value scale — only the ability to rank options against each other — which makes it usable exactly where you cannot say in advance what "good" numerically means. It is the classic secretary problem, whose optimal policy is the sample-then-commit rule with the roughly 37% split.[1] When options genuinely arrive sequentially, cannot be recalled, and you mostly care about not missing the best, it is close to the best you can do.
It misleads the moment its premises quietly fail. If rejected options can in fact be recalled, the rule needlessly throws away good early picks it could have returned to; if the sequence is not order-random — the best listings front-loaded or saved for last — the 37% calibration is off and the benchmark is set on an unrepresentative sample. It is also famously all-or-nothing: tuned for the single best, it has a real chance of ending on the fallback, which stings when a strong-but-not-top option was passed in the sampling phase. The guarding discipline is to confirm no-recall and rough order-randomness before using it, and to switch the objective toward expected quality — or to a value-threshold mechanism entirely — when "a great one" matters more than "the best one."
How it implements the components¶
observation_sequence_window— the rule is defined on the arrival sequence: it presumes options come one at a time in (roughly) random order and cannot be revisited, and its two phases are positions within that window.sampling_budget— the first fraction of the pool is an explicit exploration budget spent entirely on learning the field's standard, committing to nothing.current_best_and_fallback_state— the best option seen during sampling becomes the benchmark the commit phase must beat, and a pre-named fallback catches the case where nothing does.stopping_boundary— the boundary is positional: reject through the sampling split, then stop on the first record-beater — a rule of place-in-sequence rather than absolute value.
It carries no prior_distribution_or_base_rate and does no signal-by-signal belief revision — that is Bayesian Value-of-Information Update; and it deliberately sets no per-interval value schedule with a decision_rationale_trace, which is the Reservation Value Table.
Related¶
- Instantiates: Sequential Stopping Boundary Design — the boundary for the pure no-recall, sequential-arrival case, solved by exploration-then-commitment.
- Sibling mechanisms: Reservation Value Table · Bid Acceptance Cutoff · Bayesian Value-of-Information Update · Real-Option Exercise Boundary · Sequential Monitoring Stop Rule · Stop-Rule Postmortem
Editorial Notes¶
Form Classification¶
Form family: Decision, Gate & Allocation
Rationale: Secretary-Problem Sampling Rule operates as a case-specific gate, selection, routing, prioritization, or resource disposition because it splits a no-recall sequence into a learn-only sampling phase and a commit phase, then takes the first later option that beats everything seen so far.
Independent corroboration: The frozen evidence defines Secretary-Problem Sampling Rule as 'Splits a no-recall sequence into a learn-only sampling phase and a commit phase, then takes the first later option that beats everything seen so far', so its operative form is Decision, Gate & Allocation.
Nearest alternative: Rule, Policy & Commitment — Secretary-Problem Sampling Rule includes features of a standing rule, threshold, contractual commitment, or policy constraint governing future conduct, but its defining operation is a case-specific gate, selection, routing, prioritization, or resource disposition.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Operations Research
Origin pattern: Convergent development
Present-day reach: Multi-domain
Rationale: The sample-then-select no-recall rule is the canonical secretary problem in optimal stopping.
Related originating lineages:
- Economics & Finance — Sequential search independently applies it to hiring and offers.
- Mathematics — Probability theory materially proves its stopping properties.
Review resolution: The blind reviewers agree that operations_research is the primary origin and differ only on origin mode disagreement, domain reach disagreement. I preserve every independently explained alternate from both records rather than imposing a numeric cap. I retain convergent because the combined record shows independent disciplinary development. The broader reach of multi_domain records portability separately from historical provenance, and encyclopedia_synthesis=false preserves the affirmative synthesis judgment where either reviewer identified one.
Review outcome: Reconciled after independent review; high confidence.
References¶
[1] Gilbert, J. P., & Mosteller, F. "Recognizing the Maximum of a Sequence". Journal of the American Statistical Association 61(313), 35–73 (1966). Derives the classic secretary problem’s observe-then-select policy with an asymptotic cutoff of about 1/e, or 37 percent. registry ↩