Bayesian search theory¶

A probabilistic lost-object search using location beliefs, detection chances, and updates when results arrive.

Version: v1 · 2026-09-28 · History
Domain-specific #: 8141
Domain group: Formal Sciences
Origin domain: Operations Research
Subdomain: Search Theory → Operations Research
Aliases: Bayesian search

Core Idea¶

Bayesian search theory treats a missing object's possible locations as hypotheses with prior probabilities and models the chance of detecting it at each place. Those two ingredients can guide an initial search before any outcome exists. After a failed search, Bayes' rule discounts locations in proportion to how likely failure would have been there; imperfect detection leaves residual probability. Normalized posterior beliefs can guide further search when combined with effort and future detection prospects.

A two-cell worked beacon case changes A's probability from 0.6 to about 0.23 after an 80%-effective search of A finds nothing; the same prior and detection model could have guided which cell to inspect first. Metron used prior maps, prior unsuccessful sweeps and detection modeling in BEA's AF447 underwater search planning. That is a documented application, not a claim that a modeled probability map was an observed location. Bayesian Updating governs the post-result revision stage; location, physical detection and effort also define the pre-result search plan.

How would you explain it like I'm…

Smart Lost-Toy Hunting

When you lose a toy, you think about where it most likely is and how easy it would be to spot there. You look in the best spot first. If you look under the bed and don't find it, the bed becomes a less likely spot, but not impossible, because you might have missed it. Then you pick the next best place to look.

Smart Searching With Chances

When something is lost, Bayesian Search Theory starts by giving each possible spot a chance of holding it, and a chance of spotting it if you search there. Those two numbers help you pick where to search first. If you search a place and find nothing, you lower that place's chance, but only as much as the search was good; a sloppy search lowers it only a little. The other places go up so all the chances still add up. Then you use the new chances, plus how much effort each search costs, to decide where to go next.

Probability-Guided Search Planning

Bayesian Search Theory treats each possible location of a missing object as a hypothesis with a prior probability, and also models the chance of detecting the object if you search there and it is present. Before any search, these two ingredients already help plan where to look first. After an unsuccessful search, Bayes' rule lowers a location's probability according to how likely a miss would have been there; because detection is imperfect, some probability remains. For example, if location A starts at 0.6 and a search of A that would find it 80% of the time comes up empty, A's probability drops to about 0.23. The updated, normalized probabilities, together with the effort needed and future chances of detection, guide the next search.

Bayesian Search Theory models a search as a set of location hypotheses with prior probabilities plus a detection model giving the probability of finding the object in a cell if it is there and searched with a given effort. These ingredients guide the initial allocation of search before any result exists. When a search fails, Bayes' rule multiplies each searched cell's probability by its probability of missing, then renormalizes, so imperfectly searched cells keep residual probability. In a two-cell case with A at 0.6, an 80%-effective search of A that finds nothing leaves A at 0.12/0.52, about 0.23. Subsequent search is planned from the posterior combined with effort costs and future detection prospects. The update stage is ordinary Bayesian updating; what makes this a search theory is that location, physical detection, and effort also shape the plan before any outcome. Metron used prior maps, earlier unsuccessful sweeps and detection modeling in BEA's search planning for the AF447 wreckage; the probability map was a planning tool, not an observed location.

Structural Signature¶

Sig role-phrases:

Missing-object locations — Defines mutually accountable hypotheses about where the object could be. It is constitutive. Counterfactual: A known location removes this search uncertainty.
Prior location probabilities — Assigns initial mass across possible places before a given search. It is constitutive. Counterfactual: An uncalibrated hotspot sketch is not a probability prior.
Conditional detection chance — Estimates how likely a search would find the object if it occupied a particular place. It is constitutive. Counterfactual: A searched cell cannot be declared empty when detection is imperfect.
Observed search outcome — When a search has occurred, supplies detection or non-detection evidence for revising location beliefs. It is diagnostic. Counterfactual: A pre-search plan can still be Bayesian search theory before this evidence exists.
Posterior and next-search frame — When results arrive, normalizes revised beliefs; before and after results, compares detection effectiveness and effort for the next action. It is boundary. Counterfactual: Highest location mass need not be the cheapest or most detectable next region.

What It Is Not¶

Not a guaranteed location. A prior or posterior is conditional on model assumptions.
Not perfect area clearance. A no-find need not zero a searched cell.
Not a hotspot map alone. Detection chances must help plan the search, with updates if results arrive.
Not a route by probability alone. Detection and effort also matter.
Closest near-miss. A grid walked in fixed order with no probabilistic location or detection model is the nearest miss.

Scope of Application¶

Underwater recovery. Plan wreckage searches using explicit prior and detection assumptions.
Land search. Adapt detection models to terrain and cover when supported by evidence.
Operations research. Compare expected detection per unit effort after probability revision.
Search audit. Evaluate how past no-finds should update location beliefs.

Clarity¶

Name the missing object, location hypotheses, prior and detection model. If a search has occurred, state its result and the resulting update; a pre-search plan still qualifies. A fixed grid walk without probabilistic location or detection modeling is the nearest miss. A posterior ranking and a cost-effective next route are related but distinct outputs.

Manages Complexity¶

Bayesian search compresses disparate clues and unsuccessful sweeps into one conditional map. The compression remains assumption-bound: detection estimates and prior hypotheses can be wrong, and a searched place can retain nonzero probability. Keeping observation, inference and choice distinct prevents a model output from being misread as discovery.

Abstract Reasoning¶

Define possible locations and prior mass.
State detection chance conditional on location and search conditions.
Compare initial search options using prior mass, detectability and effort.
If a search result arrives, normalize prior times outcome likelihood to obtain a posterior.
Compare further searches by revised mass, detection and effort without hiding model limits.

Knowledge Transfer¶

The update stage uses Bayesian Updating literally when an outcome arrives, but a prior-and-detection search plan can exist before that stage. The added cargo is uncertain physical location and imperfect detection during a search. Diagnosis may share the mathematics but is not this lost-object problem. A plan that disregards results once observed lacks the method's feedback capacity.

Cross-Domain Echoes¶

See how this entry connects to another domain.

What one result can—and cannot—tell you

Examples¶

Canonical¶

A lost beacon is initially in A with probability 0.6 or B with probability 0.4. A search of A would detect a beacon there 80% of the time. After finding nothing in A, its unnormalized weight is 0.6×0.2=0.12; B's stays 0.4, so the new probability of A is 0.12/0.52≈0.23. The search has reduced A's probability without making it impossible; the next area also depends on cost and detection.

Mapped back: Missing-object locations → beacon in A or B; Prior location probabilities → 0.6 and 0.4; Conditional detection chance → 0.8 if beacon is in searched A; Observed search outcome → no-find in A; Posterior and next-search frame → A becomes about 0.23; future choice remains effort-dependent.

Applied / In Practice¶

Metron's BEA-hosted AF447 underwater wreckage analysis estimated a prior location distribution, accounted for earlier unsuccessful searches and modeled detection effectiveness to revise a probability map for subsequent search planning. BEA's sea-search report documents this use. The map was a model-conditioned guide, not an observed wreckage location or a guarantee of recovery.

Mapped back: Missing-object locations → possible AF447 underwater wreckage locations; Prior location probabilities → Metron's pre-search map; Conditional detection chance → phase-specific underwater search effectiveness; Observed search outcome → earlier unsuccessful sweeps; Posterior and next-search frame → revised map used with effort and detection limits.

Structural Tensions¶

T1 — No-Find Evidence versus Imperfect Detection. An unsuccessful sweep is weaker evidence where the sensor could easily miss the object.

Diagnostic: How detectable was the object if present?

T2 — Posterior Concentration versus Search Effort. Highest probability and highest expected detection per effort can point to different places.

Diagnostic: Is the map mistaken for a complete route decision?

Structural–Framed Character¶

Bayesian search is formally structural in its probability update but bounded by missing-object and physical-detection roles. Evaluative weight: efficient search is a goal, not a property of the posterior alone. Human-practice-bound: teams choose search effort; normalization is formal. Institutional origin: BEA's use documents an application, not the theorem's origin. Vocabulary travels: prior and likelihood travel widely; no-find sensor performance is specialist. Import versus recognize: an actual lost-ship search can instantiate the form; hunting metaphorical ideas cannot.

Bayesian Updating is a verified operation for the post-result stage, not a strict parent of a pre-result plan. Its character: a formal search-planning method with imperfect physical observation and conditional updating.

Structural Core vs. Domain Accent¶

The Bayesian update is portable; physical lost-object search supplies the specialized roles.

What is skeletal. A prior over locations and a detection model guide search; if an outcome arrives, its likelihood yields a normalized posterior. Failure to find is evidence because its probability differs by possible location.

What is domain-bound. Hypotheses are physical places, outcomes arise from a sensor and search path, and the next step involves detection effort. AF447's underwater conditions alter likelihoods, not Bayes' rule.

Why this does not clear the prime bar. Removing the object, location and detection leaves generic probabilistic planning or updating. The pre-result plan and post-result revision jointly form a specialist method rather than a strict subtype of the update operation alone.

This entry is a kind of Decision Method.

Related — Bayesian Updating. It governs location revision once a search outcome exists, not the entire pre-result method.
Related — probability. A map allocates mass but is not the whole search process.
Related — search operations. Route deployment uses prior or posterior mass and detection estimates.

Relationships to Other Abstractions¶

Current abstraction Bayesian search theory Domain-specific

Parents (1) — more general patterns this builds on

Bayesian search theory is a kind of Decision Method Domain-specific

Bayesian search theory satisfies the defining boundary of Decision Method: A decision method is a repeatable procedure that represents a decision frame, feasible alternatives, objectives or loss, evidence and uncertainty, preference or priority information, and an aggregation or search rule to recommend, rank, or adapt a course of action.

Hierarchy path (1) — routes to 1 parentless root

Bayesian search theory → Decision Method

Neighborhood in Abstraction Space¶

Bayesian search theory sits in a sparse region of the domain-specific corpus (66^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Geographic Mapping & Positioning (14 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

Hotspot map. Tell: Are detection chances used to plan the search?
Perfect clearance. Tell: Was detection certain if present?
Highest-probability route. Tell: Were effort and future detection considered?
Generic Bayesian inference. Tell: Is a missing object physically being sought?

References¶

Metron, Search Analysis for the Location of the AF447 Underwater Wreckage, BEA-hosted report (2011): https://bea.aero/fileadmin/uploads/tx_elyextendttnews/metron.search.analysis_01.pdf
BEA, Report on Sea Search Operations, AF447 (2012): https://bea.aero/fileadmin/uploads/tx_elyextendttnews/sea.search.ops.af447.05.11.2012.en_03.pdf
Frozen Wikipedia discovery revision: https://en.wikipedia.org/wiki/Bayesian_search_theory (revision 1371068351).