Does an unsuccessful search provide information?

Yes, when its detection capability is known or modeled sufficiently to interpret the failure.

Why do unsearched locations gain posterior probability?

The failure makes searched locations relatively less plausible after normalization.

Is this the same as Bayesian optimization?

No. Here the main task is locating a target with imperfect detection, rather than optimizing an expensive objective function.

AI Glossary / Statistics and probability / Intermediate

Bayesian search theory

Definition

Bayesian search theory is a framework for allocating search effort using probabilities of a target's location, models of detection, and Bayesian updates as information arrives. An unsuccessful search changes the location probabilities according to how likely that search would have detected the target. The framework supports planning under uncertainty; it does not guarantee that the target will be found.

What it means

Two uncertainties must be distinguished: where the target is and whether a particular search would detect it if it were there. A likely location can be difficult to inspect.

Failing to find something is therefore evidence of varying strength. A thorough search with reliable detection is stronger negative evidence than a brief inspection with obstructed visibility. The location probabilities and detection model should reflect this distinction.

This entry concerns locating targets, rather than internet search ranking.

How it works

Define a set of possible locations or trajectories and assign prior probabilities using available observations and explicit scenario assumptions. Include relevant possibilities outside an initially convenient search region rather than silently treating that region as certain to contain the target.

Estimate detection probability for each proposed search action conditional on target location. It may depend on equipment, visibility, coverage, target characteristics, and effort. Evaluate proposed allocations against an objective and operational constraints, including travel time and available resources.

For a stationary target in mutually exclusive cells, let p_j be the location probability and d_j the probability of detection under that search. After no detection, the posterior is p_j(1 - d_j) divided by the sum of those quantities across all cells, provided that sum is positive.

Use the updated distribution to inform the next action. Moving targets need a motion model between observations, and repeated searches require appropriate dependence assumptions. A simple one-step ranking by detection probability is not automatically optimal for an entire sequence with travel costs or changing conditions.

Three practical examples

Asset recovery: prioritize warehouse zones for finding a misplaced tagged item when detection reliability varies by storage arrangement.
Robotic inspection: allocate sensing actions across possible target locations while updating beliefs after unsuccessful observations.
Search planning research: combine location scenarios, environmental movement, and detection models when evaluating alternative search strategies.

Think of it this way

Think of looking for a misplaced key with a map of likely places and an honest assessment of how carefully each place was checked. “I looked there” carries little information without knowing whether the key would have been visible. A negative result should reduce belief according to the quality of the inspection.

Worked example

Suppose an inventory tag is certainly in one of three disjoint zones: A, B, or C. Prior probabilities are 0.50, 0.30, and 0.20. A search of A detects the tag with probability 0.80 if it is there and cannot detect it in B or C. Assume no false detection and that the search does not move the tag.

Before searching, the probability of detection is 0.50 × 0.80 = 0.40. If nothing is found, unnormalized location weights are 0.50 × 0.20 = 0.10 for A, 0.30 for B, and 0.20 for C. They sum to 0.60, the probability of no detection.

After normalization, the posterior is A = 1/6, B = 1/2, and C = 1/3. A is less likely but not ruled out. B and C increase in relative probability without the tag physically moving.

Now suppose equally costly next actions would detect a present tag with probabilities 0.80 in A, 0.60 in B, and 0.90 in C. Their conditional chances of success are 2/15, 0.30, and 0.30. B and C tie under this one-step objective, even though B has higher location probability. The repeated-A probability assumes the stated detection rate remains valid conditional on the previous failure.

Illustrative case study

Fictional case study: North Quay Logistics cannot locate a reusable tracking device. Staff combine the last recorded handover with storage records to assign initial zone probabilities. They separately estimate how likely a permitted scanner pass is to detect the device in each zone.

After an unsuccessful pass, the team records actual coverage and equipment conditions before updating the map. A blocked aisle leads to a lower assessed detection probability than originally planned, so the negative result is less decisive.

The team uses the revised probabilities to compare its next inspection options. The case illustrates disciplined use of negative evidence, not a promise that the initial location scenarios are complete or correct.

Common misconceptions and mistakes

Do not set a searched area's probability to zero unless the model genuinely gives perfect detection under that search. Also do not confuse the posterior location map, which is normalized conditional on no detection, with the cumulative probability that the search plan would have succeeded.

Avoid multiplying repeated miss probabilities without considering dependence. The same obstruction can cause repeated failures. Poorly justified priors, incorrect coverage records, or an omitted outside location can produce a confident but misleading map.

The highest-probability location is not necessarily the best next action when detection performance, cost, or travel differs.

Business use

Organizations can use the framework to make search priorities explainable and to learn from unsuccessful inspections. It encourages teams to record both where they searched and what their equipment could realistically detect.

For operational use, maintain provenance for location assumptions, search coverage, and detection estimates. Define responsibility for revising assumptions when evidence conflicts with the model. Decisions about continuing or ending a search require context and consequences; no single probability threshold is universally appropriate.

Check your understanding

What is A's posterior probability after the failed search?

Answer: 1/6.

Why do B and C tie for the next search?

Answer: Their different location and detection probabilities yield the same product, 0.30.

What could invalidate the repeated-A calculation?

Answer: Dependence that makes its new detection probability differ from the assumed 0.80 after the earlier miss.

Related terms

Decision theory
Bayesian updating
probability of detection
probability of containment
search effort allocation
negative evidence
target tracking

Learn next

study detection-versus-effort curves, moving-target models, and sequential search objectives. Practice distinguishing a normalized posterior map from accumulated detection probability and examine how uncertainty in detection estimates affects planning.

Sources

Sources: J. Van Gurley and Lawrence D. Stone, Bayesian Search for Missing Aircraft, Ships, and People, https://www.siam.org/publications/siam-news/articles/bayesian-search-for-missing-aircraft-ships-and-people/ ; J. R. Frost with the U.S. Coast Guard, The Theory of Search, https://navcen.uscg.gov/sites/default/files/pdf/Theory_of_Search.pdf . Primary explanations were consulted. The original three-zone calculation was independently checked using exact fractions.

Previous: Bayesian Ridge Regression Next: Bayesian statistics