Does this apply only to physical sensors?

No. An observation can be a software event or another recorded signal, provided its relationship to the state is modeled.

Can the assumption be made more plausible?

Sometimes. A richer state or an explicit measurement-history model can capture dependencies that were previously omitted.

Does the current observation determine the state?

No. The model describes probabilities of observations conditional on states; inference also needs prior information and does not generally identify a unique state.

AI Glossary / Graphs and symbolic AI / Advanced

Sensor Markov assumption

Definition

The sensor Markov assumption says that, given the current underlying state, the current observation is independent of earlier states and observations. It is a conditional independence assumption used to simplify probabilistic models of systems that change over time.

What it means

An observation is evidence about a state, not necessarily the state itself. A warehouse detector may report movement while the underlying question is whether a particular aisle is occupied. The report can be wrong, and the system may be uncertain about occupancy.

The assumption says that the chosen current state contains all the historical information needed to describe the distribution of today's sensor reading. It does not say that readings are unconditionally independent. Adjacent readings can be strongly associated because the underlying condition persists.

It is also different from the transition Markov assumption. The transition assumption concerns how the state evolves; the sensor assumption concerns how evidence is generated from the state. A model may satisfy one assumption more convincingly than the other.

How it works

Write X_t for the state at time t and E_t for the observation. The assumption is P(E_t | X_0:t, E_0:t−1) = P(E_t | X_t). The colon denotes a sequence of time steps. Once X_t is specified, the additional history on the left does not change the observation distribution in this model.

A filtering system first predicts the current state from its previous belief and a transition model. It then combines that prediction with the likelihood of the new observation under each possible current state. Normalization turns the resulting weights into an updated probability distribution.

The state definition is crucial. If a detector has persistent calibration drift, representing only occupancy may leave yesterday's reading informative even after occupancy is known. Adding a calibration variable, modeling correlated errors or changing the observation process may be necessary. An assumption is a modeling choice to assess, not a property guaranteed by calling a device a sensor.

Three practical examples

A robot localization model relates a current range reading to the robot's pose and map, while checking whether unmodeled sensor bias persists over time.
An equipment-monitoring system estimates a hidden operating mode from noisy measurements and examines whether recent measurement errors remain correlated.
A room-occupancy model combines predicted occupancy with a motion-detector report, distinguishing an absent report from a confirmed empty room.

Think of it this way

Think of a sealed envelope containing a complete description of the present situation relevant to an instrument. Under the assumption, opening older envelopes adds nothing to predicting the instrument's next current reading once that description is known.

The analogy depends on “complete” being relative to the observation model. Real state descriptions are often incomplete, and an instrument can retain a memory through warming, drift or internal averaging.

Worked example

Updating a probability without counting evidence twice

In an invented occupancy model, the predicted probability that an aisle is occupied is 0.30. A detector reports movement with probability 0.80 when occupied and 0.20 when empty. A movement report therefore receives occupied weight 0.30 × 0.80 = 0.24 and empty weight 0.70 × 0.20 = 0.14.

The updated occupied probability is 0.24/(0.24 + 0.14), approximately 0.632. The assumed sensor reliability has not turned into certainty; the prior distribution also matters.

Now suppose the network retransmits exactly the same physical reading. Treating that duplicate as independent new evidence would multiply by the likelihood again and produce 0.192/(0.192 + 0.028), approximately 0.873. That increase is unjustified: no new measurement occurred. The model must distinguish a new observation from a repeated message. This example isolates one update and does not assume that occupancy remains unchanged across actual later time steps.

Illustrative case study

Fictional case study: Willow Monitoring deploys a prototype state estimator for workshop equipment. Several readings in a row suggest an unusual operating condition, and the estimator becomes very confident.

The team discovers that the device publishes overlapping moving averages. Consecutive reports reuse many of the same raw measurements. A model that treats their errors as independent given the operating mode overstates how much fresh evidence each report provides.

Willow documents the averaging window and evaluates a model that accounts for the measurement process. It also checks confidence calibration against independently observed operating states. The fictional lesson is that many messages do not necessarily represent many independent observations; a convenient probability factorization needs a defensible physical interpretation.

Common misconceptions and mistakes

The sensor assumption says the world has no memory. It concerns observation dependence after conditioning on the chosen current state.
It makes successive readings independent without qualification. Dependence can remain through the underlying state process.
It requires perfect sensors. A likelihood model can explicitly represent noise and false reports.
It is the same as stationarity. A time-invariant observation distribution is a separate assumption; conditional independence alone does not require identical parameters at every time.

Business use

Businesses using streaming estimates should ask what the hidden state includes and how measurements are produced. Persistent biases, shared preprocessing, retransmissions and environmental effects can all matter to the credibility of the observation model.

Inspect errors and probability calibration rather than relying only on classification accuracy. If uncertainty estimates drive actions, excessive confidence can be consequential even when the most likely state is often correct. More elaborate modeling adds costs, so compare alternatives on representative recorded sequences and clearly describe any approximation retained.

Check your understanding

A detector's bias persists for hours, but the state contains only machine mode. Is conditional independence automatically justified?

Answer: No. Earlier readings may reveal the unmodeled bias even when the current mode is known. The state or observation model may need revision.

Related terms

Learn next

Read sensor model and Bayesian filtering. Sketch a state variable, an observation variable and a plausible persistent source of error. Explain whether that error belongs in the state or requires a different observation-dependence model.

Bayesian filtering

Sources

Russell and Norvig, Artificial Intelligence: A Modern Approach, Temporal Probability Models teaching slides: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter15a.pdf. The explicit sensor independence equation, separate transition and stationarity assumptions, and filtering derivation were reviewed. Device scenarios and numerical occupancy calculations above are original illustrations, not measured sensor specifications.

Previous: Sensor interface layer Next: Sentence extraction