Must every variable be hidden?

No. A DBN can include observed and unobserved variables.

Does dynamic mean the graph changes at every time step?

No. A fixed repeated template is common.

Are hidden Markov models related?

Yes. They can be represented as a simpler special case within the broader DBN framework.

AI Glossary / Statistics and probability / Intermediate

Dynamic Bayesian network

Definition

A dynamic Bayesian network is a probabilistic graphical model for variables that evolve across a sequence, commonly over time. It represents conditional dependencies both within a time step and between steps, allowing uncertainty about hidden states to be updated as observations arrive. The model is often abbreviated DBN. It extends Bayesian-network reasoning to sequential processes and can represent structures more elaborate than a single hidden-state chain. Here dynamic refers to the modeled process; it does not require the network structure itself to change at every step.

What it means

Many systems have important states that cannot be observed directly. A machine's condition, for example, may influence several noisy sensor readings and evolve from one period to the next. A DBN combines a model of that evolution with a model of the observations. Instead of deciding that one reading reveals the truth, it maintains probabilities over possible states. Factoring the state into related variables can express structure that would be cumbersome in one enormous state variable, although inference can still be computationally difficult.

How it works

Specify an initial distribution, a transition model and conditional distributions for observations and any other variables. A common first-order construction uses a two-slice template: the next state depends on the preceding state, with dependencies within each slice arranged acyclically. Unroll the template over the sequence to obtain an ordinary directed acyclic network. Filtering estimates the current state from observations available so far; prediction looks ahead, while smoothing uses later observations to revise beliefs about earlier states. Exact or approximate inference depends on the graph and distributions. Parameters and sometimes structure may be learned from sequential data, but the chosen dependencies remain modeling assumptions.

Three practical examples

Equipment monitoring: combine evolving component conditions with several imperfect sensor measurements.
Activity recognition: use sequential observations to infer an underlying activity that persists or changes over time.
Service operations: track uncertain workflow states from incomplete event records while distinguishing observation errors from actual transitions.

Think of it this way

Think of maintaining a running set of explanations for a changing situation. Each new time step first changes what you expect, then new evidence changes what you believe. The model supplies explicit probability rules for both changes instead of treating every observation as an isolated verdict.

Worked example

Consider a hypothetical two-state machine model with normal and fault states. Its previous fault probability is 0.20. A fault persists to the next step with probability 0.80, while a normal machine develops a fault with probability 0.10. Before seeing the next sensor reading, the fault probability is 0.20 x 0.80 + 0.80 x 0.10 = 0.24. The alarm has probability 0.90 given fault and 0.05 given normal. After an alarm, Bayes' rule gives 0.24 x 0.90/[0.24 x 0.90 + 0.76 x 0.05] = 0.216/0.254, approximately 0.8504. Predicting one more step without a new reading gives 0.8504 x 0.80 + 0.1496 x 0.10, approximately 0.6953. This simple hidden Markov model is a special case of a DBN, not an illustration of its full multivariable expressive power.

Illustrative case study

Fictional case: Birch Maintenance builds a DBN linking equipment condition, operating load and noisy observations across inspection periods. Engineers review the proposed dependencies and compare predicted probabilities with later verified outcomes. An alarm updates the probability of a fault, while an intervention decision also considers inspection costs and failure consequences. The team distinguishes retrospective smoothed estimates from probabilities available at the time of the decision. This invented case illustrates sequential reasoning and avoids claiming that the model establishes a mechanical cause or guarantees reliable maintenance decisions.

Common misconceptions and mistakes

Do not confuse a dynamic Bayesian network with a deep belief network, which can share the abbreviation DBN. A directed edge does not automatically establish causation. A first-order assumption means the modeled state carries the needed past information; it does not mean real systems have no longer memory. Missing observations can be handled probabilistically under a specified model, but informative missingness may require additional variables. Reusing future observations in a real-time evaluation creates leakage. A fixed transition model can fail when operating conditions change.

Business use

Businesses can use DBNs when decisions depend on uncertain evolving states and observations have known limitations. The explicit structure supports conversations between domain specialists and modelers about persistence, sensor reliability and dependencies. Evaluate probability calibration, decision performance and computational cost on representative sequences. Document which information was available at each decision time and monitor changes in data collection. A larger graph is not automatically better: additional dependencies require data, parameter estimation and validation, and may make inference much more expensive.

Check your understanding

In the worked example, why is the fault probability before the alarm 0.24 rather than 0.20?

Answer: The transition model accounts for both persistent faults and newly developing faults before observation evidence is incorporated.

Related terms

Learn next

compare filtering and smoothing on the same short sequence, then represent a hidden state using several interacting variables and examine how that changes inference requirements.

Sources

Sources: Murphy, Dynamic Bayesian Networks: Representation, Inference and Learning, author thesis page: https://www.cs.ubc.ca/~murphyk/Thesis/thesis.html ; Murphy, Dynamic Bayesian Networks, representation and inference chapter: https://www.cs.ubc.ca/~murphyk/Papers/dbnchapter.pdf ; Friedman, Murphy and Russell, Learning the Structure of Dynamic Probabilistic Networks: https://arxiv.org/abs/1301.7374 ; AIMA glossary source: https://github.com/aimacode/aima-glossary/blob/master/glossary.md .

Previous: Dynamic backtracking Next: Dynamic Conditional Correlation