Does the arc mean the source is certain?

Its realized value is known at decision time in the model. If it is a noisy report, uncertainty about the underlying event remains.

Must the policy use every information parent?

No. It may ignore an available variable if using it does not improve the chosen objective.

Can an earlier decision be an information parent?

Yes. Knowing a previous action can be part of the information available for a later choice.

AI Glossary / Probabilistic machine learning / Intermediate

Information arc in an influence diagram

Definition

An information arc in an influence diagram is an arrow into a decision node showing that the source variable's value is available when that decision is made. It specifies what the decision rule may use, rather than asserting that the information causes a particular choice.

What it means

Influence diagrams represent decisions, uncertain events and utilities in one graphical model. Their arrows have different meanings depending on the type of node they enter.

An arrow into a chance node helps define a probability model. An arrow into a utility node indicates an argument of the utility function. An arrow into a decision node concerns information available before choosing an action.

This distinction prevents an especially important modeling error: allowing a decision to depend on an outcome that will only become known later. A model with that extra knowledge may produce an attractive plan that no real decision-maker could follow.

How it works

A decision policy maps its available information to an action. Information parents define the variables on which that mapping may depend. If a forecast is available before choosing a delivery plan, an arc from forecast to decision permits different plans for different forecasts.

The actual future condition may remain unobserved, so it should not be included as known decision information merely because it influences the eventual outcome. The probability model can still connect that condition to the forecast and utility.

Sequential diagrams also require a clear convention about remembered observations and previous decisions. Some representations assume perfect recall or omit implied information arcs. The diagram's conventions must therefore be read alongside its timing and policy definitions.

Three practical examples

A scheduling model shows that a confirmed staff-availability report is known before assigning a shift.
A maintenance planner distinguishes an observed inspection result from an unobserved internal equipment condition.
An inventory study compares an order decision made before a demand signal with a decision allowed to observe that signal first.

Think of it this way

Think of the papers on a manager's desk at the moment a decision is due. An information arc identifies a paper the manager can consult. It does not require the manager to use it or establish that its contents are perfectly accurate about the world.

The analogy has a boundary: the model knows the value of the observation variable, such as a reported forecast, while uncertainty about the event being forecast can remain.

Worked example

The value of knowing the state before choosing

Consider a fictional equipment decision with two possible states: normal with probability 0.6 and stressed with probability 0.4. Action R yields utility 10 in the normal state and −10 in the stressed state. Action S yields utility 4 in either state.

Without observing the state, expected utility for R is 0.6 × 10 + 0.4 × (−10) = 2. Action S gives 4, so S is optimal among these choices.

If perfect state information is available before acting, the policy can choose R when normal and S when stressed. Expected utility becomes 0.6 × 10 + 0.4 × 4 = 7.6. The gross value of perfect information is 7.6 − 4 = 3.6 utility units.

Adding the state-to-decision information arc represents that changed availability. It does not change the state probabilities or utilities. If obtaining the observation costs 1 utility unit, the net value in this example is 2.6. A noisy inspection would require its own observation node and conditional probabilities; it cannot be represented honestly as perfect knowledge of the true state.

Illustrative case study

Fictional case: a dispatch dashboard

A fictional logistics team builds a diagram for choosing a dispatch plan. Its first version lets the plan depend on actual next-day loading delays. Review reveals that only a forecast is available when dispatch commitments are made.

The team replaces the information assumption with an observed forecast variable and models its relationship to actual delays. It then evaluates policies that can genuinely be implemented at the decision time.

The revised diagram may show a lower expected utility than the unrealistic first version. That is a correction to the information assumption, not evidence that the operational team became less capable. The case contains no actual shipment results.

Common misconceptions and mistakes

An information arc is not automatically a causal arrow. It says that a value can be used by the decision policy.

Availability is not the same as usefulness. A variable can be observed but fail to change the optimal action. Conversely, an informative signal can matter greatly even when it is imperfect about the underlying state.

Omitting an arc requires attention to the diagram's conventions. Some notations include implied memory rather than drawing every earlier observation into every later decision. A reader should not infer forgetting without checking the representation. The model must still prevent access to genuinely unavailable future information.

Business use

Use information arcs to make decision timing and data availability explicit. Ask which report, sensor reading or approval is actually accessible when each action must be selected. Distinguish the reported value from the hidden condition it describes.

This is particularly useful when an AI system recommends actions from forecasts. A model evaluation can accidentally use information that only appears after the decision, creating optimistic results. Reviewing the information parents helps expose that error. The diagram can also support a disciplined comparison of whether acquiring an additional observation is worth its cost.

Check your understanding

A model observes an inspection report before choosing an action, but the equipment's true condition is hidden. Which variable should have the direct information arc into the decision?

Answer: The report. The true condition can influence the report and outcomes through the probability and utility model, but it should not be represented as directly observed unless that is actually the assumption.

Related terms

Decision network
Chance nodes
Utility
Influence diagram
value of information
decision policy
chance node
perfect recall

Learn next

Draw separate nodes for a hidden condition and its noisy measurement. Then compare the expected value of a policy with no measurement, an imperfect measurement and perfect information.

Sources

Author-written technical reference: Poole and Mackworth, Artificial Intelligence: Foundations of Computational Agents, Decision Networks: https://www.cs.ubc.ca/~poole/aibook/html1e/ArtInt_219.html

The equipment utilities and dispatch-dashboard case are original teaching examples. The calculated information value is conditional on those invented assumptions.

Previous: InfoNCE lower bound Next: Information bottleneck method