Why are decision nodes often drawn as rectangles?

That is a common influence-diagram convention distinguishing choices from chance and utility nodes. The model's semantics matter more than the drawing shape.

Can a decision node have a probability table?

A stochastic policy can be represented that way, but its entries specify chosen action probabilities conditional on available information.

Does selecting an action guarantee its intended outcome?

No. The action may influence chance variables whose outcomes remain uncertain.

AI Glossary / Graphs and symbolic AI / Intermediate

Decision nodes

Definition

Decision nodes represent choices that an agent or decision maker can control in a decision model. In decision networks and influence diagrams, they specify available actions and the information on which those actions may depend, rather than uncertain events to be predicted.

What it means

A delivery delay may be uncertain, while choosing a delivery method is an action. Representing both as ordinary random variables can obscure the difference between observing what happened and selecting what will be done.

A decision node makes the controllable choice explicit. Its possible values are alternatives such as “standard route,” “reserve capacity” or “postpone.” The model evaluates those alternatives through their effects on uncertain outcomes and utility.

The term also appears in decision trees. In a decision-analysis tree, a decision node is a choice point. In a machine-learning tree, an internal node usually tests a feature rather than choosing an action. These meanings should be distinguished when discussing a diagram.

How it works

Define the decision's alternatives, timing and decision maker. Identify what information is available before the choice, including earlier observations and decisions that can be remembered. In a standard influence diagram, incoming informational arcs indicate that context.

A policy maps each available information situation to an action. If there are no observed parents, the policy may be one fixed action. If a signal is observed first, the policy can specify a different action for each signal value. A stochastic policy can assign action probabilities, but those probabilities describe a choice mechanism, not an uncontrolled natural event.

During evaluation, uncertain outcomes are averaged using their probabilities, while controllable alternatives are compared according to the chosen objective, commonly maximum expected utility. Confusing these operations can produce the wrong answer: averaging equally across actions assumes random choice rather than optimizing.

The chosen action must remain feasible. Timing, authority and resource constraints need to be represented in the model or its admissible policy set. Adding a decision node cannot give an agent practical control over something it merely observes.

Three practical examples

A dispatch model represents the chosen service level as a decision node and traffic conditions as chance variables.
An inspection workflow represents whether to collect an additional observation before selecting a later action.
A software agent uses a policy to choose automatic handling or human review from the evidence available at that moment.

Think of it this way

Think of a control panel with buttons you can actually press. The weather display beside it reports something uncertain; it is not another button. A policy explains which button to press for the displayed information.

The analogy stops short of assuming every displayed fact is available in time or perfectly accurate. The model must distinguish the true condition from the signal the decision maker can see.

Worked example

Choosing rather than averaging actions

Suppose a toy dispatch decision has three alternatives. Under normal conditions, their utilities are Standard=6, Reserve=3 and Emergency=0. Under disruption, their utilities are Standard=-2, Reserve=3 and Emergency=4. Assume disruption probability 0.25 and no advance signal.

Expected utilities are Standard: 0.75×6+0.25×(-2)=4; Reserve: 3; Emergency: 0.75×0+0.25×4=1. The decision node therefore selects Standard under this expected-utility model.

Averaging the three expected utilities gives (4+3+1)/3=8/3, but that is the value of an equal-probability random policy, not the optimal choice.

Now assume the condition is revealed perfectly before the choice. The policy becomes Standard when normal and Emergency when disrupted. Its expected utility is 0.75×6+0.25×4=5.5. If the revelation arrives after dispatch, that policy is unavailable. The same alternatives and utility table lead to different feasible policies depending on the information timing.

Illustrative case study

A fictional delivery coordinator, Alder Dispatch, creates a model with a “traffic state” parent for its service-choice node. The implementation silently uses the actual future traffic state from a simulation.

The team realizes that dispatchers only have a forecast when choosing. It replaces the unavailable information with the forecast and models the uncertainty between forecast and actual conditions. The policy table now maps observed forecasts to actions.

It also distinguishes a recommendation from an authorized operational change. A human review decision remains a separate step where required by the workflow. The model's action alternatives are aligned with what the executing system can actually choose, so policy evaluation no longer assumes powers or information it lacks.

Common misconceptions and mistakes

A decision node is simply a chance node with a convenient name. A decision represents control and is evaluated through a policy.
Every incoming arrow describes a cause of the decision. In standard influence diagrams it indicates information available for the choice.
A decision node always selects one action for all circumstances. A policy can vary with the observed context.
The model can use any variable that exists in the database. The variable must be available at the decision time and permitted in the policy.

Business use

Decision nodes are a useful design tool for AI systems because they separate prediction from responsibility for action. They force a team to state which choices the system can make and what evidence it may use.

For each node, record the alternatives, timing, available observations, authority and objective. Check the policy on situations where signals disagree, inputs are missing or an action becomes unavailable. When evaluating past records, exclude information learned only after the original decision; otherwise the apparent policy quality can depend on hindsight.

Check your understanding

A diagram lets the first decision depend on a test result produced only after that decision. Is the policy implementable as drawn?

Answer: No. The information timing is inconsistent. The test must occur earlier, or the first policy must be evaluated without that result.

Related terms

Decision network
Policy
Chance nodes
Utility nodes
Information sets
Expected utility
Action
influence diagram
chance node
utility node
information set

Learn next

Take the dispatch table and replace perfect advance information with a noisy forecast. Write a policy table whose rows are forecast values, not actual future states. Identify which conditional probabilities are needed to compare the actions in each row.

Policy

Sources

Russell and Norvig, AIMA materials on action nodes and expected-utility choice: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter16.pdf

Ali Jenzarli, Information/Relevance Influence Diagrams, explains the informational meaning of arcs into standard decision nodes: https://arxiv.org/abs/1302.4963

The dispatch utilities and Alder Dispatch case are original teaching illustrations.

Previous: Decision network Next: Decision region