Is a decision network the same as an influence diagram?

The terms are often used interchangeably, although particular formalisms and software may impose different details.

Can a network include several utility nodes?

Yes, but their combination must be defined consistently, such as through an explicitly justified total-utility function.

Does the optimal policy always choose one fixed action?

No. It may choose different actions for different observations available at the relevant decision point.

AI Glossary / Graphs and symbolic AI / Advanced

Decision network

Definition

A decision network is a graphical model that combines uncertain variables, available decisions and utilities to support choosing actions under uncertainty. Often called an influence diagram, it extends a Bayesian-network-style representation with explicit choices and outcome preferences.

What it means

A Bayesian network can help predict what is likely. A decision network additionally asks what to do, given those predictions and the consequences of the available actions.

Three roles are kept separate: chance nodes represent uncertain quantities, decision nodes represent controllable choices, and utility nodes represent how outcomes are valued. Their relationships show which information and consequences matter to the decision.

The model is not just a picture. Chance variables need probability specifications, decisions need alternatives and information assumptions, and utilities need a defined preference scale. A graph without those quantities may help frame a discussion, but it cannot by itself calculate the preferred action.

How it works

Construct the graph around a precise decision. Connect chance variables according to the probabilistic dependencies represented by the model. Connect decisions to the consequences they can influence, and connect relevant quantities to utility nodes.

In a standard influence diagram, an arrow into a decision node identifies information available when that decision is taken. This differs from an arrow into a chance node, which helps specify its conditional probability model. Accidentally supplying a decision with future information can make its policy appear better than any implementable one.

For a simple single decision, evaluate the expected utility of each available action using the evidence known at decision time. Choose the action with the greatest expected utility under the model. With several decisions, the solution is a strategy: a collection of policies that specify actions for the information situations that can arise.

Sequential evaluation must respect timing and information access. Particular diagram classes and solution algorithms make additional assumptions, so an optimization claim should identify the model and method rather than implying every decision-network problem is equally easy.

Three practical examples

A maintenance team models uncertain equipment condition, a repair choice and the consequences of repair or continued operation.
A planning group compares whether obtaining an additional inspection before acting could improve its decision.
An AI workflow separates a predicted problem probability from the action chosen after accounting for review effort and error consequences.

Think of it this way

Think of a map with three kinds of landmarks: things that may happen, choices you can make and outcomes you care about. The connecting lines explain how these roles interact.

Unlike an ordinary route map, different arrows have different meanings. Seeing information before a choice is not the same relationship as causing an outcome after the choice.

Worked example

A small maintenance network

Let Fault be a chance variable with probability 0.20. Let Action be Repair or Continue. Let the utility be the negative of a toy loss measure, so lower loss means higher utility.

Assume Repair always removes the fault and incurs loss 3. Continue incurs loss 20 if a fault exists and 0 otherwise. With no inspection information, Repair has expected utility -3. Continue has expected utility -(0.20×20+0.80×0)=-4. Repair is preferred.

The network needs the fault uncertainty, the action alternatives and a utility function depending on Fault and Action. It must not give the action node the actual Fault value unless the decision maker observes it.

If a perfect, free inspection becomes available before choosing, the policy can repair when Fault=true and continue otherwise. Expected loss is 0.20×3+0.80×0=0.6, giving utility -0.6. The improvement is 2.4 utility units. This different result comes from a changed information structure, not a changed prior fault probability.

Illustrative case study

A fictional machine-service company, Maple Maintenance, builds a decision network for an inspection workflow. Its first diagram includes an information arrow from the actual hidden condition directly into the repair decision.

A reviewer notices that technicians only see a noisy indicator, not the condition itself. The team replaces that information assumption with an observed indicator node and supplies a model of its reliability. The resulting policy can now depend only on the evidence available in practice.

The company also reviews the utility table with operations staff. A technically correct probability model would still recommend unsuitable actions if downtime or unsuccessful repair consequences were omitted. The diagram becomes a shared record of those assumptions rather than a claim that the software knows the right business objective automatically.

Common misconceptions and mistakes

A decision network is just a Bayesian network with different labels. Decisions and utilities have different semantics from chance variables.
More arrows always improve realism. An incorrect information arrow can give the policy knowledge it will never have.
The most probable outcome determines the best action. Less likely outcomes can matter greatly when their consequences differ.
Utility values are probabilities. They express preferences or consequences and do not need to sum to one.

Business use

Decision networks are useful when an AI prediction must feed into a choice with asymmetric consequences. They make the assumptions connecting evidence, actions and outcomes visible.

Document each decision's timing, allowed actions and available information. Validate probability inputs separately from preference inputs, and examine sensitivity near action-switching thresholds. If the network was fitted to observational records, do not assume that changing an action has the same effect as merely conditioning on a historically correlated variable; the action model needs a justified interpretation.

Check your understanding

A repair policy uses the true equipment condition, but only a noisy sensor reading is available before the choice. What is wrong with that model?

Answer: It grants unavailable information to the decision. The policy must depend on the observed reading and the uncertainty remaining about the condition.

Related terms

Bayesian network
Chance nodes
Decision nodes
Utility nodes
Expected utility
Influence diagram
chance node
decision node
utility node
value of information

Learn next

Draw the Fault - Action - Utility example and add a noisy indicator observed before Action. Specify the indicator's false-positive and false-negative behavior, then identify which probability calculations must change before comparing the action policies.

Sources

Russell and Norvig, AIMA Chapter 16 materials, introduces decision networks and expected-utility evaluation: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter16.pdf

Ali Jenzarli, Information/Relevance Influence Diagrams, clarifies informational arrows into decisions: https://arxiv.org/abs/1302.4963

The maintenance numbers and Maple Maintenance case are original illustrations.

Previous: Decision lists Next: Decision nodes