Can a utility node use a formula?

Yes. A formula can replace a table when it accurately represents the desired mapping.

Can there be several utility nodes?

Yes, with an explicit and justified rule for combining their contributions.

Does the node choose an action?

No. It supplies values used by the decision procedure to compare actions or policies.

AI Glossary / Graphs and symbolic AI / Advanced

Utility nodes

Definition

Utility nodes are components of a decision network, or influence diagram, that assign numerical desirability values to combinations of their parent variables. Their parents may include decisions and uncertain outcomes. They represent how consequences are valued, while chance nodes describe uncertainty and decision nodes represent choices.

What it means

A probabilistic model can predict what might happen without specifying whether those outcomes are desirable. A utility node adds the preference model needed to compare possible decisions. Its output may account for service, resources, and inconvenience.
The node typically contains a table or formula rather than a probability distribution. Entries may be negative and do not need to sum to one. A utility of 80 does not mean an outcome is 80% likely.

How it works

A utility node receives the values of its parents and returns the corresponding utility. In common diagram conventions it is drawn as a diamond. An incoming edge identifies an argument of its utility function; it does not by itself describe a causal mechanism.
To evaluate a decision, the solver combines the utility values with the relevant outcome probabilities and sums over uncertainty. If actions affect those probabilities, their effects must be represented in the model. Evidence available before a decision can change the probability estimates and therefore the preferred action.
Some models use several utility nodes for separate contributions. Their combination must be specified, often as an additive total when that representation is justified. Summing overlapping consequences or unrelated scales can produce an inappropriate objective even if the numerical calculation is correct.

Three practical examples

An event planner combines uncertainty about an external power outage with a decision to arrange backup power. A utility node scores the resulting service and resource consequences for each combination.
A maintenance system predicts equipment failure and compares intervention options. Its utility model evaluates downtime, maintenance effort, and disrupted operations rather than assuming that minimizing failure probability alone captures the business objective.
A scheduling assistant models uncertain travel time and a choice of appointment. A utility node represents the user's preferences over the resulting schedule, including missed commitments where relevant.

Think of it this way

A utility node resembles a scoring desk beside a planning board. Other parts of the board describe choices and uncertain events. The desk assigns a value once their relevant values are specified. The decision procedure then averages those scores using probabilities; the scoring desk itself does not forecast the events.

Worked example

Consider a fictional event with an external power-outage chance of 0.15. Decision G is whether to arrange backup power. Chance variable O records whether the external outage occurs. Both point to a utility node U(G,O). Arranging backup does not change the probability of the external outage; it changes the consequences in this toy model.
Without backup, utility is 100 if no outage occurs and -60 if one occurs. With backup, utility is 88 without an outage and 78 with an outage. These are illustrative preference values that already include the backup's resource burden. They are not probabilities or monetary prices.
Expected utility without backup is 0.85 × 100 + 0.15 × (-60) = 76. With backup it is 0.85 × 88 + 0.15 × 78 = 86.5. The model prefers arranging backup.
If updated evidence lowers the outage probability to 0.03, the same utility table gives 95.2 without backup and 87.7 with backup. The preferred decision changes even though the utility node's values remain unchanged.
For outage probability p, the scores are 100-160p and 88-10p. They tie at p=0.08, both scoring 87.2. The assumptions exclude backup malfunction; representing that possibility would require additional outcome detail and probabilities, rather than treating the table as a guarantee of service.

Illustrative case study

In a fictional events company, two analysts build separate utility nodes for service quality and backup resources. For the worked-example assumptions, the service contribution is 100 without an external outage. During an outage it is -60 without backup and 90 with backup. The resource contribution is -12 when backup is arranged and zero otherwise.
Adding these contributions reproduces the four values in the single-node table: 100, -60, 88, and 78. Reviewers then discover that another proposed penalty would count the same backup burden twice. They remove that duplication before comparing decisions.
Reviewers can distinguish preference contributions from the estimated outage probability. It also shows why adding more utility nodes is not automatically an improvement. Each contribution needs a justified relationship to the total.

Common misconceptions and mistakes

“Utility nodes contain probability tables.” They contain desirability values or functions; chance nodes carry probability information.
“An edge into utility means a physical causal effect.” It indicates that the utility calculation depends on that parent.
“Maximum expected utility guarantees the best realized outcome.” An uncertain event can still produce an unfavorable result.
“Separate utility nodes can always be summed.” The scales, decomposition, and treatment of interactions must support the sum.
“Changing a forecast requires rewriting the utility table.” Often only the probability model changes, as the worked example shows.

Business use

Utility nodes make the objectives of a decision-support model inspectable. Review their parent variables and ask whether important consequences are missing or counted more than once. Check whose preferences the numbers represent and how those values were obtained.
Keep probability estimation separate from preference assessment so each can be challenged on its own evidence. Test decisions around thresholds where modest input changes reverse the recommendation. Ensure the model uses only information available when the decision is made. If a requirement is mandatory, represent it appropriately rather than assuming that a finite penalty will always prevent its violation under every combination of other scores.

Check your understanding

Why does the original example choose backup?

Answer: Its expected utility is 86.5, exceeding 76.

What changes when outage probability falls to 0.03?

Answer: The expectation and preferred decision change; the utility table stays fixed.

Why is the extra backup penalty inappropriate in the case study?

Answer: It repeats a consequence already included in the resource contribution.

Related terms

Decision network
Chance nodes
Decision nodes
Utility function
Expected utility
influence diagram
chance node
decision node
value of information
multiattribute utility

Learn next

Build a small decision network and distinguish utility-function inputs from information available before a decision. Then examine how an additional observation can change the selected action.

Sources

University of California, Berkeley, CS 188, Decision Networks: https://inst.eecs.berkeley.edu/~cs188/textbook/vpis/decision-networks.html Russell and Norvig, AIMA, Rational Decisions lecture slides: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter16.pdf Poole and Mackworth, Preferences and Utility: https://artint.info/3e/html/ArtInt3e.Ch12.S1.html

Previous: Utility independence Next: Utility theory