Is the highest-utility outcome always the best action?

No. An action's full distribution of possible outcomes determines its expected utility.

Can utility values be negative?

Yes. Their origin can be shifted without changing choices, provided the transformation preserves the expected-utility scale.

Does maximizing expected utility ensure the preferred outcome happens?

No. It selects according to a probabilistic criterion; an unfavorable outcome can still occur.

AI Glossary / Graphs and symbolic AI / Intermediate

Expected utility

Definition

Expected utility is the probability-weighted average of the utility assigned to possible outcomes. Utility represents how desirable an outcome is according to a specified preference model. Expected utility combines those preferences with uncertainty, allowing an agent to compare actions whose consequences are not known in advance.

What it means

Probability answers how likely an outcome is. Utility answers how desirable it is. Expected utility keeps these roles separate: a highly desirable outcome may have little influence if it is very unlikely, while an undesirable but probable outcome may strongly affect the comparison.

Utility need not equal money. It can represent preferences concerning time, quality, comfort, or combinations of consequences, provided the scale has the interpretation required by the decision model.

A simple ranking of outcomes is not sufficient for calculating expected utility. Numerical spacing matters when outcomes are uncertain. Assigning ranks 1, 2, and 3 arbitrarily and averaging them can imply preferences over lotteries that the decision-maker never intended.

How it works

Specify the available actions and the relevant possible outcomes. For each action, estimate probabilities conditional on that action and the information available. Assign utility values using the chosen preference model.

Multiply each outcome's utility by its probability and add the products. An expected-utility-maximizing agent selects an available action with the highest resulting total, subject to the model's constraints.

Classical expected utility theory explains when preferences over uncertain outcomes can be represented this way. Its consistency conditions are assumptions about the preference structure, not evidence that every person always behaves according to the model.

Utilities can be rescaled by multiplying by a positive constant and adding a constant without changing expected-utility rankings. Arbitrary nonlinear transformations generally do not preserve those rankings. This distinguishes a utility scale for risky choices from a scale that merely orders certain outcomes.

Three practical examples

1. A planning agent compares routes whose travel times and comfort levels vary, using an explicitly defined preference model for the possible outcomes.

2. A service team compares delivery options that differ in reliability and speed, after deciding how those consequences should be valued.

3. A simulated agent evaluates actions that can lead to success, partial completion, or failure, with probabilities and utilities defined separately for each outcome.

Think of it this way

Think of evaluating meals using both how likely each dish is to arrive and how much you would value receiving it. The favorite dish does not automatically make an uncertain option best if it is rarely available.

The analogy has limits. Informal satisfaction scores may lack the numerical interpretation needed for expected utility. The model requires a meaningful preference scale, not simply attractive labels or an arbitrary collection of ratings.

Worked example

Consider a toy decision over nonnegative payoff amounts. Option A gives 36 units for certain. Option B gives 100 units with probability 0.5 and 0 with probability 0.5.

The expected payoff of A is 36, while B's is 50. If utility is linear in the payoff, B has the larger expected utility.

Now stipulate a different preference model: U(x) = the square root of x. Under this model, A has expected utility √36 = 6. Option B has expected utility 0.5 × √100 + 0.5 × √0 = 5. Thus A is preferred under the specified nonlinear utility function.

Notice that applying utility to B's expected payoff gives √50, approximately 7.071. That is not B's expected utility. In general, the utility of an average outcome differs from the average of outcome utilities.

Finally, rescale utilities as U'(x) = 2U(x) + 3. A's expected utility becomes 15 and B's becomes 13, preserving the ranking. The square-root function is an invented teaching assumption, not a claim about anyone's measured preferences or an appropriate utility for a particular real financial decision.

Illustrative case study

Fictional case: Maple Support uses an automated routing model with a numerical score for each possible service outcome. During review, staff discover that the numbers were simply ordinal labels for poor, acceptable, and excellent outcomes.

The team had been averaging those labels as though their spacing expressed agreed trade-offs under uncertainty. It revisits the preference model and separates probability estimation from outcome valuation. The case illustrates that precise arithmetic cannot repair an unjustified interpretation of the input utility scale.

Common misconceptions and mistakes

Do not equate expected utility with expected monetary value. They coincide in ranking monetary lotteries when the relevant utility is positively affine in money, but not in general.

Do not calculate utility only after averaging outcomes unless the utility function permits that simplification.

Avoid treating utilities as objective measures of moral worth or universal happiness. They encode a specified preference model for a decision context.

Do not assume probabilities are preferences. Changing an outcome's desirability should not silently change the estimated likelihood that it occurs, and new evidence can change probabilities without changing utility values.

Business use

Expected utility can provide a coherent structure for comparing uncertain business actions when their consequences and trade-offs are clearly defined. It makes explicit which outcomes matter and how uncertainty enters the choice.

Document whose preferences the utility function represents, how it was established, and which consequences it includes. Check sensitivity to uncertain probabilities and disputed trade-offs. Requirements that must always hold should remain explicit constraints rather than being accidentally traded away through a poorly chosen score.

Check your understanding

In the worked example, why is √50 not the expected utility of option B?

Answer: Because 50 is its expected payoff. Expected utility averages the utilities of the actual possible outcomes, giving 0.5 × 10 + 0.5 × 0 = 5.

Related terms

Learn next

Review probability-weighted averages, then compare linear and nonlinear utility functions on the same outcome distribution. Study positive affine transformations and certainty equivalents to understand which numerical changes preserve preferences and which alter them.

Sources

Primary sources: Russell and Norvig, rational decisions and utility scales: https://aima.cs.berkeley.edu/slides-pdf/chapter16.pdf

Poole and Mackworth, Preferences and Utility: https://artint.info/3e/html/ArtInt3e.Ch12.S1.html

The payoff comparison, transformation calculation, and fictional case are original instructional examples. The utility function is stipulated rather than empirically estimated.

Previous: Expected SARSA Next: Expected value