Is utility theory the whole of decision theory?

No. Decision-making under uncertainty also requires beliefs, available actions, and information about consequences.

Does risk aversion violate expected utility?

No. Suitable utility functions can represent risk-averse preferences.

Can a model be useful without perfectly describing behavior?

Yes, if its purpose and assumptions are clear and appropriate to the task.

AI Glossary / Graphs and symbolic AI / Advanced

Utility theory

Definition

Utility theory studies how preferences can be represented numerically and used to reason about choices. In AI, it supplies the preference component of decision theory: utilities describe desirability, while probabilities describe uncertainty. Expected-utility theory connects suitable consistency assumptions about preferences to evaluating uncertain alternatives by their probability-weighted utilities.

What it means

An agent can predict possible outcomes yet still need a basis for choosing among them. Utility theory asks what preference structure would make its choices coherent and how that structure can be represented. The numbers express a particular decision-maker's preferences.
Distinguish two questions. A descriptive question asks how people actually choose. A normative question asks what follows if certain principles of consistent choice are accepted. Expected-utility theory is useful as a normative framework even though actual behavior can depart from its assumptions. Calling a model rational in this technical sense does not establish that its objectives are ethical, complete, or appropriate.

How it works

In a standard finite-outcome setting, the theory considers both outcomes and lotteries over them. Relevant consistency conditions include the ability to compare alternatives, transitive preferences, and continuity. Independence requires a preference comparison to remain consistent when both alternatives are mixed with the same additional lottery in the same proportion.
Under the appropriate expected-utility axioms, preferences admit a representation using a utility function: an uncertain alternative is evaluated by summing each outcome's utility multiplied by its probability. The representation is unchanged in its choices by positive affine changes of utility scale.
These are assumptions with consequences, not facts that every person's choices must satisfy. A modeler must also specify outcomes carefully: consequences that the decision-maker cares about cannot be omitted and then recovered merely by applying the theorem.

Three practical examples

An AI planning system compares strategies that can lead to several service outcomes. Utility theory helps separate beliefs about their likelihood from preferences about the outcomes, supporting review.
A preference interview asks a manager to compare a certain intermediate result with an uncertain best-or-worst result. The answer supplies information about uncertain choices that a simple ranking of the three certain results would not provide.
A scheduling assistant encounters apparently inconsistent user choices. Reviewing the outcome descriptions may reveal an omitted consequence, such as who must rearrange childcare. The team investigates the missing context.

Think of it this way

Utility theory resembles a set of consistency rules for a decision-maker's comparisons. The rules do not decide what the person should care about. They show how certain accepted comparisons constrain other comparisons, allowing a numerical representation to summarize the resulting preference structure.

Worked example

Consider three complete, fictional service outcomes: Best, Middle, and Worst. The decision-maker ranks them in that order and is indifferent between Middle for certain and a lottery with a 60% chance of Best and a 40% chance of Worst.
Normalize U(Best)=1 and U(Worst)=0. Under the expected-utility assumptions, that indifference sets U(Middle)=0.6. The value 0.6 records an elicited preference; it is not the probability that a middle-quality result will occur.
Option A gives a 50% chance of Best and a 50% chance of Middle. Its expected utility is 0.5 × 1 + 0.5 × 0.6 = 0.8. Option B gives a 70% chance of Best and a 30% chance of Worst, so its expected utility is 0.7. The model prefers A.
The same conclusion follows by replacing Middle in A with its indifferent lottery. Best then occurs with probability 0.5 + 0.5 × 0.6 = 0.8, and Worst with probability 0.5 × 0.4 = 0.2. This reduction illustrates the consistency behind the calculation.
If the elicited value of Middle were instead 0.2, A would score 0.6 and B would remain at 0.7. The certain ranking Best above Middle above Worst could stay unchanged while the preferred uncertain option reverses. Utility spacing requires more information than that ranking alone.

Illustrative case study

In a fictional fulfillment business, an analyst asks managers to rate three delivery outcomes and then averages the ratings across uncertain plans. Reviewers discover that the ratings were intended only as an ordering, with no discussion of risk or intermediate consequences.
The team replaces this shortcut with explicit comparisons between certain outcomes and lotteries. It describes each outcome consistently, including delay, damage, and customer effort. Where responses conflict, the team checks whether scenarios were misunderstood or whether the assumed preference rules are unsuitable.
The resulting model documents its assumptions and unresolved disagreements. It supports a transparent comparison of plans rather than presenting arbitrary ratings as a theorem-backed decision rule. The business also keeps mandatory operating requirements separate from preferences that may legitimately be traded off.

Common misconceptions and mistakes

“Utility theory says everyone maximizes money.” Preferences can include nonfinancial consequences, and utility need not be linear in money.
“A utility theorem proves the objective is good.” It concerns representability and consistency under assumptions, not moral approval.
“Certain rankings determine every risky choice.” The worked example shows otherwise.
“Observed human choices always follow expected utility.” Descriptive deviations motivate other models and closer examination of context.
“The independence axiom means attributes are statistically independent.” It is a consistency rule for mixtures of alternatives, distinct from probability independence and multiattribute utility independence.

Business use

Utility theory helps make the logic of an automated decision explicit. Start with the decisions, complete outcome descriptions, available information, and stakeholder perspective. Then establish whether the assumed preference framework is suitable before estimating numerical values.
Use understandable comparisons and retain uncertainty about answers. Test whether plausible changes in utilities or outcome probabilities alter the recommendation. A calculation can be perfectly correct while depending on preferences that were never actually agreed. Document both the model's normative assumptions and any evidence about observed choices so reviewers can see what the system is intended to prescribe and what it merely predicts.

Check your understanding

Why is Middle assigned utility 0.6?

Answer: The decision-maker is indifferent to the stated 60/40 best-or-worst lottery.

Why does A score 0.8?

Answer: Half its probability is on utility 1 and half on utility 0.6.

What does the theorem fail to establish?

Answer: That the specified objectives include every relevant concern or are ethically appropriate.

Related terms

Learn next

Examine the expected-utility axioms through concrete lottery comparisons. Then compare normative decision models with descriptive accounts of actual human choice.

Expected utility

Sources

Poole and Mackworth, Preferences and Utility: https://artint.info/3e/html/ArtInt3e.Ch12.S1.html Russell and Norvig, AIMA, Rational Decisions lecture slides: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter16.pdf AIMA glossary, utility theory: https://github.com/aimacode/aima-glossary

Previous: Utility nodes Next: Utterance