Is this just adding weighted scores?

No. Weighted addition is one useful form, but the theory also addresses conditions for that form, interactions, and uncertain outcomes.

Are utility values probabilities?

No. Utilities represent preferences; probabilities represent uncertainty. Expected utility combines the two for a decision.

Must everyone use the same weights?

No. Different people or decision contexts can have different preferences. Combining them is an additional problem that requires an explicit approach.

AI Glossary / Graphs and symbolic AI / Intermediate

Multiattribute utility theory

Definition

Multiattribute utility theory is a framework for representing preferences over outcomes described by several attributes, such as delivery speed, reliability, and cost. It studies how these attributes jointly determine utility and how to compare actions when their outcomes are uncertain. A weighted score is one possible model, but its suitability depends on the preferences and assumptions being represented.

What it means

Many choices cannot be described adequately by one measurement. A faster delivery arrangement may cost more or fail more often. Multiattribute utility theory makes these trade-offs explicit rather than assuming that a single raw metric captures everything that matters.
The attributes describe outcomes; the utility function represents the decision maker's preferences over those outcomes. Probabilities describe uncertainty about which outcome will occur. These are different ingredients. A team may agree on the predicted outcomes yet disagree on their desirability, or share preferences while disagreeing on the evidence. Keeping these disagreements separate makes a decision easier to examine.

How it works

Begin by identifying meaningful attributes and specifying their ranges and direction of improvement. Elicit preferences through comparisons, trade-offs, and, when uncertainty matters, suitable lottery comparisons. Check whether preference or utility independence assumptions justify a simpler mathematical representation.
An additive model combines separate attribute utility contributions. More general models allow interactions: the value of improving one attribute can depend on the level of another. For uncertain actions, evaluate the utility of each possible joint outcome and average using its probability. Do not automatically substitute average attribute values into a nonlinear utility function; that can change the answer.

Three practical examples

A service designer can compare delivery arrangements across predictable arrival time, successful completion, and operating expense.
A personal scheduling assistant can represent a user's trade-offs among travel effort, time with friends, and quiet time, provided those preferences are elicited rather than presumed.
A robot-task planner can balance completion quality and resource consumption while treating non-negotiable operating restrictions as constraints on feasible actions.

Think of it this way

Think of choosing a meal while considering flavor, preparation effort, and variety. A preference model explains how those aspects combine for one diner. The analogy is limited because formal utility under uncertainty requires consistent comparisons of uncertain outcomes, and another diner need not share the same model or trade-offs.

Worked example

Consider an invented delivery-service decision with three normalized attribute utilities: affordability A, speed S, and reliability R, each between 0 and 1. Assume the decision maker has assessed and accepted the additive utility U = 0.40A + 0.35S + 0.25R for this restricted decision.
Option Blue has attribute utilities (0.8, 0.6, 0.9), giving U = 0.32 + 0.21 + 0.225 = 0.755. Option Gold has (0.5, 0.9, 0.95), giving U = 0.20 + 0.315 + 0.2375 = 0.7525. Blue ranks slightly higher under the stated model.
Now suppose another option produces Blue's complete outcome with probability 0.6 and Gold's complete outcome with probability 0.4. Its expected utility is 0.6 × 0.755 + 0.4 × 0.7525 = 0.754. This is an illustrative preference calculation, not a measurement of real services. The tiny difference also makes sensitivity analysis useful: modest changes in assessed trade-offs or outcome estimates could change the ranking.

Illustrative case study

In a fictional equipment-rental business, an AI assistant ranks dispatch plans using price alone. Staff explain that customers also care about arrival windows and equipment readiness. The analyst builds an explicit outcome description and conducts preference discussions with the people responsible for the service promise.
A proposed additive model performs poorly on scenarios where fast arrival is useless without a complete equipment set. The team records this interaction and revises the model rather than forcing the original weights to fit every case. It compares recommendations with held-out preference examples before using them in a supervised planning workflow. No business improvement is claimed from this fictional exercise.

Common misconceptions and mistakes

Weights are not universal measures of importance. Their interpretation depends on attribute scales, ranges, utility functions, and the elicitation method. Rescaling an attribute without adjusting its contribution can silently change the decision rule.
Also distinguish statistical independence from independence of preferences. Two attributes can be statistically correlated while satisfying a particular preference condition, or uncorrelated while interacting strongly in utility. Finally, a higher model score means preferred under that model. It does not prove that all affected people benefit or that omitted considerations are unimportant.

Business use

For an organization, the main benefit is an inspectable account of trade-offs. Decision owners can review which outcomes matter, how uncertainty is represented, and which assumptions support a recommendation. This helps locate disagreement before a scoring system becomes embedded in software.
Keep the model's scope narrow enough to validate. Record attribute definitions, preference owners, ranges, constraints, and sensitivity checks. Revisit the model when the decision context changes. An AI system can help organize comparisons, but automatically inferring enduring human preferences from a few clicks can produce an unjustified sense of precision.

Check your understanding

Under the worked model, what is the utility of attribute utilities (0.7, 0.8, 0.6)?

Answer: 0.40 × 0.7 + 0.35 × 0.8 + 0.25 × 0.6 = 0.28 + 0.28 + 0.15 = 0.71. This result follows from the assumed additive utility model, not from the attribute values alone.

Related terms

Learn next

Study preferential independence and utility independence next. Then change the worked example's affordability weight while adjusting the other weights to keep the total at one. Identify when the ordering reverses and explain why a ranking without its preference assumptions is incomplete.

Utility independence

Sources

Russell and Norvig, Artificial Intelligence: A Modern Approach, Chapter 16 teaching slides, especially multiattribute utility, preference independence, and utility independence: https://aima.cs.berkeley.edu/slides-pdf/chapter16.pdf . The numerical example and fictional case are original illustrations; they are not empirical results from the source.

Previous: Multiarmed bandit Next: Multiband blending