Is a multiplicative utility function just a geometric mean?

No. Product-based multiattribute utility formulations can include shifts, scaling constants, and interaction terms beyond that particular aggregation rule.

Must the attributes be statistically independent?

Utility independence is a different assumption about preferences. Expected utility can require the joint outcome distribution.

Is multiplication always better than addition?

No. The appropriate form depends on the preferences and assumptions being represented, rather than mathematical complexity alone.

AI Glossary / Mathematics and optimization / Advanced

Multiplicative utility function

Definition

A multiplicative utility function combines utilities for several attributes through a product-based expression that includes interactions between those attributes. In multiattribute decision theory, this form can represent suitable preferences over uncertain outcomes under utility-independence assumptions. It is more specific than multiplying arbitrary business scores, and its parameters must reflect the preferences being modeled.

What it means

A decision may involve several attributes, such as service quality and delivery convenience. An additive utility model adds separate contributions. A multiplicative model can make the contribution of improving one attribute depend on the level of another.

The word “multiplicative” does not necessarily mean that overall utility is simply the product of raw attribute values. Standard formulations use single-attribute utility functions, scaling constants, and normalization. Expanding a product produces both individual terms and interaction terms. This entry uses the multiattribute decision-theory sense associated with the supplied AIMA source, rather than an arbitrary geometric scoring rule.

How it works

Define the attributes and their relevant worst and best outcomes. Elicit single-attribute utility functions and examine whether the required preference-independence assumptions are credible. Utility independence concerns preferences over lotteries in one subset of attributes while the other attributes are fixed. It does not mean the attributes are statistically independent in actual outcomes.

Choose and calibrate an appropriate multiplicative representation. One useful illustrative two-attribute expression is U(u,v)=((1+u)(1+v)−1)/3=(u+v+uv)/3, with u and v between zero and one. It is normalized to zero at (0,0) and one at (1,1). This example's constants are chosen for teaching, not prescribed for every decision maker.

For uncertain outcomes, calculate utility for each joint outcome and then average using its probability. Do not generally substitute average attributes into the utility formula: product terms mean that dependence between attributes can affect expected utility.

Three practical examples

1. Product configuration: compare designs when a customer's preference for an improvement in ease of use depends on the level of another product attribute.

2. Service planning: assess options involving several uncertain service outcomes using a utility model whose structure has been checked with the relevant decision makers.

3. Decision-support research: compare additive and multiplicative preference models to understand which assumptions change a recommended choice and which outcomes need further preference assessment.

Think of it this way

Think of adjusting two controls whose effects interact. Increasing one control may have a larger effect when the other is already high, so adding two isolated effects misses part of the picture.

The analogy captures interaction, but it does not explain which preferences a person should have or establish that every pair of desirable attributes is complementary. Interaction can take different forms and signs under different parameterizations. A valid utility model requires more than an appealing metaphor.

Worked example

Use the illustrative normalized function U(u,v)=(u+v+uv)/3. Option A has attribute utilities u=1 and v=0.2, so U(A)=(1+0.2+0.2)/3=1.4/3, approximately 0.467. Option B has u=v=0.6, giving U(B)=(0.6+0.6+0.36)/3=0.52.

An equal-weight additive score (u+v)/2 gives both options 0.6. The multiplicative example prefers B because its interaction term is larger. This is a consequence of the chosen preference model, not proof that B is objectively superior.

Now consider a lottery with equal probabilities of (u,v)=(1,1) and (0,0). Its expected utility is 0.5×1+0.5×0=0.5. Both attributes have mean 0.5, but evaluating the mean pair gives U(0.5,0.5)=1.25/3, approximately 0.417. The two calculations differ because E[uv]=0.5 while E[u]E[v]=0.25.

If instead the lottery offers (1,0) and (0,1) equally, the mean attributes remain 0.5 and 0.5, but expected utility becomes 1/3. Joint outcome structure matters even when the marginal means match.

Illustrative case study

Fictional case study: a software team builds a preference-based comparison tool for service packages. Its first version multiplies two satisfaction ratings without documenting their scales or explaining why multiplication matches customer preferences.

A review replaces the automatic rule with an explicit elicitation exercise. Participants compare fixed outcomes and simple uncertain choices, while analysts test whether an additive or a justified multiplicative representation is suitable. The tool records the chosen model and shows when plausible parameter changes reverse rankings. No customer preference or business improvement is assumed in advance.

Common misconceptions and mistakes

Do not multiply raw dollars, minutes, and survey ratings and call the result a calibrated utility function. Attribute scales and utility transformations have different roles.

Do not confuse utility independence with independence of random variables. The former concerns conditional preference orderings; the latter concerns a probability distribution.

Do not assume every multiplicative representation forces overall utility to zero when one single-attribute utility is zero. In the worked function, U(1,0)=1/3. Also avoid interpreting 0.8 utility as literally twice the satisfaction of 0.4; utility normalization does not supply that psychological ratio meaning.

Business use

This concept is useful when a decision-support system must represent explicit tradeoffs between attributes and uncertainty. It can make hidden interaction assumptions visible, but it also creates a responsibility to explain and test them.

Document the attribute ranges, elicitation method, normalization, probability assumptions, and sensitivity of recommendations. Keep hard requirements separate when they truly are nonnegotiable. A scoring model should not silently permit compensation for a requirement that the organization intended to enforce as a constraint.

Check your understanding

For the worked function, does a lottery's expected utility necessarily equal utility evaluated at its mean attribute utilities?

Answer: No. The uv term depends on joint outcomes. The two equal-probability examples have identical means but different expected utilities.

Related terms

Learn next

compare additive and product-based scores using fixed attribute scales. Then construct two lotteries with the same attribute means but different joint outcomes, and explain which additional preference and probability information a defensible decision model needs.

Sources

Sources: Russell and Norvig, Rational decisions, multiattribute utility and stochastic preference structure: https://aima.cs.berkeley.edu/slides-pdf/chapter16.pdf . Relevant primary slides were read. The supplied GitHub glossary did not expose readable entry content. The normalized two-attribute function and all numerical comparisons are original illustrations of product interaction, not a universal calibration formula or an elicited stakeholder utility model.

Previous: Multiplicative noise Next: Multiplicative weight update method