AI Glossary / Graphs and symbolic AI / Intermediate
Additive value function
Definition
An additive value function evaluates an option by adding separate contributions from its attributes. A common form is V(x) = sum of wᵢvᵢ(xᵢ), where each vᵢ converts an attribute into a value score and each wᵢ scales its contribution. The result represents preferences under the model's assumptions, rather than an objective measure of universal worth.
What it means
Many decisions involve several considerations, such as purchase cost, reliability, and ease of maintenance. An additive model gives each consideration an explicit contribution to a combined score, making the trade-offs easier to inspect.
The word “additive” describes the overall combination. Individual attribute value functions can be nonlinear: an improvement from unacceptable to adequate reliability may be valued differently from an equally sized improvement at an already excellent level.
This entry concerns multiattribute decision analysis. An additive value function in reinforcement learning can refer to a different decomposition of predicted returns. Also, ranking certain outcomes with a value function does not by itself justify using the same numbers as utilities for risky lotteries.
How it works
Define the decision attributes and their relevant ranges. Specify how each raw measurement becomes a value score, including which direction is better. Then establish scaling factors or weights that express the trade-offs between changes across those ranges.
Evaluate each option using the same functions and weights, and add its contributions. A higher total indicates a preferred option under that model. Some formulations absorb weights into the component functions, so a formula without visible weights can still be additive.
The central restriction is that there are no explicit interaction terms between attributes. Appropriate preference-independence conditions, with the usual representation assumptions, support additive forms. Independence here concerns preferences, not whether measured attributes are statistically correlated.
Check whether a trade-off changes when another attribute changes. If a feature is valuable only when another feature is present, a plain additive model may misrepresent that interaction. Mandatory requirements should also be handled explicitly rather than accidentally traded away through a favorable total score.
Three practical examples
1. A procurement team compares printers using assessed values for operating cost, output quality, and maintenance access. It can inspect which contribution drives each option's total.
2. A project portfolio review scores candidate initiatives on agreed attributes after applying mandatory eligibility rules. Sensitivity analysis shows whether small weight changes alter the ranking.
3. A personal planning assistant compares accommodation options using location, space, and flexibility. It should check whether the value of extra space depends strongly on who will share it before assuming additivity.
Think of it this way
Think of a scorecard with separate columns whose points are added. The total is easy to reproduce because each column's contribution remains visible.
The analogy is limited because a convenient scorecard is not automatically a faithful preference model. If two features only work well together, awarding independent points can miss the combination's value. The meaning of the scales and the trade-offs matters as much as the arithmetic.
Worked example
Consider an original toy comparison using three already-assessed value scores on a 0-100 scale: affordability, serviceability, and capacity. The weights are 0.4, 0.35, and 0.25. These are assumptions for illustration, not recommended procurement weights.
Option A scores 80, 60, and 70. Its total is 0.4 × 80 + 0.35 × 60 + 0.25 × 70 = 32 + 21 + 17.5 = 70.5. Option B scores 60, 90, and 65, totaling 24 + 31.5 + 16.25 = 71.75. B leads by 1.25 points.
Now increase the affordability weight to 0.5, reduce serviceability to 0.25, and retain capacity at 0.25. A scores 72.5, while B scores 68.75. The preference reverses. The exercise shows that the original narrow lead depends on the assumed trade-offs, not that either option is inherently best. All scores must refer to consistent attribute definitions and ranges.
Illustrative case study
Fictional case: Willow Studios uses an additive model to choose editing workstations. A candidate receives high scores for a graphics card and an inexpensive enclosure. During review, the team discovers that the enclosure cannot accommodate the selected card.
Compatibility becomes an explicit feasibility check before scoring. The team then uses the additive model to compare the remaining valid configurations and reviews whether performance attributes overlap. This invented case illustrates why a favorable sum cannot rescue an invalid combination, without claiming a measured purchasing outcome.
Common misconceptions and mistakes
Additivity does not require each component function to be a straight line. It requires their contributions to combine by addition without explicit cross-attribute interactions.
Weights are not meaningful as vague importance labels detached from attribute ranges. Changing a scoring scale can alter a contribution unless the model is adjusted consistently.
Do not confuse preference independence with statistical independence. Nor should correlated or overlapping criteria be counted repeatedly without considering what preference they represent. Finally, a score difference is not automatically a probability of success, a monetary benefit, or a measure of confidence in the ranking.
Business use
Additive value functions can support transparent comparisons when the attributes and trade-offs have been carefully established. They allow reviewers to trace a recommendation to its component judgments instead of accepting an unexplained overall score.
Before relying on the result, test plausible alternative weights and uncertain attribute estimates. Review close rankings, possible interactions, and duplicated considerations. Document who supplied the preferences and whether they represent the intended decision maker. A mathematically correct total can still answer the wrong question if the underlying value judgments are mismatched.
People also ask
Must the weights sum to one?
Normalized weighted models often use that convention, but a general additive representation can absorb scaling into its component functions.
Does additive mean linear in every raw measurement?
No. The attribute value functions can be nonlinear even though their contributions are added.
Can one excellent attribute compensate for a poor one?
In a compensatory additive model, yes. If compensation is unacceptable, represent the relevant limit explicitly.
Check your understanding
A model adds separate points for two components that cannot operate together. Does the high total show that the combined option is suitable?
Answer: No. Compatibility must be checked. The additive score ranks options under its assumptions and cannot establish feasibility when a necessary constraint is missing.
Related terms
- Preference independence
- Weighted sum
- Sensitivity analysis
- Utility function
- Value function
- multiattribute utility
- trade-off
- constraint
Learn next
Next, examine preferential independence and compare value functions with expected-utility models. Recalculate a small scorecard under alternative weights, then look for an interaction that a simple sum cannot express.
More in Graphs and symbolic AI
- Actor - An actor, in planning, is an entity that performs actions within the modeled environment.
- Adaptive utility - Adaptive utility is a decision-theoretic approach in which uncertainty about preferences can be reduced through experience or information.
- Add list - An add list is the part of a STRIPS-style planning action that specifies the facts to be made true in the resulting state.
- Admissible heuristic - An admissible heuristic is a search estimate that never overestimates the true minimum remaining cost of reaching a goal.
- Adversary argument - An adversary argument is a proof technique that demonstrates a limitation of an algorithm or class of algorithms by constructing difficult inputs or responses.
- Aggregation - Aggregation, in planning, is the representation of multiple relevantly interchangeable objects or resources as a quantity or group instead of distinguishing...
Sources
Russell and Norvig, official Artificial Intelligence: A Modern Approach slides, preference structure for deterministic attributes and additive value representation: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter16-6pp.pdf
UK Government Analysis Function, primary practitioner guidance on attribute scales, weighting, and additive models: https://analysisfunction.civilservice.gov.uk/policy-store/an-introductory-guide-to-mcda/
Primary explanations read. Certain-outcome value and lottery utility are distinguished. Examples are original.
Previous: Additive Rule Next: Address space