Must preferences actually change?

No. This formulation can learn about stable preferences that were initially uncertain.

Is every trial worthwhile?

No. Its information value must be considered alongside its costs and consequences.

Can an AI infer preferences perfectly from behavior?

No. Observed behavior needs interpretation and may reflect constraints or noise as well as preferences.

AI Glossary / Graphs and symbolic AI / Intermediate

Adaptive utility

Definition

Adaptive utility is a decision-theoretic approach in which uncertainty about preferences can be reduced through experience or information. Rather than assume that the value of every possible outcome is already known, the model represents uncertainty about the utility function and updates that understanding. This entry uses the preference-learning sense.

What it means

Ordinary uncertainty concerns what will happen: whether a delivery will arrive on time, for example. Preference uncertainty concerns how an outcome will be valued. Someone may know exactly what a new work arrangement involves yet remain unsure how much they will appreciate it after trying it.

Adaptive utility makes room for that second uncertainty. A decision can have both an immediate consequence and an informational consequence that affects later choices.

This need not mean that underlying preferences are changing. In the approach described here, a stable but incompletely known preference can become better understood. Other literature uses adaptive utility for actual changes in valuation. Those meanings should be distinguished instead of treating every changing score as the same theory.

How it works

Specify a family of possible utility functions, often indexed by a parameter θ, together with beliefs about which parameter describes the decision maker's preferences. Under a compatible scaling, outcomes can be evaluated by averaging their utilities over that uncertainty.

For example, an uncertain parameter might represent the trade-off between convenience and flexibility. Information from an experience or an explicit comparison can change the beliefs over that parameter, leading to a revised evaluation of future options.

In sequential decisions, learning has potential value because later actions can respond to what is discovered. An initially less attractive trial can sometimes be worthwhile, but only after accounting for its costs, consequences, and the decisions it can inform.

The utility scales, feedback interpretation, and update rule need justification. A numerical average across arbitrarily scaled utility functions is not automatically meaningful. Nor should a single observed choice be treated as an infallible reading of preference: constraints, mistakes, and context may also explain behavior.

Three practical examples

1. A workspace assistant helps a user compare quiet individual space with a shared studio. A short trial provides evidence about the user's preferences before a longer commitment.

2. A product-configuration tool remains uncertain about a customer's trade-off between portability and screen size. Carefully chosen comparisons help it improve recommendations without assuming the first answer reveals everything.

3. A planning assistant offers a reversible workflow pilot when a team is unsure how much it values flexibility relative to a fixed routine. Feedback informs future planning while mandatory operating requirements remain in place.

Think of it this way

Think of trying a sample before ordering a full meal. You are learning how much you like the taste, even when the ingredients were already known.

The analogy is limited. A formal model requires a defined uncertainty representation and a defensible interpretation of the evidence. Enjoying one sample does not establish a permanent preference across every context, and an AI should not infer broad personal priorities from a single narrow experience.

Worked example

Consider an original toy decision with two possible preference types, T1 and T2, initially assigned probabilities 0.5 each. On an assumed comparable utility scale, option A has utilities 10 under T1 and 2 under T2. Option B has utility 6 under either type.

Before learning, A's expected utility is 0.5 × 10 + 0.5 × 2 = 6, so A and B tie. Suppose new evidence has likelihood 0.8 under T1 and 0.2 under T2. Bayes' rule gives posterior probability 0.8 for T1 and 0.2 for T2.

After the evidence, A's expected utility becomes 0.8 × 10 + 0.2 × 2 = 8.4, while B remains 6. A is now preferred under this toy model. The options themselves did not change; the belief about which utility function applies did. The likelihoods and utilities are teaching assumptions, not empirical estimates or evidence that the user has objectively measurable happiness units.

Illustrative case study

Fictional case: Larch Design creates an assistant to help teams select collaboration routines. A manager initially expresses uncertainty between frequent short meetings and fewer longer sessions. The assistant records that uncertainty and proposes a limited, reversible trial approved by the team.

After the trial, participants report that interruptions matter more than expected. The assistant revises its recommendation and shows the feedback behind the revision. It does not interpret the trial as permission to change unrelated working conditions. This invented case illustrates preference learning and scope control, not a measured organizational benefit.

Common misconceptions and mistakes

Adaptive utility is not simply changing the objective whenever an optimization result is inconvenient. The update should be linked to evidence and a stated model.

Learning what someone values differs from persuading them to value what the system already delivers. The latter changes the problem rather than resolving preference uncertainty.

Do not equate utility uncertainty with uncertain outcomes, or assume that better personalization guarantees better decisions. Incorrect prior assumptions, poor feedback models, and incompatible utility scales can produce precise-looking but misleading recommendations. The term also should not be confused with any product that happens to use “adaptive utility” in its name.

Business use

The concept is useful when a decision process encounters genuinely unfamiliar outcomes or uncertain trade-offs. It encourages systems to preserve uncertainty, consider informative reversible steps, and explain why recommendations change.

Business teams should identify whose preferences are represented, which information can update them, and which requirements remain fixed. Evaluate whether the learning process improves relevant later decisions rather than merely fitting previous responses. Give users a way to correct inferred preferences and distinguish temporary contextual choices from broader priorities.

Check your understanding

A recommendation changes after new evidence about a user's trade-offs, while the available outcomes remain unchanged. Is that necessarily inconsistent?

Answer: No. Updating beliefs about the relevant utility function can rationally change the ranking, provided the evidence and update model are appropriate.

Related terms

Utility function
Preference learning
Decision theory
Preference elicitation
Bayesian updating
value of information
reward learning

Learn next

Next, compare uncertainty about outcomes with uncertainty about preferences. Then examine whether a proposed trial can change a later decision enough to justify its cost.

Sources

Houlding and Coolen, Adaptive Utility and Trial Aversion, primary accepted manuscript, definitions and preference-uncertainty discussion: https://www.tara.tcd.ie/bitstreams/eed1b967-33f6-44ce-a8eb-69a2b27ef498/download

Shah and colleagues, Benefits of Assistance over Reward Learning, related AI treatment of uncertainty over objectives: https://aima.cs.berkeley.edu/~russell/papers/neurips20ws-assistance.pdf

Primary passages read. Scope: learning uncertain preferences, distinct from changing valuation. Original numerical example.

Previous: Adaptive Synthetic Sampling Next: Add gate