Does MEU ignore risk?

No. Risk preferences can be represented in the utility function, although a poorly chosen function can fail to capture them.

Is the action with the highest success probability always selected?

No. Costs and the utilities of success and failure also matter.

Can new information change the choice?

Yes. It can change outcome probabilities or support a conditional policy, provided it is available before the relevant decision and its acquisition costs are included.

AI Glossary / Graphs and symbolic AI / Beginner

Maximum expected utility

Definition

Maximum expected utility, or MEU, is the principle of choosing the available action or policy with the highest probability-weighted utility across its possible outcomes. Utility represents the decision-maker's preferences, including relevant trade-offs. The preferred choice depends jointly on what may happen, how likely it is, and how desirable each outcome would be.

What it means

The most likely outcome is not always the most important one for a decision. An uncommon but costly failure can outweigh a common modest benefit. Expected utility brings both probability and consequence into the same comparison.

Utility also need not equal money or a raw performance score. A decision-maker may value reliability, time, comfort, or other outcomes differently, and risk preferences may make utility nonlinear in a measured quantity. Therefore maximizing expected utility is not automatically the same as maximizing expected revenue, average speed, or the probability of the single most favorable outcome.

How it works

For each feasible action a, identify outcomes o and their probabilities under that action and the available evidence e. Calculate EU(a|e)=ΣP(o|a,e)U(o,a), then choose an action attaining the largest value. Continuous outcomes use an integral rather than a finite sum.

The probabilities must describe what would happen if the action were taken, not merely correlations among historical observations. Utilities must be comparable across the alternatives. Hard constraints can first restrict the feasible action set.

For sequential decisions, evaluate policies that specify actions based on information available at each stage. A policy cannot legitimately depend on an observation that arrives only after its decision. The result is optimal relative to the stated model and preferences, not an unconditional guarantee about the real world.

Three practical examples

Delivery planning: compare a faster route with a more reliable route using travel uncertainty and the consequences of missing a deadline.
AI assistance: decide whether to answer immediately or request clarification by comparing response quality, delay, and the chance of misunderstanding.
Inspection planning: assess whether additional information is worth obtaining before choosing between acceptance, rework, and further testing.

Think of it this way

Think of comparing several baskets of possible outcomes. Each outcome receives a weight for how likely it is and a score for how desirable it is. The basket with the highest weighted score wins.

The analogy has limits: the scores cannot simply be arbitrary rankings if their numerical spacing is used in averages. The calculation needs a meaningful utility representation of preferences over uncertain prospects, along with a defensible probability model.

Worked example

An event team chooses between two equipment-delivery plans. Plan Fast has probability 0.8 of arriving on time and 0.2 of arriving late. Plan Reliable has probabilities 0.95 and 0.05. Suppose an on-time outcome has utility 100 and a late outcome utility −100. Fast has no extra burden; Reliable incurs a fixed utility cost of 10 in either outcome.

Fast's expected utility is 0.8×100+0.2×(−100)=60. Reliable's is 0.95×90+0.05×(−110)=80. MEU chooses Reliable even though its best-case utility is lower.

Now let the late-outcome utility before the fixed cost be −L. Fast has expected utility 80−0.2L. Reliable has 85−0.05L. Their difference is 5+0.15L, so Reliable remains preferable for every nonnegative L under these assumptions. If its fixed cost were instead 30, its expected utility would be 65−0.05L, and it would become preferable only when L exceeds 100. These hypothetical utility calculations reveal which trade-off controls the choice.

Illustrative case study

Fictional case: a museum develops an AI assistant for visitor questions. A confident answer is not its only objective: incorrect accessibility information is especially undesirable, while asking unnecessary follow-up questions creates inconvenience.

The team defines outcome categories, estimates probabilities from relevant evaluations, and compares answering with clarification or staff referral. Reviewers find that one utility table underweights the cost of sending visitors to an unavailable facility, so they revise the preference model before deployment. The fictional case illustrates that an optimizer can faithfully maximize an unsuitable objective; it does not claim a quantified improvement or a universal utility scale.

Common misconceptions and mistakes

Expected utility is not the utility of the expected outcome when the utility function is nonlinear. For example, averaging possible waiting times first can conceal preferences about occasional extreme delays.

MEU also does not guarantee that the selected action's realized outcome will be best. An unlucky result can occur after a well-supported decision. Conversely, a fortunate result does not validate poor probability estimates. Avoid double-counting costs already included in utilities, comparing scores on inconsistent scales, or treating precise arithmetic as evidence that uncertain inputs are precise.

Business use

The principle helps teams make trade-offs explicit. It reveals whether disagreement comes from different forecasts, different preferences, or different feasible choices. Sensitivity analysis can show whether modest changes in an assumption would reverse the recommendation.

For deployment, record the evidence used, action-dependent outcome estimates, utility definitions, constraints, and decision timing. Evaluate the consequences that matter, rather than only predictive accuracy. When stakeholders disagree about preferences, preserve that disagreement visibly instead of presenting a single aggregated score as if it were an objective fact about what everyone values.

Check your understanding

an action produces utility 10 with probability 0.9 and −100 with probability 0.1. Another produces utility 0 for certain. Which maximizes expected utility?

Answer: the certain outcome, because the first action's expected utility is 9−10=−1, below 0. Its high chance of a positive result does not by itself make it preferable.

Related terms

Expected utility
Utility function
Decision theory
Decision network
Policy
Maximin
value of information
risk preference
Bayesian decision theory

Learn next

change one probability and one utility in the worked example separately, and identify when the preferred action changes. Then study value of information: compare deciding now with a policy that chooses after receiving a signal, subtracting the signal's acquisition cost.

Policy

Sources

Poole and Mackworth, “Making Decisions Under Uncertainty,” single-decision and sequential-policy lecture material: https://artint.info/slides/ch09/lect2.pdf. Their textbook section “Expected Utility of a Policy”: https://artint.info/html1e/ArtInt_221.html. Relevant primary definitions and decision-timing passages were read. Delivery calculations and museum case are original teaching examples.

Previous: Maximum entropy spectral estimation Next: Maximum Flow