liveBook · Manning

Expected utility

As explained in Timeless Algorithms, The Foundational Ideas

Expected utility is the probability-weighted average desirability of the possible outcomes of an action.

A probability alone does not determine what to do. A physician considering treatment must weigh the chance of disease against both the benefit of treatment and the harm of treating someone unnecessarily. Expected utility combines the likelihood and desirability of an action’s possible outcomes into one measure for comparing choices. section 7.1 section 7.1.1

For an action a, expected utility is EU(a) = Σθ u(a, θ)p(θ | data). Here θ is a possible state of the world, u(a, θ) is the utility of the action–state outcome, and p(θ | data) is its posterior probability given the evidence. Multiply each outcome’s utility by its probability and sum across states; the preferred action maximizes this sum. For continuous states, an integral replaces the sum. section 7.2.2

Utility represents relative desirability on a common scale. It need not be monetary: benefits, costs, risks, and discomfort can all matter. The resulting recommendation depends on the actions considered, the estimated probabilities, and the utilities assigned. section 7.1 section 10.2.6

Treatment after a positive test

Suppose the posterior probability of disease is 0.09. With the illustrative utilities shown, treatment has expected utility 0.09(100) + 0.91(−10) = −0.1; no treatment has expected utility 0.09(0) + 0.91(100) = 91. On these assumptions, no treatment is preferred. A different assessment of the consequences could change the comparison. section 7.1.1 section 10.2.6

Table 7.1 Utilities for each combination of action and disease state.
State of the world
Treat
Do not treat
Disease present
+100 (successful cure)
0 (no treatment, poor outcome)
Disease absent
-10 (unnecessary treatment)
+100 (avoid discomfort)

The utility table makes the consequences of each treatment–disease combination explicit. section 7.1.1

The same choice can be expressed through opportunity loss: compare each outcome with the best outcome available in that state, then weight the losses by state probabilities. Minimizing expected opportunity loss identifies the same preferred action as maximizing expected utility. Decision trees extend the reasoning to sequences of choices and uncertain events; value-of-information analysis asks whether acquiring more evidence is worthwhile. section 7.3.2 section 7.6

A preferred action is not guaranteed to produce the best outcome. Expected utility makes a choice coherent with the available evidence and stated preferences, but uncertainty remains. section 7.6

How Timeless Algorithms, The Foundational Ideas covers it

How deep it goes
Develops the probability-weighted formulation, defines its terms, and compares actions using worked numerical calculations.
Examples
Works through a treatment decision after a positive test and a product launch versus delay under uncertain demand.
What you can do after
Evaluate actions under uncertainty by comparing expected utilities without mistaking a preferred decision for a guaranteed favorable outcome.
How it connects
Connects expected utility to Bayesian inference, decision trees, opportunity loss, and value-of-information analysis.

Also explained in

Compare how 3 books explain expected utility →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Timeless Algorithms, The Foundational Ideas. The section links point to the parts of the book it draws on.