Decisions & optimisation
Decision theory explained
Decision theory breaks a difficult choice into actions, possible developments, probabilities and valuations. An introduction with a payoff matrix, expected utility and a fully worked decision tree.
At a glance
- Decision theory breaks a choice into actions, states, probabilities and valuations, making it clear what the decision actually depends on.
- Expected value is the right criterion only for risk-neutral decision-makers. Expected utility accounts for attitudes to risk, measurable through the certainty equivalent.
- Decision trees solve multistage decisions backwards and show the maximum value of additional information.
Whether a business is expanding a production line, a doctor is recommending a treatment or a team is planning a product launch, a choice must be made before it is clear how the world will develop. Decision theory gives this situation structure. It asks which actions are available, which states might occur, how likely they are and how the possible outcomes should be valued—and uses these elements to determine which choice fits our assumptions and preferences (Peterson 2017).
There are two perspectives. Normative decision theory describes how a consistent decision-maker should choose. Descriptive decision theory studies how people actually decide. This article focuses mainly on the normative side and ends by showing where people systematically depart from it.
The building blocks of a decision
Every decision problem under uncertainty has five parts:
- Actions we can choose between,
- States of the world that we do not control,
- Outcomes resulting from the action and state,
- Probabilities of the states—if we can specify them,
- Preferences, meaning a valuation of the outcomes.
For example, a business is considering a small, medium or large expansion of a production line. Demand over the coming years may be low, medium or high. The payoff matrix shows the profit for each combination in thousands of euros. The numbers are constructed to illustrate the methods; they are not measurements.
| Expansion | Low demand | Medium demand | High demand |
|---|---|---|---|
| Small | 20 | 30 | 40 |
| Medium | 0 | 50 | 80 |
| Large | −50 | 30 | 140 |
No row is best in every column. That is precisely what makes the decision difficult: which option wins depends on which state occurs.
Deciding without probabilities
If we cannot or do not want to assign probabilities to the states, we are deciding under uncertainty without specified probabilities. Several classical rules address this situation:
- Maximin: choose the action whose worst outcome is best—the pessimist’s rule.
- Maximax: choose the action with the best best outcome—the optimist’s rule.
- Hurwicz: weight the best and worst outcomes using an optimism parameter , giving .
- Minimax regret: for each state, calculate how much is lost relative to the best action in that state, then choose the action with the smallest maximum regret (Savage 1954).
- Laplace: treat all states as equally likely and compare averages.
For the table above, this gives:
| Rule | Small | Medium | Large | Choice |
|---|---|---|---|---|
| Maximin (worst outcome) | 20 | 0 | −50 | Small |
| Maximax (best outcome) | 40 | 80 | 140 | Large |
| Hurwicz with | 30 | 40 | 45 | Large |
| Minimax regret (maximum regret) | 100 | 60 | 70 | Medium |
| Laplace (average) | 30 | 43.3 | 40 | Medium |
Regret is calculated column by column: the best outcomes in the three states are 20, 50 and 140. Choosing a small expansion and then experiencing high demand forgoes ; that is the largest regret in that row.
Expected value
Once probabilities can be assigned to the states—from data, models or a reasoned assessment—each action can be summarized by its expected value:
If the business assumes probabilities of 0.3, 0.5 and 0.2 for low, medium and high demand, the results are:
By expected value, the medium expansion is the best choice. The reasoning rests on the law of large numbers: making the same decision many times under identical conditions produces an average close to the expected value. For a large, one-off decision, however, expected value omits something important—the spread. The standard deviations of the outcomes are €7,000 for “small”, €29,100 for “medium” and €65,900 for “large”. The large expansion may mean a €50,000 loss, and that matters more to many decision-makers than its expected value reveals.
Expected utility: when risk matters
The St. Petersburg paradox
The problem originated with Nicolaus Bernoulli; in 1738 his cousin Daniel Bernoulli proposed the solution still cited today (Bernoulli 1738/1954). A fair coin is tossed until heads first appears. If that happens on toss , you receive ducats. The expected value is infinite because every term contributes exactly one ducat:
Yet hardly anyone would pay more than a small amount to play. Bernoulli’s explanation is that the value of additional money decreases with wealth. Valuing gains logarithmically gives finite expected utility, , making the game worth as much as a certain gain of ducats (ignoring existing wealth).
The idea of expected utility
John von Neumann and Oskar Morgenstern showed that if preferences over lotteries satisfy certain consistency axioms—completeness, transitivity, continuity and independence—there is a utility function such that the preferred lottery is precisely the one with higher expected utility (von Neumann & Morgenstern 1944):
Leonard Savage extended this approach to subjective probabilities (Savage 1954). The shape of the utility function describes the attitude towards risk. If is concave, : a certain amount equal to the expected value is worth at least as much as the lottery itself. This is risk aversion.
Certainty equivalent and risk premium
The certainty equivalent (CE) of a risky alternative is the certain amount with the same utility, . The difference between expected value and certainty equivalent is the risk premium: the expected gain a decision-maker is willing to give up to eliminate the risk.
The exponential utility function is widely used in decision analysis because it requires only one parameter, risk tolerance (Howard & Abbas 2016):
With a risk tolerance of thousand euros, the production-line expansion gives:
| Expansion | Expected value | Standard deviation | Certainty equivalent | Risk premium |
|---|---|---|---|---|
| Small | 29.0 | 7.0 | 28.8 | 0.2 |
| Medium | 41.0 | 29.1 | 36.7 | 4.3 |
| Large | 28.0 | 65.9 | 9.0 | 19.0 |
The medium expansion remains ahead, but the large one loses substantial value: its certainty equivalent is only €9,000 because the possible €50,000 loss weighs heavily. If the business were even more cautious, the ranking would change—below a risk tolerance of approximately €34,000, the small expansion would be the best choice.
Decision trees: thinking in stages
Many decisions are not made all at once. Often, we can first gather information and then decide. A decision tree represents these sequences: squares are decisions, circles are chance events, and outcomes appear at the ends (Raiffa 1968).
Backward induction
A decision tree is solved from right to left: calculate the expected value at each chance node, and select the best branch at each decision node (Clemen & Reilly 2014).
- Launch immediately: .
- Test result: a positive result occurs with probability . By Bayes’ theorem, the probability of success then rises to ; after a negative result, it falls to .
- After a positive test, launching yields —so launch. After a negative test, it would yield —so stop.
- Market test overall: , less the test cost: .
The root therefore has value . The best strategy is to test first, launch if the result is positive, and stop if it is negative. Its expected value is €48,000 higher than launching immediately.
What is information worth?
The tree answers a second, often more important question: how much can information cost? The expected value of perfect information (EVPI) compares two worlds. If the team knew for certain beforehand whether the product would succeed, it would launch only in the success case, yielding an expected . Without information, the value is 60. Thus:
No test, study or dataset can improve this decision by more than €180,000. The imperfect market test is worth thousand euros: this is the expected value of sample information (EVSI). It exceeds the €30,000 test cost and captures roughly 43% of the maximum possible information value. Both concepts originate in the Bayesian decision theory of Howard Raiffa and Robert Schlaifer (Raiffa & Schlaifer 1961).
How people actually decide
Expected utility is a normative model. People systematically depart from it. Maurice Allais demonstrated this in 1953 with a famous pair of choices (Allais 1953). In the usual textbook version:
- Choice 1: A = €1 million for certain; B = a 10% chance of €5 million, 89% of €1 million, and 1% of nothing.
- Choice 2: C = an 11% chance of €1 million, otherwise nothing; D = a 10% chance of €5 million, otherwise nothing.
Many people prefer A and D. This is incompatible with expected utility, because both comparisons reduce to the same inequality:
Daniel Kahneman and Amos Tversky brought such findings together in Prospect Theory (Kahneman & Tversky 1979): people evaluate outcomes relative to a reference point, losses weigh more heavily than equally large gains, and small probabilities are overweighted. This does not refute normative theory; it indicates where decisions need support—in explicitly formulating probabilities, reference points and attitudes to risk.
Good decision, bad outcome
In our example, the team launches after a positive test—and still faces a 36% probability of failure. Would the decision then have been bad? No. Ronald Howard emphasizes the distinction between decision quality and outcome quality: a decision is good if it fits the available information, alternatives and preferences; the outcome also depends on chance (Howard & Abbas 2016).
Three practical rules follow:
- Document assumptions. Recording the success probability and test accuracy makes it possible to check later which assumption was wrong.
- Learn across many decisions. Whether probabilities are well calibrated only becomes clear across a series of cases.
- Compare options before committing resources. A decision tree shows where additional information offers the greatest value.
Conclusion
Decision theory requires discipline rather than exotic mathematics: write down actions, states, probabilities and valuations separately. Expected value is the natural starting point; expected utility adds the attitude towards risk; decision trees show when it is worth learning first and acting afterwards. The article on optimization explains how to find the best feasible action when there are millions of options rather than three. Bayesian statistics explains where the probabilities come from.
Frequently asked questions
What is the difference between expected value and expected utility?
Expected value averages the outcomes themselves, weighted by their probabilities. Expected utility instead averages the utility of those outcomes. If the utility function is concave, large losses weigh more heavily than equally large gains. Expected utility therefore represents risk aversion, while expected value does not.
How do you calculate a decision tree?
Work from right to left: at each chance node, calculate the expected value (or expected utility) of the subsequent branches; at each decision node, choose the branch with the best value. The root then gives the value of the best strategy, and the branches chosen along the way constitute that strategy.
What does the expected value of perfect information (EVPI) mean?
EVPI is the difference between the expected outcome if you knew which state would occur before deciding, and the best expected outcome without that information. No study, test or dataset can be worth more for this decision.
Is a decision bad if the outcome is bad?
Not necessarily. A decision is good if it fits the information, alternatives and preferences available at the time. Chance can give even a good decision a bad outcome, so evaluate the process and assumptions, not just the result.
Sources and further reading
- Bookvon Neumann, J. and Morgenstern, O. (1944): Theory of Games and Economic Behavior. Princeton University Press.The axiomatic foundation of expected utility appears in the appendix to the second edition (1947).
- PaperBernoulli, D. (1954): Exposition of a New Theory on the Measurement of Risk. Econometrica 22(1), 23–36 (English translation of the Latin original from 1738).
- BookSavage, L. J. (1954): The Foundations of Statistics. Wiley.
- BookRaiffa, H. and Schlaifer, R. (1961): Applied Statistical Decision Theory. Division of Research, Harvard Business School.
- BookRaiffa, H. (1968): Decision Analysis: Introductory Lectures on Choices under Uncertainty. Addison-Wesley.
- PaperAllais, M. (1953): Le comportement de l’homme rationnel devant le risque: critique des postulats et axiomes de l’école américaine. Econometrica 21(4), 503–546.
- PaperKahneman, D. and Tversky, A. (1979): Prospect Theory: An Analysis of Decision under Risk. Econometrica 47(2), 263–291.
- BookClemen, R. T. and Reilly, T. (2014): Making Hard Decisions with DecisionTools (3rd edition). South-Western Cengage Learning.A practical textbook on decision trees, utility functions and the value of information.
- BookPeterson, M. (2017): An Introduction to Decision Theory (2nd edition). Cambridge University Press.A concise, philosophically rigorous overview of normative decision theory.
- BookHoward, R. A. and Abbas, A. E. (2016): Foundations of Decision Analysis. Pearson.
Content last reviewed on 25 September 2026. Please let us know about any errors: impressum@nyxai.com.