--- title: "Decision theory" source: "https://systems-analysis.info/eng/Decision_theory" wiki: "systems-analysis.info/eng" article: "Decision_theory" language: "en" categories: - "Category:Decision theory" - "Category:English" - "Category:Science" revision_id: 118 wiki_created_at: 2026-09-06T22:17:37Z wiki_modified_at: 2026-09-06T22:17:37Z downloaded_at: 2026-09-07T22:21:08Z --- # Decision theory **Decision theory** is an interdisciplinary field devoted to the development of methods, models, and procedures for rational and well-founded choice among one or more courses of action under conditions of limited resources, multiple objectives, uncertainty, and risk. In modern scholarship, decision theory is structured along three complementary lines of inquiry: normative (how an ideally rational agent *ought* to choose), descriptive (how people *actually* make decisions), and prescriptive (what tools can help a real decision-maker approximate rational choice). This tripartite distinction, established in the work of Bell, Raiffa, and Tversky (1988), defines the architecture of the discipline. ## Definition Decision-making is the conscious process of selecting the best — in some appropriate sense — alternative from a given set of feasible options. In professional settings, this process constitutes a specific form of purposeful human activity, particularly significant in management, design, planning, and other goal-directed endeavors. The purpose of decision theory is to provide the **decision-maker** (DM) with the means to: - formalize the problem, - identify the set of feasible alternatives, - assess possible consequences, - justify the choice of preferred alternatives in light of stated objectives and constraints. ## Aims and scope The overarching aim of decision theory is to develop instruments for rational choice. Its specific tasks include: - constructing models of preferences and objectives; - characterizing conditions of uncertainty, risk, and conflict; - formalizing the process of evaluating alternatives; - developing procedures for comparison, ranking, aggregation, and selection; - analyzing the robustness and justifiability of decisions; - designing models and algorithms for decision support. ## Normative theory and axiomatic foundations Normative decision theory addresses the question of how an ideally rational agent *ought* to act. Its starting point is the **decision matrix**: rows correspond to alternatives (acts), columns to states of the world, and cells to outcomes. The task is to select the row that is optimal according to a given criterion. The most elementary criterion is **dominance**: alternative $A$ dominates alternative $B$ if it yields an outcome no worse under every state and strictly better under at least one. Under risk, this principle generalizes to **stochastic dominance**: an alternative stochastically dominates another if its cumulative distribution function is nowhere higher. The dominance principle is accepted by virtually all normative models and serves as a minimal condition of rationality. ### Expected utility theory The foundation of normative theory is **expected utility theory** (EU theory), originating in the work of John von Neumann and Oskar Morgenstern, *Theory of Games and Economic Behavior* (1944; the axiomatic treatment of expected utility appeared in the second edition, 1947). The very idea of utility as a measure distinct from monetary payoff was proposed by Daniel Bernoulli (1738) in response to the **St. Petersburg paradox** — a problem in which the expected monetary value of a gamble is infinite, yet no one is willing to pay an arbitrarily large sum to participate. The idea of deriving subjective probabilities and utilities from preferences was first advanced by F. P. Ramsey (1926). According to expected utility theory, a rational agent selects the alternative that maximizes the expected value of a utility function over all possible outcomes: $EU(a) = \sum\limits_{i}p_{i} \cdot u(x_{i})$, where $a$ is an alternative, $p_{i}$ is the probability of outcome $x_{i}$, and $u$ is the utility function. The key axioms are completeness, transitivity, continuity, and **independence** of preferences. The last axiom (if $A \succ B$, then a mixture of $A$ with any $C$ is preferred to a mixture of $B$ with $C$ in the same proportions) is the most contested: the Allais paradox demonstrates its violation in the context of objective probabilities. The Ellsberg paradox, in turn, violates the related **sure-thing principle** in Savage's axiom system, which deals with subjective probabilities. When the axioms are satisfied, a **representation theorem** guarantees the existence of a numerical utility function: the agent's preferences can be represented as EU maximization. It is important to distinguish two types of utility. **Ordinal utility** merely ranks alternatives and is unique up to an arbitrary monotone transformation. **Cardinal utility** permits comparison of differences between evaluations and is unique up to a positive affine transformation ($u^{\prime} = \alpha \cdot u + \beta$, $\alpha > 0$). It is cardinal utility that figures in the expected utility formula. The shape of the utility function $u(x)$ reflects the DM's **attitude toward risk**: a concave function corresponds to risk aversion, a convex function to risk seeking, and a linear function to risk neutrality. ### Subjective expected utility An extension of the classical model is **subjective expected utility** (SEU), due to L. J. Savage (1954), in which probabilities are not given objectively but are formed by the DM on the basis of personal beliefs. Savage proposed his own axiom system, from which both a subjective probability measure and a utility function are derived simultaneously. A closely related result was obtained by F. J. Anscombe and R. J. Aumann (1963), who developed an axiomatic framework accommodating both objective and subjective probabilities. A distinct strand is **subjective probability as a measure of belief**, originating with B. de Finetti (1937): probability is defined through the agent's willingness to accept bets, and de Finetti's theorem establishes the coherence of subjective assessments. ### Interpretations of probability and Bayesianism The foundations of decision theory are inseparable from the debate over the **interpretation of probability**. The classical interpretation (P.-S. Laplace) defines probability through the equipossibility of outcomes. The frequentist (objectivist) interpretation (R. von Mises, J. Neyman, E. Pearson) treats probability as the limit of relative frequency in a repeating series of experiments. The subjectivist interpretation (F. P. Ramsey, B. de Finetti, L. J. Savage) defines probability as the agent's degree of belief, derivable from preferences. The logical interpretation (J. M. Keynes, R. Carnap) views probability as an objective measure of the degree to which evidence confirms a hypothesis. This division gives rise to two competing approaches in decision theory. **Bayesian decision theory** requires the agent to hold a complete prior distribution over the state space; decisions are made by maximizing subjective expected utility, and beliefs are updated via Bayes' rule. **Non-Bayesian decision theory** (including the frequentist statistical theory of Neyman–Pearson and Wald's minimax approach) dispenses with a single prior distribution and evaluates decision rules by their worst-case or frequentist properties. Modern approaches to ambiguity — such as multiple-priors models — occupy an intermediate position. The central pragmatic arguments in favor of the Bayesian axioms are the **Dutch book argument** — if the agent's subjective probabilities violate the axioms, a set of bets can be constructed under which the agent is guaranteed to lose — and the **money pump** — if preferences are intransitive, the agent can be made to pay for an endless cycle of exchanges. De Finetti (1937) proved that coherent degrees of belief satisfy the axioms of probability; a separate diachronic Dutch book argument (P. Teller, 1973; D. Lewis, 1999) justifies updating beliefs via Bayes' rule. Critiques of Bayesianism include concerns about the completeness of preferences, sensitivity to the choice of prior, and the problem of ambiguity, discussed below. ### Causal and evidential decision theory In the philosophy of decision-making, two rival approaches address the connection between acts and outcomes. **Evidential decision theory** (EDT; R. C. Jeffrey, 1965) evaluates acts by the conditional probability of desired outcomes given the act. **Causal decision theory** (CDT; D. Lewis, 1981; J. Joyce, 1999) requires that the *causal* connection between act and outcome be considered, rather than mere correlation. The distinction manifests in problems involving common causes (Newcomb's problem) and remains an active area of debate. ## Statistical decision theory Statistical decision theory (J. Neyman, A. Wald, J. O. Berger) frames choice as the problem of constructing an optimal **decision rule** — a function mapping observed data into the action space. Its central concepts are: - **loss function** $L(\theta,d)$ — the loss incurred by action $d$ when the true state of nature is $\theta$; - **risk function** $R(\theta,\delta) = E_{\theta}\lbrack L(\theta,\delta(X))\rbrack$ — the expected loss of decision rule $\delta$ for a given $\theta$; - **admissibility** — a decision rule is admissible if no other rule has risk no greater for all $\theta$ and strictly less for at least one; - **Bayes rule** — minimizes the average risk with respect to a prior distribution over $\theta$; - **minimax rule** — minimizes the maximum risk over all possible $\theta$. Statistical decision theory connects normative choice theory with mathematical statistics and underpins Bayesian inference, estimation theory, and hypothesis testing. An important result is the **complete class theorem**: under certain regularity conditions, every admissible decision rule is Bayes or a limit of Bayes rules. This means that any reasonable decision rule can be approximated by a Bayes rule for some choice of prior — a fundamental bridge between the frequentist and Bayesian paradigms. Within statistical decision theory, parameter estimation, hypothesis testing, and the construction of confidence regions are treated as special cases of choosing an optimal decision rule under the corresponding loss function: squared error loss for estimation, zero–one loss for testing. ## Descriptive theory Descriptive theory investigates how people *actually* make decisions and documents systematic departures from normative models. ### Bounded rationality H. A. Simon (1955) showed that real decision-makers cannot enumerate and evaluate all alternatives. Instead of maximizing utility, people employ the principle of **satisficing** — selecting the first option that meets an acceptable aspiration level. ### Prospect theory D. Kahneman and A. Tversky (1979) proposed a model in which the DM evaluates outcomes not by their absolute magnitude but by their deviation from a reference point. The value function in prospect theory is asymmetric (S-shaped): losses are felt more acutely than equivalent gains (**loss aversion**). Probabilities are transformed nonlinearly: small probabilities are overweighted and large probabilities are underweighted. In 1992, Tversky and Kahneman proposed **cumulative prospect theory**, extending the model to an arbitrary number of outcomes and ensuring compatibility with stochastic dominance. ### Cognitive heuristics, biases, and paradoxes Other major results of descriptive theory include: - **cognitive heuristics** — simplified judgment strategies (availability, representativeness, anchoring); - **cognitive biases** — systematic errors, notably the framing effect (violation of the description invariance axiom), the sunk cost fallacy, and overconfidence; - **paradoxes** — experimental violations of normative axioms (the Allais paradox, the Ellsberg paradox). ### Naturalistic decision making The **naturalistic decision making** (NDM) program (G. Klein) studies how experienced professionals — surgeons, firefighters, pilots, military commanders — make decisions in real-world conditions: under time pressure, incomplete information, and high stakes. Unlike laboratory research on cognitive biases, NDM emphasizes pattern recognition and situational awareness (the recognition-primed decision model). ## Decision-making conditions F. Knight (1921) identified three basic types of conditions: certainty, risk, and uncertainty. An extended classification adds a fourth type — conflict — in which the outcome depends on the actions of a rational adversary. ### Decision making under certainty The DM knows exactly which outcome each alternative will produce. The problem reduces to optimizing an objective function over the set of feasible solutions. This category encompasses the classical problems of mathematical programming — linear, nonlinear, and integer. ### Decision making under risk The outcomes of alternatives depend on states of the world whose probabilities are known or can be estimated. The basic tool is maximization of expected utility (or expected monetary value, EMV). Typical models include decision trees, Bayesian analysis, and the Bayes criterion. ### Decision making under uncertainty The probabilities of states of the world are unknown and cannot be reliably estimated. Choice criteria reflect different attitudes toward the unknown: - **Wald's criterion** (maximin) — select the alternative with the best worst-case outcome; a strategy of extreme pessimism; - **maximax criterion** — select the alternative with the best best-case outcome; a strategy of extreme optimism; - **Hurwicz criterion** — a weighted combination of optimism and pessimism with parameter $\alpha \in \lbrack 0,1\rbrack$; - **Savage's criterion** (minimax regret) — minimize the maximum foregone payoff; - **Laplace criterion** — assume equal probability across states and maximize the average payoff. For formalizing choice under vague, linguistic information, the framework of **fuzzy set theory** (L. Zadeh, 1965) is widely used, particularly in modern MCDM methods (Fuzzy AHP, Fuzzy TOPSIS). #### Decision making under ambiguity The Ellsberg paradox showed that decision-makers systematically distinguish situations with known and unknown probabilities, exhibiting **ambiguity aversion**. This observation stimulated the development of models generalizing classical EU theory: - **Choquet expected utility** (D. Schmeidler, 1989) — probabilities are replaced by a non-additive capacity, and integration is performed in the sense of Choquet; - **Maxmin expected utility** (I. Gilboa, D. Schmeidler, 1989) — the DM considers a set of plausible probability distributions and maximizes EU under the worst-case distribution; - **Multiple priors model** — a generalization in which the DM operates with an entire family of probability measures rather than a single one. These models constitute an active area of research in modern decision theory, linking axiomatic theory, the economics of uncertainty, and the philosophy of probability. ### Decision making under conflict The outcome depends on the actions of a rational adversary pursuing its own objectives. The formal apparatus is **game theory**. Key results include the **minimax theorem** (J. von Neumann, 1928), guaranteeing the existence of optimal **mixed strategies** (randomized action choices) in zero-sum games, and **Nash equilibrium** (J. Nash, 1950) — a strategy profile in which no player can benefit by unilaterally changing their strategy. Game-theoretic problems vary along several dimensions: zero-sum (strictly competitive) games; general-sum noncooperative games; cooperative games in which players can form binding agreements (bargaining problems, Nash bargaining solution); repeated games, where strategic behavior changes through the possibility of punishment and cooperation; and evolutionary games, modeling strategy adaptation in populations. Canonical examples include the **Prisoner's Dilemma** (conflict between individual and collective rationality), the **Battle of the Sexes** (a coordination problem), and the **Stag Hunt** (choosing between safe and risky cooperation). In games of incomplete information (J. Harsanyi, 1967), players do not know each other's types; the solution concept is **Bayesian Nash equilibrium**. The relationship between game theory and decision theory is bidirectional: a game against nature (a single DM facing a passive environment) is a special case of a game, while strategic interaction requires each player to solve a choice problem that accounts for the rationality of the others. ## Dynamic and intertemporal choice ### Sequential decisions When decisions are made in multiple stages, the key tool is **dynamic programming**, based on the principle of optimality due to R. Bellman (1957): an optimal policy has the property that every continuation is itself optimal relative to the state reached at the current stage. In stochastic multi-stage problems, dynamic programming combined with a probabilistic transition model yields the framework of **Markov decision processes** (MDPs). When the state is only partially observable, the model generalizes to **partially observable Markov decision processes** (POMDPs). MDPs and POMDPs constitute one of the fundamental formalisms for sequential choice under uncertainty and form the foundation of modern reinforcement learning. ### Intertemporal choice and time preferences A distinct class of problems involves choosing among outcomes distributed over time. The classical **discounted utility** model (P. Samuelson, 1937) assumes exponential discounting of future benefits at a constant rate. Descriptive research, however, has shown that people systematically deviate from the exponential model, exhibiting **hyperbolic discounting** and the associated time inconsistency of preferences (G. Ainslie; D. Laibson). Intertemporal choice occupies a central place in behavioral economics, finance, and health economics. ## Decision analysis **Decision analysis** is a prescriptive discipline that grew out of the work of H. Raiffa and R. A. Howard in the 1960s, aimed at providing systematic assistance to the real-world decision-maker. Unlike normative theory, which formulates the ideal, and descriptive theory, which documents reality, decision analysis offers a practical procedure for overcoming cognitive limitations. The main stages of decision analysis are: - **problem structuring** (problem framing) — defining objectives, criteria, alternatives, and the boundaries of the problem; - **preference and probability elicitation** — extracting from the DM a utility function, criterion weights, and probabilistic assessments; - **model building** — influence diagrams, decision trees, Bayesian networks; - **sensitivity analysis** — determining which model parameters materially affect the choice; - **value of information** — assessing whether it is worth gathering additional data before making the decision. The output of decision analysis is not an automatic answer but a structured justification of choice, enhancing its transparency and reproducibility. Decision analysis methods are applied in medicine, energy, risk assessment, strategic management, and public policy. The practical software embodiment of decision analysis is the **decision support system** (DSS) — an integrated platform combining models, databases, and a DM interface. ## Multiple criteria decision analysis In most practical problems, alternatives are evaluated on several — typically conflicting — criteria. Multiple criteria decision analysis (MCDA / MCDM) is one of the most highly developed branches of decision theory. The central concept is **Pareto optimality**: an alternative is Pareto optimal (non-dominated) if no other alternative is at least as good on all criteria and strictly better on at least one. The set of Pareto-optimal alternatives forms the efficient frontier, from which the DM makes the final selection. The main methods include: - **Analytic Hierarchy Process (AHP)** (T. Saaty) — hierarchical decomposition of the problem with pairwise comparisons of alternatives and criteria; - **ELECTRE family** (B. Roy) — sequential elimination of alternatives based on outranking relations (concordance / discordance); - **PROMETHEE** (J.-P. Brans) — ranking of alternatives based on pairwise preference flows; - **TOPSIS** (C.-L. Hwang, K. Yoon, 1981) — selection of the alternative closest to the ideal solution and farthest from the anti-ideal; - **VIKOR** (S. Opricovic, 1998) — compromise ranking based on closeness to the ideal solution and maximum group agreement; - **Multi-Attribute Utility Theory (MAUT)** (R. L. Keeney, H. Raiffa, 1976) — construction of an additive or multiplicative utility function over multiple criteria. ## Social choice When a decision is made by a group, the problem of aggregating individual preferences into a collective preference arises. The central result in this area is **Arrow's impossibility theorem** (K. Arrow, 1951): no aggregation rule can simultaneously satisfy several natural requirements (unrestricted domain, unanimity, independence of irrelevant alternatives, non-dictatorship). The Gibbard–Satterthwaite theorem complements this result by showing that any non-dictatorial voting procedure with three or more alternatives is susceptible to strategic manipulation. Relaxing Arrow's conditions has permitted the construction of theories that circumvent the impossibility. A. Sen (1970) showed that even a minimal requirement of individual liberties conflicts with Pareto optimality (the liberal paradox). J. Harsanyi (1955) showed that an impartial observer, equally likely to find themselves in the position of any participant, arrives at the maximization of the sum of individual utilities (a utilitarian justification). Practical methods of social choice include various voting procedures: plurality voting (relative majority), the Borda count (rank-order scoring), Condorcet's rule (the pairwise majority winner, if one exists), and approval voting. Each procedure has its advantages and vulnerabilities: the Condorcet paradox shows that pairwise majority can be cyclic, and the Gibbard–Satterthwaite theorem proves the inevitability of strategic manipulation. Expert methods (brainstorming, the Delphi method, panel methods) complement formal procedures in situations where preferences are difficult to formalize. ## Place within the sciences Decision theory emerged as a distinct field of knowledge in the second half of the twentieth century in close connection with the development of systems analysis, cybernetics, control theory, economics, and applied mathematics. It integrates ideas and methods drawn from: - [systems analysis](https://systems-analysis.info/eng/Systems_analysis "Systems analysis"); - [operations research](https://systems-analysis.info/eng/Operations_research "Operations research"); - mathematical statistics and statistical decision theory; - game theory; - optimal control theory; - computer science and artificial intelligence; - cognitive psychology; - behavioral economics; - philosophy of science and epistemology. Each of these disciplines approaches the problem of choice from its own methodological perspective, as applied to objects of varying nature — from technical systems to human behavior. The result is an integrative scientific field that studies decision-making as a universal process of purposeful action. The theory does not replace the process of thought or the volitional act of choice, but provides instruments for its rationalization and justification. ## External links - Decision theory — Wikipedia ## See also - [Operations research](https://systems-analysis.info/eng/Operations_research "Operations research") - [Systems analysis](https://systems-analysis.info/eng/Systems_analysis "Systems analysis") ## Bibliography - Bernoulli, D. (1738). Specimen Theoriae Novae de Mensura Sortis. *Commentarii Academiae Scientiarum Imperialis Petropolitanae*, 5, 175–192. (English trans.: *Econometrica*, 1954, 22(1), 23–36.) - von Neumann, J., & Morgenstern, O. (1944). *Theory of Games and Economic Behavior*. Princeton University Press. (EU axiomatics in 2nd ed., 1947.) - Savage, L. J. (1954). *The Foundations of Statistics*. Wiley. - Ramsey, F. P. (1926). Truth and Probability. In *The Foundations of Mathematics and Other Logical Essays* (pp. 156–198). London: Kegan Paul, 1931. - de Finetti, B. (1937). La prévision: ses lois logiques, ses sources subjectives. *Annales de l'Institut Henri Poincaré*, 7(1), 1–68. - Anscombe, F. J., & Aumann, R. J. (1963). A Definition of Subjective Probability. *Annals of Mathematical Statistics*, 34(1), 199–205. - Wald, A. (1950). *Statistical Decision Functions*. Wiley. - Arrow, K. J. (1951). *Social Choice and Individual Values*. Yale University Press. - Luce, R. D., & Raiffa, H. (1957). *Games and Decisions: Introduction and Critical Survey*. Wiley. - Simon, H. A. (1955). A Behavioral Model of Rational Choice. *Quarterly Journal of Economics*, 69(1), 99–118. - Harsanyi, J. C. (1955). Cardinal Welfare, Individualistic Ethics, and Interpersonal Comparisons of Utility. *Journal of Political Economy*, 63(4), 309–321. - Harsanyi, J. C. (1967–68). Games with Incomplete Information Played by "Bayesian" Players, I–III. *Management Science*, 14(3), 159–182; 14(5), 320–334; 14(7), 486–502. - Jeffrey, R. C. (1965). *The Logic of Decision*. McGraw-Hill. (2nd ed., University of Chicago Press, 1983.) - Raiffa, H. (1968). *Decision Analysis: Introductory Lectures on Choices under Uncertainty*. Addison-Wesley. - Teller, P. (1973). Conditionalization and Observation. *Synthese*, 26(2), 218–258. - Lewis, D. (1999). Why Conditionalize? In *Papers in Metaphysics and Epistemology* (pp. 403–407). Cambridge University Press. - Keeney, R. L., & Raiffa, H. (1976). *Decisions with Multiple Objectives: Preferences and Value Tradeoffs*. Wiley. (Reprinted: Cambridge University Press, 1993.) - Kahneman, D., & Tversky, A. (1979). Prospect Theory: An Analysis of Decision under Risk. *Econometrica*, 47(2), 263–291. - Saaty, T. L. (1980). *The Analytic Hierarchy Process*. McGraw-Hill. - Berger, J. O. (1985). *Statistical Decision Theory and Bayesian Analysis* (2nd ed.). Springer. - Bell, D., Raiffa, H., & Tversky, A. (Eds.). (1988). *Decision Making: Descriptive, Normative, and Prescriptive Interactions*. Cambridge University Press. - Schmeidler, D. (1989). Subjective Probability and Expected Utility without Additivity. *Econometrica*, 57(3), 571–587. - Gilboa, I., & Schmeidler, D. (1989). Maxmin Expected Utility with Non-Unique Prior. *Journal of Mathematical Economics*, 18(2), 141–153. - Tversky, A., & Kahneman, D. (1992). Advances in Prospect Theory: Cumulative Representation of Uncertainty. *Journal of Risk and Uncertainty*, 5(4), 297–323. - Sen, A. K. (1970). *Collective Choice and Social Welfare*. Holden-Day. (Expanded ed., Penguin, 2017.) - Bellman, R. E. (1957). *Dynamic Programming*. Princeton University Press. - Nash, J. F. (1950). Equilibrium Points in N-Person Games. *Proceedings of the National Academy of Sciences*, 36(1), 48–49. - Klein, G. (1998). *Sources of Power: How People Make Decisions*. MIT Press. - Peterson, M. (2017). *An Introduction to Decision Theory* (2nd ed.). Cambridge University Press. - Parmigiani, G., & Inoue, L. (2009). *Decision Theory: Principles and Approaches*. Wiley. - Gilboa, I. (2009). *Theory of Decision under Uncertainty*. Cambridge University Press. - Howard, R. A., & Abbas, A. E. (2015). *Foundations of Decision Analysis*. Pearson. - French, S. (1986). *Decision Theory: An Introduction to the Mathematics of Rationality*. Ellis Horwood.