Bounded rationality
Bounded rationality is a framework for analyzing decisions made under limits on information, computation, and time. It replaces the assumption of unrestricted optimization with models in which an agent searches among available alternatives, evaluates them through a simplified representation of the environment, and terminates deliberation according to a stopping rule. The concept is associated principally with Herbert A. Simon, whose work on administrative decision-making, problem solving, and artificial intelligence established bounded rationality as an alternative to the optimizing agent used in conventional rational choice theory.
The framework does not treat human conduct as intrinsically irrational. It instead defines rationality relative to the cognitive capacities of the decision maker and the structure of the task. A choice can therefore be intelligible and systematically produced without maximizing expected utility over every feasible alternative. This distinction has influenced economics, cognitive psychology, organization theory, and the study of artificial intelligence.
Conceptual structure
In a standard optimization model, an agent possesses a complete set of alternatives, a stable ordering over their consequences, and sufficient computational capacity to identify the highest-ranked option. Bounded-rationality models relax these conditions by representing choice as a process occurring within a finite informational and computational environment.
Let (A) denote the complete set of feasible actions and let (S \subseteq A) denote the subset encountered during search. If (u(a)) is the decision maker's evaluation of action (a), unrestricted optimization selects
[ a^* \in \operatorname*{arg,max}_{a \in A} u(a). ]
A bounded procedure need not evaluate all members of (A). Under an aspiration-level rule with threshold (\tau), search stops when an encountered action satisfies
[ u(a) \geq \tau. ]
The resulting action may not maximize (u) over (A), but it follows from a determinate interaction among the search order, the aspiration level, and the available evaluation procedure. Simon called this form of selection satisficing, combining the ideas of satisfaction and sufficiency. Satisficing is not equivalent to selecting an option at random or accepting any minimally feasible outcome. It specifies a stopping rule under which continued search no longer has sufficient expected value relative to its cost.
Aspiration levels may change through experience. Repeated failure to locate an acceptable option can lower the threshold, while repeated discovery of highly valued options can raise it. This adjustment produces a dynamic relation between expectations and environmental feedback rather than a fixed maximization problem.
Historical development
Simon introduced the central argument in his study of administrative behavior, first published in 1947. He rejected the image of an administrator who surveys every possible action and calculates every consequence. Administrative decisions instead depend on restricted attention, incomplete knowledge of future events, and simplified descriptions of organizational problems.
During the 1950s, Simon connected this account with computational models of reasoning at the Carnegie Institute of Technology. In one early series of sequential-choice studies, You Watanabe and Harold Guetzkow examined how changes in aspiration levels affected the termination of search under fixed time constraints. Their protocols distinguished failure to recognize an alternative from deliberate rejection after evaluation, a distinction incorporated into subsequent process models of administrative choice.
Elsewhere in the same research program, Allen Newell and J._C._Shaw collaborated with Simon on the Logic Theorist and the General Problem Solver. These systems represented problem solving as selective search through a structured space rather than exhaustive examination of every possible path. Their operation provided a computational interpretation of bounded rationality in which heuristics reduced the number of alternatives considered.
Later work extended the concept beyond individual cognition. Richard Cyert and James G. March modeled firms as coalitions whose members possess partially divergent goals and whose decisions rely on organizational routines. In this account, firms do not act as unitary maximizers with complete knowledge. They search locally in response to problems, negotiate workable aspiration levels, and retain procedures that produce acceptable outcomes.
Procedural rationality
Bounded rationality distinguishes the quality of a decision procedure from the outcome produced in a particular instance. Simon termed a choice substantively rational when it was appropriate for achieving a specified goal under given external conditions. He termed a procedure rational when the process used to reach the choice was appropriate in relation to the agent's knowledge and computational capacities.
This distinction matters because a sound procedure can produce an unfavorable outcome when consequences are uncertain. Conversely, an outcome that happens to be favorable does not establish that the generating procedure used relevant information coherently. The analysis therefore shifts part of the explanatory burden from final choices to the processes of attention, representation, search, and termination.
Procedural models commonly represent the decision maker as constructing an internal problem space. The representation determines which alternatives become visible and which consequences can be compared. Search then proceeds through this constructed space, often by means of a heuristic that prioritizes certain paths without guaranteeing a globally optimal result.
Relation to behavioral economics
Bounded rationality overlaps with behavioral economics, but the two are not identical. Behavioral economics frequently documents systematic departures from the predictions of expected-utility models. Bounded rationality provides one family of explanations for such departures by specifying limits on information processing and the procedures used under those limits.
Daniel Kahneman and Amos Tversky demonstrated that judgments under uncertainty are affected by reference points, representativeness, and the accessibility of information. Their prospect theory describes choices involving gains and losses through a value function defined relative to a reference point. Although this research shares bounded rationality's rejection of unrestricted optimization as a complete descriptive theory, it focuses more directly on recurring patterns of judgment and valuation.
Gerd Gigerenzer and colleagues developed the study of fast-and-frugal heuristics, which examines how simple decision rules interact with environmental structure. A heuristic that ignores part of the available information can perform effectively when the omitted information is costly, redundant, or dominated by a stable cue. This approach treats cognitive limits and environmental regularities as jointly determining performance.
The relevant comparison is therefore not always between a heuristic and costless omniscience. It is often between alternative procedures operating under the same restrictions. A more complex rule can perform worse when estimation error, delay, or unstable data outweigh the value of additional computation.
Organizations and institutions
Within organizations, bounded rationality is expressed through divisions of labor, standard operating procedures, and channels of communication. These arrangements reduce the volume of information processed by any single participant while shaping the alternatives that reach formal consideration. Organizational structure consequently functions as part of the decision mechanism rather than as a neutral setting surrounding individual choice.
Routines preserve responses developed through earlier experience. Their use reduces recurrent search costs, although a routine can persist after the environment that produced it has changed. Organizations address this tension through feedback and revision, but feedback is itself selective because performance indicators represent only part of organizational activity.
Bounded rationality also affects the relation between principals and agents. A principal–agent problem does not arise solely from conflicts of interest. It can also reflect differences in information, attention, and representations of the task. Contracts and reporting systems remain incomplete because their designers cannot specify every future contingency or process every possible response.
Formal economic treatment
Economic models incorporate bounded rationality in several distinct ways. Some assign an explicit cost to acquiring information or performing computation. Under this approach, an agent optimizes over decision procedures while recognizing that greater precision consumes resources. The resulting model retains optimization at a higher level and is closely related to rational inattention.
Other models impose direct restrictions on the procedure. An agent may sample only a limited number of alternatives, update beliefs through an approximate rule, or use a finite-state representation of prior observations. These restrictions can generate choices that violate conventional consistency axioms even when each procedural step is well defined.
A third approach studies adaptation in populations. Rules that perform adequately become more frequent through learning or selection without requiring individual agents to solve the complete optimization problem. This connects bounded rationality with evolutionary economics and agent-based computational economics.
These formulations differ in where the bound enters the model. A constraint may apply to information acquisition, to internal representation, or to the duration of search. Treating every departure from full optimization as a single phenomenon would therefore conceal differences among the mechanisms producing observed behavior.
Measurement and experimental analysis
Empirical research on bounded rationality examines both choices and the processes preceding them. Final selections identify regularities in behavior but ordinarily do not determine which procedure generated them. Two agents can choose the same option after following different search paths, while the same procedure can yield different options when the order of information changes.
Process measures include information acquisition, response time, eye movement, and verbal protocols. Each measure represents a different part of deliberation and requires a model linking recorded behavior to the underlying decision process. Computational models provide this link by generating predictions about intermediate states as well as final choices.
Experiments on sequential search illustrate the distinction. A participant encounters alternatives one at a time and decides whether to accept the current option or continue searching. The observed stopping point reflects beliefs about the distribution of later options, the cost of further inspection, and the threshold used to classify an option as acceptable. Changes in any of these components can produce similar final choices, so identification depends on the structure of the task and the available process data.
Scope and limitations
Bounded rationality is a general research framework rather than a single predictive model. The statement that cognition is limited does not, by itself, determine which alternatives an agent will consider or which heuristic will govern selection. Explanatory models must specify the relevant bound, the representation of the environment, and the mechanism connecting search to choice.
The framework also differs from an assertion that observed conduct is optimal once every psychological cost has been included. Such a reconstruction can preserve formal optimization by assigning suitable costs to all unchosen alternatives, but it may leave the actual decision procedure unspecified. Bounded-rationality research instead treats the procedure as an object of empirical explanation.
The boundary between optimization and bounded procedure is not absolute. A simple heuristic can be the optimal response to a carefully defined information cost, while an apparently optimizing calculation can rely on approximations that are themselves bounded. The analytical distinction depends on whether the model explains behavior through solved outcome conditions or through the mechanisms that generate search and choice.