Bounded Rationality — When Optimization Fails

The standard model of human decision-making assumes that people evaluate all available options, predict their consequences, and choose the one that maximizes their outcome. This model is elegant. It is also wrong.

Herbert Simon identified the problem in 1947, in a book that would eventually earn him the Nobel Prize in Economics. The model of man used by economists, he wrote, posits “complete information, unlimited time and computational capacity, and an unambiguous ordering of goals.” None of these conditions holds in the real world. People do not have complete information. They do not have unlimited time or computational capacity. Their goals are often ambiguous and conflicting.

Simon called the gap between the standard model and actual behavior “bounded rationality.” The term captures a simple observation: rationality is constrained. Any decision-maker operates with limited information, finite computational resources, and a clock that does not stop. The rational choice, in the full optimization sense, is often impossible to compute.

The optimization model

The optimization model of decision-making appears in economics, game theory, and much of modern AI. It works like this. A decision-maker faces a set of options. Each option has a known or estimable outcome. The decision-maker assigns a value to each outcome and selects the option with the highest value.

The model requires three things. The options must be enumerated. Their outcomes must be predictable. The values must be comparable. When these conditions hold, optimization is straightforward. A sorting algorithm finds the maximum. A linear program finds the best allocation of resources. A reinforcement-learning agent finds the policy that maximizes expected reward.

The model breaks when the space of options is too large to enumerate, when outcomes are too uncertain to predict, or when the values are too complex to compute. A chess player cannot evaluate every possible game to its conclusion. The number of possible games exceeds the number of atoms in the observable universe. A supply-chain manager cannot predict every future disruption. A physician cannot weigh every possible treatment against every possible long-term outcome.

The optimization model does not say these problems are unsolvable. It says that, in principle, the optimal solution exists and a sufficiently powerful agent could find it. Simon argued that this principle is irrelevant when the agent is human, or when the agent is any system with finite computational resources.

The 1947 formulation

Simon’s original formulation appeared in Administrative Behavior: A Study of Decision-Making Processes in Administrative Organization, published in 1947 by John Wiley & Sons. The book was short by academic standards. It contained roughly one hundred thousand words. Its influence was disproportionate to its size.

The central argument was that the organizations Simon studied – corporations, government agencies, universities – do not optimize. They use procedures, routines, and rules of thumb that produce acceptable outcomes without evaluating every possible alternative. Managers do not calculate the utility of every decision. They follow established patterns, modify them when they fail, and move on.

Simon wrote that “the selection of a course of action is just as much a question of finding or inventing the alternative selected as it is of appraising alternatives.” In other words, the process of generating options is as important as the process of evaluating them. The standard optimization model assumes the options are given. Simon showed that they are not.

The book also introduced the idea that decision-making is a search process through a space of possibilities. The search is guided by heuristics – rules of thumb that prune the space and focus attention on promising branches. The search stops when an acceptable option is found, not when the best option is identified.

Satisficing

Simon introduced the term “satisficing” in his 1955 paper “A Behavioral Model of Rational Choice,” published in the Quarterly Journal of Economics, and developed it further in his 1956 paper “Rationality as Process and as Product of Thought,” published in the Review of Economic Studies. The word is a portmanteau of “satisfice” and “satisfice,” but the meaning is not a compromise between satisfaction and optimization. It is a distinct decision criterion.

Satisficing works by establishing an aspiration level – a threshold of acceptability – and searching for the first option that meets or exceeds that threshold. The search stops when the threshold is met. The selected option is not necessarily the best available option. It is the first option that is good enough.

The aspiration level is not fixed. It adapts to experience. If options meeting the threshold are easy to find, the threshold rises. If options meeting the threshold are hard to find, the threshold falls. This adaptive mechanism is crucial. It means that satisficing does not produce a fixed level of performance. It produces a dynamic equilibrium between expectations and experience.

Simon demonstrated this through a series of thought experiments and empirical observations. A buyer who finds a satisfactory price for a product does not continue searching for the absolute lowest price, even if such a price might exist. The cost of continued search – time, effort, uncertainty – outweighs the marginal benefit of a better price. A job candidate who receives a satisfactory offer does not wait for the perfect offer, even if one might exist. The cost of waiting is a period of unemployment, which has its own costs.

The key insight is that satisficing is not a failure of rationality. It is a response to the computational constraints that any real decision-maker faces. Optimization is rational only when the cost of finding the optimum is less than the benefit of having it. When the cost exceeds the benefit, satisficing is the rational choice.

The computational view

Simon’s contribution was not just the observation that people do not optimize. It was the argument that any computational system – biological or artificial – must approximate optimization when the exact solution is intractable.

He developed this argument through his work on problem-solving and artificial intelligence. In the 1950s and 1960s, Simon and Allen Newell built some of the first artificial intelligence programs, including the Logic Theorist and the General Problem Solver. These programs did not search exhaustively through all possible solutions. They used heuristics to guide the search. The heuristics were not guaranteed to find the optimal solution. They were designed to find a good solution in a reasonable amount of time.

Simon wrote that “human problem-solving is a search process – a search through a problem space of possible states and operators.” The search is bounded by the finite memory and finite processing speed of the human mind. Any accurate model of human problem-solving must account for these bounds.

This was a radical shift in the study of cognition. Instead of asking what the optimal solution to a problem is, researchers began asking how people actually find solutions given their cognitive constraints. The field of cognitive science emerged from this shift. So did modern algorithm design.

Heuristics and algorithms

The distinction between heuristics and algorithms is central to bounded rationality. An algorithm is a procedure that guarantees a correct or optimal solution. A heuristic is a procedure that usually finds a good solution but does not guarantee optimality.

Simon argued that heuristics are not inferior to algorithms. They are adapted to the constraints of the systems that use them. A heuristic for routing delivery trucks may not find the shortest possible route, but it finds a route that is short enough and can be computed in seconds rather than hours. A heuristic for diagnosing disease may not identify every possible condition, but it identifies the most likely ones quickly enough to be useful.

The trade-off is explicit. Heuristics sacrifice guaranteed optimality for computational feasibility. Algorithms sacrifice feasibility for guaranteed optimality. Both are rational within their domains.

This trade-off appears throughout computer science. The A* search algorithm guarantees optimality when an admissible heuristic is used, but it can require exponential memory for large problems. Greedy best-first search uses less memory but does not guarantee optimality. Both are useful in different contexts. Linear programming finds the optimal solution to a resource-allocation problem, but it cannot handle non-linear constraints. Heuristic methods like genetic algorithms or simulated annealing can handle non-linear constraints but do not guarantee optimality.

The choice between heuristic and algorithm is not a technical detail. It is a fundamental design decision that reflects the constraints of the problem domain.

Bounded rationality in economics

The standard economic model of rational choice assumes that agents have stable preferences, consistent beliefs, and the ability to compute optimal decisions. Bounded rationality challenges all three assumptions.

Preferences are not always stable. They are constructed in the moment, influenced by framing effects, context, and the way options are presented. Beliefs are not always consistent. People hold contradictory beliefs and do not update them according to Bayesian principles. Optimal decisions are not always computable. Even when preferences and beliefs are well-defined, the computation required to find the optimal decision can exceed the resources of the decision-maker.

Simon did not argue that economics should abandon the optimization model entirely. He argued that the model should be supplemented with a theory of how decisions are actually made under constraints. The optimization model describes the environment in which decisions are made. Bounded rationality describes the decision-maker.

The distinction is important. An optimization model of a market assumes that all participants are rational. A bounded-rationality model of a market assumes that participants use heuristics that approximate rationality. Both models can produce similar aggregate outcomes. But they produce different predictions when the heuristics systematically deviate from rationality.

This is where behavioral economics entered the picture. Daniel Kahneman and Amos Tversky demonstrated that people systematically violate the axioms of rational choice through heuristics like availability, representativeness, and anchoring. Their work was complementary to Simon’s. Simon focused on the computational constraints that make optimization infeasible. Kahneman and Tversky focused on the systematic biases that heuristics introduce.

The connection to information theory

Bounded rationality shares a structural similarity with rate-distortion theory. Rate distortion asks how much information can be preserved when the communication channel has limited capacity. The answer is a trade-off between fidelity and rate. Higher fidelity requires higher rate. Lower rate produces lower fidelity.

Bounded rationality asks how close to optimal a decision can be when the decision-maker has limited computational capacity. The answer is a trade-off between fidelity and capacity. Higher fidelity requires higher capacity. Lower capacity produces lower fidelity.

The mathematics is similar. Both problems involve minimizing a distortion measure subject to a resource constraint. In rate distortion, the constraint is on bits. In bounded rationality, the constraint is on computation. Both problems show that the optimal strategy involves lossy compression of information.

This connection is not accidental. Both rate distortion and bounded rationality describe how systems operate under scarcity. Rate distortion describes scarcity of bits. Bounded rationality describes scarcity of computation. The same mathematical structure appears in both domains.

What bounded rationality does not say

Bounded rationality is often misunderstood. It does not say that people are irrational. It says that rationality is bounded. A satisficing decision-maker is rational within the bounds of their information and computational resources.

It does not say that optimization is never possible. It says that optimization is possible only when the problem is small enough, the constraints are simple enough, and the time horizon is short enough. For many real-world problems, these conditions do not hold.

It does not say that heuristics are always good enough. Some heuristics perform surprisingly well. The recognition heuristic – choosing the option you recognize over the one you do not – can outperform complex models in certain environments. Other heuristics perform poorly. The availability heuristic – judging probability by how easily examples come to mind – can produce severe errors when the most memorable examples are not the most probable.

The point is not that heuristics are always good or always bad. The point is that their performance depends on the environment. A heuristic that works well in one environment may fail catastrophically in another. Bounded rationality requires an understanding of both the decision-maker and the environment.

What remains uncertain

Whether bounded rationality can be formalized to the same degree as optimization is an open question. Optimization has a clean mathematical foundation. Linear programming, dynamic programming, and game theory provide precise solutions to well-defined problems. Bounded rationality has fewer precise results. The concept is clear. The mechanisms are described. But the mathematical formalism is less developed.

Gigerenzer and the Heuristics Project at the Max Planck Institute has argued that fast and frugal heuristics can outperform complex models in uncertain environments. This claim is controversial. Critics argue that the outperformance is limited to specific environments and does not generalize. Proponents argue that the outperformance reveals something fundamental about the relationship between heuristics and environmental structure.

Whether the adaptive aspiration mechanism in satisficing can be modeled with the same precision as utility maximization remains an open question. Simon described the mechanism qualitatively. He did not provide a complete mathematical theory of how aspiration levels adapt over time.

The relationship between bounded rationality and artificial intelligence is also unresolved. Modern AI systems, particularly deep learning, approximate optimization through gradient descent. They do not find the global optimum. They find a local optimum that is good enough for the task at hand. This is a form of bounded rationality. But the systems are not constrained by the same cognitive limitations that Simon studied. They are constrained by computational resources, training data, and architectural choices. Understanding how these constraints map onto Simon’s framework is an open area of research.

Sources

• Herbert A. Simon, Administrative Behavior: A Study of Decision-Making Processes in Administrative Organization (John Wiley & Sons, 1947, 4th ed. 1997). Original formulation of bounded rationality and the critique of the economic model of man.
• Herbert A. Simon, “A Behavioral Model of Rational Choice,” Quarterly Journal of Economics 69(1) (1955), 99-118. Formal introduction of bounded rationality as an alternative to expected-utility theory.
• Herbert A. Simon, “Rationality as Process and as Product of Thought,” Review of Economic Studies 43(1) (1976), 1-16. Development of satisficing, aspiration levels, and the computational view of decision-making. Nobel lecture.
• Herbert A. Simon and Allen Newell, “Computer Science as Empirical Inquiry: Symbols and Search,” Communications of the ACM 19(3) (1976), 113-123. The computational view of cognition and heuristic search.
• Amos Tversky and Daniel Kahneman, “Judgment Under Uncertainty: Heuristics and Biases,” Science 185(4157) (1974), 1124-1131. Systematic biases in human judgment and the relationship to bounded rationality.
• Gerd Gigerenzer and Peter M. Todd, Fast and Frugal Heuristics: The Smart Tools Series (MIT Press, 1999). The Max Planck Institute research program on fast and frugal heuristics and their performance in uncertain environments.