Bounded Rationality

Classical economics painted a picture of the rational agent: a person who considers all available information, correctly calculates probabilities, weighs costs and benefits without error, and chooses the option that maximizes their utility. This model is mathematically elegant and produces clean predictions. It is also a description of a creature that does not exist. In 1955, the economist and cognitive scientist Herbert Simon introduced the concept of bounded rationality — the idea that human decision-making is rational, but within limits. Those limits are not failures of human dignity; they are the inevitable constraints of operating in a complex world with finite time and finite cognitive resources.

The Three Bounds

Simon identified three fundamental constraints on rationality. Information bounds: Decision-makers rarely have access to complete information about options, consequences, or probabilities. In most real decisions, gathering complete information would cost more — in time and effort — than the decision is worth. A student choosing a college cannot research every aspect of every institution. A physician making a diagnosis cannot order every conceivable test. Decision-making must proceed under genuine uncertainty. Cognitive capacity bounds: Even with complete information, the human mind cannot hold and process it all simultaneously. Working memory is limited (roughly four chunks of information at a time in modern estimates). Attention is finite and selective. Complex calculations — like computing the true expected utility of a job offer that involves uncertain salary growth, uncertain quality of life, and uncertain career outcomes over decades — exceed what the mind can carry out without aids. Time bounds: Real decisions have deadlines. A firefighter deciding how to enter a burning building cannot run an exhaustive analysis of all structural options. A trader deciding whether to sell a position while the market is moving cannot wait for certainty. Time pressure is a genuine constraint, and the rational response to time pressure is not to attempt perfect analysis (which is impossible) but to find strategies that produce good-enough decisions quickly.

Simon's central empirical claim was that people satisfice rather than optimize. Optimization means finding the best possible option from all available options — the global maximum. Satisficing (a portmanteau of satisfy and suffice) means searching through options until you find one that meets an acceptable threshold — one that is good enough — and stopping there. Consider job searching. An optimizer would evaluate every possible job, assign precise utility values to each, and select the maximum. A satisficer establishes an aspiration level ('I want a job in my field, with a salary above X, in a city I'd be willing to live in') and accepts the first offer that meets those criteria. The satisficing strategy will almost never find the globally optimal job — but it will find a good job reliably, at a cost in time and cognitive effort that is actually sustainable. Critically, the aspiration level is adjustable. If no jobs meeting the threshold appear, the satisficer lowers the threshold. If many options easily exceed the threshold, they raise it. This dynamic adjustment makes satisficing adaptive: it tracks the real distribution of available options rather than a fixed abstract ideal.

Gerd Gigerenzen and Ecological Rationality

Simon's framework was extended by the psychologist Gerd Gigerenzen, who argued more forcefully for the rationality of heuristics. Gigerenzen's concept of ecological rationality claims that a heuristic is rational not in the abstract but relative to the environment in which it is used. A strategy that ignores most available information and uses only one good cue can outperform complex optimization strategies when data is scarce, when the future is uncertain, or when overfitting to available data would lead to poor generalization. Gigerenzen's research group demonstrated this with the 'take the best' heuristic: when comparing two options on a set of features, look up the features in order of their predictive validity, and as soon as one feature distinguishes between the options, choose the option with the better value on that feature — and stop. Ignore all other features entirely. In experiments across many real-world prediction tasks, 'take the best' performed as well as or better than complex multiple-regression models, particularly with small samples. The reason: complex models are prone to overfitting noise in small datasets, while a simple one-cue strategy does not. The heuristic's simplicity is a feature, not a bug. This connects directly to AI and machine learning: the phenomenon of overfitting in machine learning models mirrors the logic Gigerenzen identified. A model with too many parameters fits training data perfectly but performs poorly on new data. A simpler model generalizes better. The principle that 'less can be more' in cognition and in algorithmic design is the same insight from different directions.