Field station 01 · bounded minds, branching worlds

Every choice grows a forest you cannot fully search.

Decision trees make choices visible: a question, a fork, a stopping place. Bounded rationality asks why real people rarely inspect every branch—and why a smaller path can be either a practical intelligence or a dangerous shortcut.

  • No account
  • No tracking
  • State stays in this browser
The fork and the canopy A central path divides. One side grows many interconnected branches; the other takes a short route to a bright exit.

01 / Live understory

The Branching Choice Forest

Two fictional cooling pilots compete for one small grant. You control which cues enter attention, how much search you can afford and what counts as a decisive difference. The exhibit compares two transparent rules over the same made-up data.

Illustrative · unvalidated

This is a learning experiment, not a psychological test, policy recommendation or model of you. Its scores are invented to expose mechanics.

The decision

Fund one neighbourhood heat-relief pilot

Banyan Canopyfixed shade at one busy square
Lichen Libraryportable shade kits loaned locally

Set your value lens, attention budget and stopping threshold, then reveal cue cards in any order. Run the comparison to see two explainable outputs.

1. Choose a value lens
4. Spend attention: reveal cues in your chosen order

0 of 3 attention tokens used

Reveal at least one cue, then compare. Your explanation will grow here.

Local-only state: settings are stored in your browser’s local storage. Nothing is transmitted. Reset clears it.

02 / Field notes

The tree is not the territory

A decision tree is a representation of questions and consequences. Bounded rationality is a theory family about choosing when information, time, computation and attention are limited. Their overlap is useful—but only if we keep the meanings straight.

Established

Bounded rationality

Herbert Simon challenged models that assume a chooser can enumerate every option, know every consequence and optimise without search cost. His 1955 model replaced that impossible view with organisms whose environments and computational capacities jointly shape choice. “Bounded” does not mean irrational. It means that the procedure is part of the explanation.

Established

Satisficing

A satisficing rule searches until it finds an option that clears an aspiration level: good enough on the dimensions that matter. The aspiration can move with experience, scarcity and what the search reveals. Satisficing is not automatically lazy or wise; its quality depends on the threshold, the order of search and the cost of a missed alternative.

Established

Decision tree

In a designed decision aid, a tree is a sequence of if–then questions. In machine learning, a tree is fitted from examples by choosing splits that partition a target outcome; Quinlan’s ID3 and the CART tradition are foundational forms. A transparent path can still encode poor labels, biased samples or an unsafe objective.

The causal story in this exhibit

What changes a result?

  1. 01Attention selects cues.

    Uninspected information cannot affect either displayed result.

  2. 02A rule integrates them.

    One rule weights and adds; the other searches and stops.

  3. 03Order meets threshold.

    The shortcut can exit before contradictory evidence appears.

  4. 04An output follows.

    The recommendation is fully caused by fictional ratings, your settings and the stated arithmetic—not a hidden profile.

Perspective A · error and bias

Shortcuts can predictably mislead

The heuristics-and-biases programme associated with Amos Tversky and Daniel Kahneman documents systematic departures from normative benchmarks in judgment under uncertainty. Availability can make memorable events seem more frequent; anchoring can make an arbitrary starting value exert too much pull. From this perspective, a stopping rule deserves scrutiny because ignored evidence may matter, especially when stakes are high and base rates or probabilities are knowable.

The strength of this lens is diagnostic: it asks which benchmark was violated and whether the error repeats. Its limit is that laboratory tasks may not reproduce the structure, feedback or learned skill of a real environment. A deviation from one formal norm does not, by itself, prove that a person’s process is foolish in every context.

Perspective B · ecological fit

Ignoring information can be adaptive

Fast-and-frugal and ecological-rationality research asks when a simple rule fits the statistical structure of an environment. If samples are small, relationships unstable or information expensive, estimating many weights can fit noise. A one-reason rule may be more robust, cheaper to communicate and easier to audit. The claim is conditional, not mystical: a heuristic succeeds when its building blocks match the environment.

The strength of this lens is that it treats search cost and environmental uncertainty as real. Its limit is the danger of celebrating simplicity without specifying where it works. “Less can be more” is an empirical hypothesis for a defined task and comparison set—not permission to ignore inconvenient evidence.

These perspectives compete most usefully when they make their benchmark, environment and loss function explicit. Is the goal logical consistency, calibrated probability, fewer false negatives, speed, accountability—or some negotiated combination?

03 / Two documented clearings

Where simple paths met real institutions

Neither case establishes a general law. Each shows how a transparent decision procedure interacted with a particular problem, dataset and organisation.

CASE 01Rural hospitalMichigan · 1997 report
Documented study

Chest-pain triage: the tool changed attention before it changed calculation

Lee Green and David Mehr studied decisions about placing patients with suspected acute cardiac ischaemia in a coronary care unit or a monitored nursing bed at a 100-bed rural hospital. Residents were shown probability charts from a Heart Disease Predictive Instrument, then encouraged to consult the cards during alternating weeks. The trial-year analysis included 89 patients, alongside earlier and later comparison periods.

The card’s actual use did not change coronary-care utilisation during the intervention. Yet utilisation had already fallen after a conference introduced the instrument, before the alternating-week trial. The authors proposed that learning the relevant clinical factors may have focused physicians’ attention, allowing a simpler practical strategy without explicit probability calculation.

What can be learned: a decision aid may work through changed cue attention, not only through literal computation. What cannot be concluded: this small, single-service study does not show that a simple tree is safe for all cardiac triage, nor that contemporary clinicians should replace validated protocols. The causal interpretation was proposed by the authors and the timing complicated the planned trial.

Read the PubMed record
CASE 02Bank regulationUnited Kingdom · 2014 paper
Documented analysis

Bank failure: when a sparse tree competed with richer models

A Bank of England Financial Stability Paper by David Aikman and colleagues examined simplicity versus complexity in financial regulation. Using the global financial crisis as an empirical setting, the authors reported that some simple indicators outperformed more complex risk-weighted metrics for predicting individual bank failure. When cues were combined, fast-and-frugal decision trees performed comparably to more information-intensive regression approaches in their analysis.

The proposed mechanism is a bias–variance and communication trade-off. With limited observations and fat-tailed, changing risks, a flexible model can estimate noise as if it were signal. A sparse tree estimates fewer relationships and makes its alert logic inspectable. But an interpretable rule can still miss novel failure modes, and historical crisis performance does not guarantee future calibration.

What can be learned: model complexity should earn its keep out of sample, and communicability matters in governance. What cannot be concluded: simple indicators always dominate, or one crisis supplies a timeless regulatory recipe. The paper is a policy research analysis, not a randomised intervention.

Read the Bank of England paper page

04 / Evidence, not folklore

Prune four tempting myths

Myth: more information always improves a choice.

Evidence: information helps only if it is relevant, reliable and integrated without excessive cost or overfitting. Sometimes more cues help; sometimes they add noise.

Myth: a decision tree is objective because every branch is explicit.

Evidence: people choose the target, training data, cue order, thresholds and error costs. Visibility enables critique; it does not remove values.

Myth: satisficing means settling for a bad outcome.

Evidence: satisficing means stopping when an aspiration level is met. Whether the level is defensible depends on stakes, alternatives and search cost.

Myth: a simple rule is naturally fair.

Evidence: simplicity can aid audit, yet a single discriminatory cue or unequal threshold can reproduce harm with great efficiency.

05 / The ethics of the cut

Every branch excludes a future

Trees feel accountable because a person can trace a route. That advantage is real but incomplete. In hiring, credit, health, education or public benefits, the leaf of a tree can distribute opportunity. Before deployment, ask who defined success, whose records became training data, which errors are tolerated and whether affected people can understand and contest the result.

Do not use a teaching tree as a diagnostic instrument. Do not infer motives or stable traits from a path someone took under artificial constraints. For high-impact decisions, a transparent rule should sit inside governance: domain validation, subgroup error analysis, human review with genuine authority, documentation, monitoring for drift and an appeal channel. A human in the loop is not a safeguard if they merely rubber-stamp the output.

Sometimes the ethical action is to refuse the tree. If the target is illegitimate, the labels unreliable or the institution unwilling to repair harms, better prediction can intensify the wrong system.

06 / Guided reflection · 12 minutes

Map a decision you actually face

  1. Name the decision without hiding a conclusion.

    Write “choose a transport plan for six months,” not “justify buying a car.”

  2. Separate options, cues and outcomes.

    An option is what you can do; a cue is evidence available now; an outcome arrives later.

  3. Set one non-negotiable guardrail.

    Identify a threshold that protects against unacceptable harm. Say who defined it.

  4. Write two rival procedures.

    One weighs several cues. One checks them in order and stops. Predict where they disagree.

  5. Price one more branch.

    What would added information cost in time, money or delay? What error might it prevent?

  6. Schedule a review.

    Record what evidence would make you revise the rule rather than defend the result.

Final prompt: Which limit is shaping this choice—missing information, limited attention, irreducible uncertainty, or an institutional deadline?

Boundary marker

What this does not prove

Your route through the Branching Choice Forest does not measure rationality, intelligence, personality, values or decision skill. A changed result does not show that one strategy is universally superior. The pilot ratings, weights and stopping rule are fictional; the exhibit has not been validated against real funding decisions. It demonstrates a conceptual dependency: when attention, cue order, weights or thresholds change, a transparent procedure can change its output. Transfer to any real decision requires relevant evidence, error costs and domain review.

07 / Questions from the trail

Frequently asked, carefully bounded

Are decision trees the same as human decisions?

No. A tree is a representation or procedure. Human choice also involves perception, memory, emotion, social context, habits and goals that may change during deliberation. A tree can externalise part of a decision and make assumptions inspectable; it should not be mistaken for a complete model of a mind.

Does bounded rationality mean people are bad at thinking?

No. It rejects the idea that unbounded optimisation is a realistic description of ordinary choice. People and institutions use procedures adapted to limits. Some procedures produce systematic errors; others exploit useful environmental regularities. Evaluation needs a task, a benchmark, an environment and a cost function.

When is a fast-and-frugal tree a good idea?

It is a candidate when time or data are scarce, a few cues carry stable signal, errors and stopping rules can be specified, and transparency is valuable. It still needs comparison against credible alternatives on unseen data. High stakes increase—not reduce—the need for validation and monitoring.

Why can a complex model perform worse than a simple one?

A flexible model can learn accidental patterns in a small or shifting sample. This is overfitting: excellent recall of the past, weak transfer to new cases. Simpler models restrict what can be fitted, sometimes lowering variance at the cost of some bias. “Can” matters: abundant representative data and stable relationships may reward complexity.

Is an interpretable tree automatically trustworthy?

No. Interpretability tells us that a route can be inspected. Trust also requires legitimate goals, accurate data, appropriate validation, fair error distribution, security, monitoring and recourse. A clearly explained harmful rule remains harmful.

Can I use the exhibit’s result for a real project?

Use it as a worksheet for questions, never as a recommendation. The scenario and ratings are invented. For a real project, gather context-specific evidence, consult affected people, define error costs and test any proposed rule. Health, legal, credit or employment decisions need qualified domain governance.

08 / Further reading

A trail of primary and scholarly sources

Links and bibliographic details below were checked against publisher, journal, institutional or National Library of Medicine records on 28 October 2026.

  1. 1955

    Simon, Herbert A. “A Behavioral Model of Rational Choice.” The Quarterly Journal of Economics 69(1): 99–118. The foundational formal statement behind bounded rationality and satisficing. Oxford Academic

  2. 1974

    Tversky, Amos, and Daniel Kahneman. “Judgment under Uncertainty: Heuristics and Biases.” Science 185(4157): 1124–1131. The classic programme on systematic judgment errors. DOI record

  3. 1984

    Breiman, Leo, Jerome H. Friedman, Richard A. Olshen, and Charles J. Stone. Classification and Regression Trees. The foundational CART monograph on fitted tree models, pruning and honest accuracy estimates. Stanford record

  4. 1986

    Quinlan, J. R. “Induction of Decision Trees.” Machine Learning 1: 81–106. Describes ID3, information-based splitting, noise and incomplete data. Springer Nature

  5. 1996

    Gigerenzer, Gerd, and Daniel G. Goldstein. “Reasoning the Fast and Frugal Way: Models of Bounded Rationality.” Psychological Review 103(4): 650–669. Formalises one-reason decision mechanisms and tests comparative performance. DOI record

  6. 1997

    Green, Lee A., and David R. Mehr. “What Alters Physicians’ Decisions to Admit to the Coronary Care Unit?” Journal of Family Practice 45(3): 219–226. A prospective decision-support study and the first case discussed above. PubMed

  7. 2003

    Kahneman, Daniel. “Maps of Bounded Rationality: Psychology for Behavioral Economics.” American Economic Review 93(5): 1449–1475. A major synthesis of intuitive judgment, accessibility and choice. American Economic Association

  8. 2011

    Gigerenzer, Gerd, and Wolfgang Gaissmaier. “Heuristic Decision Making.” Annual Review of Psychology 62: 451–482. Reviews formal heuristic models and the conditions under which ignoring information can help or harm. Annual Reviews

  9. 2014

    Aikman, David, et al. “Taking Uncertainty Seriously: Simplicity versus Complexity in Financial Regulation.” Bank of England Financial Stability Paper No. 28. The source for the banking case above. Bank of England

You have reached an exit, not the end.

Carry a better question out of the forest.