From “should” to “is”
Everything so far was normative: what an optimal agent should do. But our systems don't act alone — they advise people.
Decision support systems typically do not operate in isolation. They support the decision-making of humans — and users could ignore, misinterpret, or misuse a recommendation, even if it were perfectly rational.
So before we can support a human decision-maker, we have to answer a question we've dodged for four lectures: are humans even rational? The honest answer — backed by decades of experiments — is often not.
If we know humans are predictably irrational, we must model how they actually decide. Then we can design the computational part of a DSS to counterbalance their biases — so the whole system (human + machine) ends up rational, even if the human alone isn't. A DSS that ignores human psychology can be perfectly correct and still fail.
These limitations aren't disabilities — they're natural, a product of how our species evolved. Everyone has them. The job of behavioral decision theory is to map them.
Humans aren't always rational
Three famous experiments where real people reliably break the rules of expected utility.
① Preference reversal
Two lotteries: A = a moderate payoff with high probability (like a safe bond); B = a big payoff with low probability (like a lottery ticket). Ask people to choose and they pick A. Ask them to price each one and they price B higher. That's contradictory — a stable utility function would rank them the same way both times. The preference flips with how you ask: it's not representation-invariant.
② The Allais paradox
Violates the independence axiom (preferences shouldn't depend on an outcome that's identical across both options). In one framing people prefer a sure million to a gamble; in a mathematically equivalent framing they flip their preference. The "irrelevant" shared part turns out not to be irrelevant to humans.
③ The Ellsberg paradox — try it yourself
This one you can feel directly. Two urns of 100 balls. Pick a bet, then a second bet:
Win €1000 if you draw the winning colour. For each bet, pick the urn you'd draw from.
Most people prefer the urn with known odds in both bets. This is ambiguity aversion: humans distinguish between risk — known probabilities — and ambiguity — unknown probabilities — and tend to prefer known risks, even when the unknown option is the better bet.
These failures aren't universal — change the task and people may behave rationally. But the fact that they happen at all, predictably, forces us to revise the theory. Two readings: either humans are irrational, or our rationality axioms are too strong for any real agent. Either way, a DSS designer can't assume the user is a perfect expected-utility maximiser.
Bounded rationality
Maybe humans aren't broken — maybe "be perfectly optimal" was always an impossible bar. Herbert Simon's idea (it won him a Nobel).
Classical rationality says: consider every policy, compute each one's expected utility, pick the max. Bounded rationality keeps the goal but shrinks the menu — maximise expected utility only over the policies that are actually computable with the agent's limited resources.
Agents — human or machine — are limited by cognitive resources (not enough mental/computational power), time (a doctor can't run value iteration before treating a patient), and incomplete information (you rarely know the exact probabilities). Perfect rationality may simply be computationally impossible — recall that Nash equilibria are hard, and value iteration on a huge state space is infeasible.
Two cracks. (1) It's barely surprising — any algorithm on a computer is already restricted to computable policies. (2) Even if each policy is cheap, the act of maximising over a huge set of them can itself be infeasible. And it never says which policies count as "feasible." We need a model that drops global optimisation entirely.
Procedural rationality & heuristics
Ariel Rubinstein's shift: stop judging the outcome, study the procedure — the actual rule an agent uses to turn information into an action.
A state is described by cues — features (a patient's symptoms, a city's name). A heuristic is a simple decision rule that takes some cues and returns an action that approximately maximises utility. It replaces optimisation with a rule — no scanning of all options.
Four heuristics the professor highlights — tap each:
Satisficing
Set an aspiration level θ (the least you'll accept). Go through the options in order and pick the first one that's "good enough" (utility ≥ θ). Sequential search + early stopping. The order of options changes the result → models framing.
Take-the-best
Given cues one at a time, in order. Check them until you hit the first discriminating cue — enough to decide — then stop. (A doctor: cough? not enough → fever? now I can decide.) The order of cues matters.
Recognition
If you recognise one of the cues as a known situation, discard the rest and decide on that alone. "Which city is bigger — Madrid or Viterbo?" You recognise Madrid → you judge it bigger.
Elimination by aspects
Repeatedly pick a cue (by your own sense of importance) and eliminate every option that fails it, until one remains. Unlike take-the-best, you choose the order of cues.
They're cheap: roughly linear complexity, tiny memory, and small information requirements (they decide before seeing everything). A limited brain — or limited hardware — can actually run them. The price: they're only approximations.
Cognitive biases
A heuristic makes an error ε(s) versus the optimal choice. When that error is systematic — its average over situations isn't zero — the agent is biased.
If errors were random, they'd cancel out (sometimes you do better, sometimes worse). But heuristic errors lean one way, so they don't cancel — they accumulate into a bias. Biases are predictable, repeatable mistakes baked into how we think.
The professor groups them into families. You don't need to memorise the list — just recognise that shortcut-based thinking produces systematic errors. Tap for examples:
Judgement under uncertainty
Availability (judge probability by how easily it comes to mind — a plane crash you saw on the news). Representativeness (judge by stereotype — "a nurse is probably female"). Anchoring, overconfidence (Dunning–Kruger), optimism (underrate risks → skip insurance).
Probabilistic reasoning
Conjunction fallacy: rating "female nurse" as more likely than "nurse" — impossible, the subset can't beat the whole. Gambler's fallacy: thinking random events "self-correct" (red is due!). Confirmation bias.
Memory & attention
Framing effect (the wording changes the choice — exactly what satisficing/take-the-best predict). Recency bias (over-weight the latest info).
Temporal & social & choice
Sunk-cost fallacy & status-quo bias (temporal). Authority bias & halo effect (social). Ambiguity aversion, preference reversal, loss aversion (choice).
Ecological rationality
If heuristics are systematically bad, why do humans keep using them? Because in the right environment, they're not bad at all — sometimes they beat "optimal."
Rationality is not a property that holds in general. The rationality of an agent depends on the actual environment in which the agent is inserted. A heuristic that's bad on average may be near-optimal — or better — in the specific environments an agent really faces.
Judge a decision rule by its average performance over the distribution of environments the agent actually meets — not over all possible worlds. In environments that are uncertain, noisy, and changing (the real world!), heavily optimised procedures overfit; simple heuristics generalise and can win. There is no universally optimal rule.
Fast-and-frugal heuristics are the rules that win this way. Four traits: sequential information search · limited integration (only some cues) · early stopping · no global optimisation. Take-the-best and recognition both qualify.
A fast-and-frugal tree (ER triage)
The flagship example: a tiny decision tree where each cue can immediately exit, and only a few cues are ever checked. A real one for "is this chest-pain patient high-risk?" — note it uses no probabilities at all, just risk:
The cues near the top matter most — they filter everything below. Doctors and soldiers really do decide this way under pressure, and it works well in that environment.
Ecological rationality is the criterion — "is this rule good in the environments we actually face?" Fast-and-frugal heuristics are the mechanisms we evaluate against it.
How do experts decide well under time pressure without optimising? They recognise a familiar situation from limited cues, retrieve a plausible action, mentally simulate its consequences — and if it's good enough, do it; otherwise try the next.
Klein's Recognition-Primed Decision model is literally recognition heuristic + satisficing + mental simulation. It's a special case of ecological rationality: the "environment" is whatever a given expert faces under pressure.
Prospect theory
Every model so far still assumed a stable utility — a fixed way of valuing outcomes. The paradoxes say even that is wrong. Kahneman & Tversky's answer rebuilds valuation itself.
(1) A value function, not utility. We value outcomes as gains and losses relative to a reference point, not in absolute terms. (2) Weighted probabilities, not raw ones. We over-weight small probabilities and under-weight large ones.
The value function — play with it 📉📈
The curve is S-shaped: concave for gains (risk-averse), convex for losses (risk-seeking), and steeper on the loss side — that asymmetry is loss aversion. Drag the sliders:
The 0 in the middle is your reference point. Right = gains, left = losses. Watch how a loss "feels" bigger than an equal gain.
Prospect theory also bends probabilities through a weighting function w(p): tiny probabilities feel bigger than they are (why we buy lottery tickets and insurance), and near-certain ones feel smaller. The value of a "prospect" is Σ w(p)·v(outcome − reference) — same shape as expected utility, but neither piece is what classical theory assumed. It fits real human choices far better.
The Exam Lab
Behavioral theory is "softer" than the math lectures — so the exam rewards clear intuition and good examples. Every answer below is written that way.
Open questions, discuss the topic (same shape as the other lectures). ① name the idea · ② explain it in plain words · ③ give a concrete example (a paradox, a heuristic, a bias). You're not asked to reproduce formulas — you're asked to show you understand why humans deviate from rationality and what that means for designing a DSS.
Why does a decision-support system need a model of how humans decide, if it can already compute the rational decision? Discuss.
Describe one experiment showing humans violate rationality (e.g. the Ellsberg paradox), and say which assumption it breaks. Discuss.
Explain bounded rationality and how procedural rationality / heuristics go further. Discuss.
What is a heuristic, and how does it produce a cognitive bias? Give examples. Discuss.
If heuristics are biased, why use them? Explain ecological rationality and fast-and-frugal heuristics. Discuss.
Explain prospect theory and how it differs from expected utility — the value function and probability weighting. Discuss.
• Why must a DSS model human (not just rational) decision-making?
• Ellsberg / preference reversal / Allais — what does each one break?
• Bounded rationality in one sentence — and one reason it's not enough.
• What's a heuristic? Name two and what each depends on (order of options vs order of cues).
• Heuristic → bias: when does an error become a bias? Give two biases.
• Ecological rationality: why can a "biased" heuristic be good? What are fast-and-frugal heuristics?
• Prospect theory: the value function (reference point + loss aversion) and probability weighting.