one life to live · metaphor 37 of 100
Tonight: the restaurant you love, or the one you've never tried? Multiply by every dinner, friendship, job, and city you have left, and it becomes the question of how to spend a finite life among options you can only rate by living them.
The twenty-five-year-old who moves to a new city every year and the sixty-year-old who orders the same dish at the same corner table get filed under different temperaments: the adventurer, the creature of habit. Watch more carefully and they may be playing the same strategy, correctly. A discovery is an investment, and like any investment it is priced by the time remaining to collect on it. A city that fits, found at twenty-five, pays out over ten thousand evenings; the same discovery at eighty pays out over a few hundred. Neither of them is braver. One of them simply has more future to sell to.
Mathematics keeps this problem in a drawer labeled the multi-armed bandit: a row of slot machines, each with a hidden payout rate you can only learn by feeding it coins you never get back. Below, the machines are four restaurants and the coins are evenings. Every visit returns a noisy rating — a great kitchen has off nights; a mediocre one gets lucky. Play the town yourself first. Then hand it to the strategies, and watch the arithmetic take sides.
hidden means drawn uniformly from 5.0–8.0 · each visit returns μ + gaussian noise (σ = 2.0) · every number on this page is simulated live — nothing is scripted
…
best ε by horizon: computing…
the bandit's dilemma
Choose any restaurant above and you collect two things at once. The first is dinner: tonight's enjoyment, banked and finished. The second is a data point: your estimate of that place sharpens by one noisy observation. Exploitation maximizes the first currency — go where the running average is highest, and tonight is as good as you know how to make it. Exploration maximizes the second — go where your knowledge is thinnest, and tonight will probably be worse; that is the fee — but every evening after tonight inherits a better map.
The dilemma is structural, not psychological: one action, two returns, and the returns mature at different times. Exploitation pays now; exploration pays only across whatever future remains, which is why noise makes it treacherous. A single visit lies freely. In this town, judging all four restaurants by one night each crowns the wrong favorite roughly two times in five — and a wrong favorite, revisited faithfully, generates no further evidence against itself.
what to try
Play the town twice, and notice the evening you stopped experimenting — the night curiosity quietly handed the keys to habit. That switch has a name in the strategies tab: ε, the fraction of evenings you still give to chance. Now let the machines play. Watch always the favorite (ε = 0): it tastes each place once, crowns a winner off a single noisy night, and never looks back. In a large share of towns it settles on a restaurant that is genuinely good — and not the best — forever. That is the cost of premature certainty, and note its most dangerous property: from inside, it feels like nothing. No alarm rings. It feels like a pleasant routine.
Then drag the horizon. At 10 evenings remaining, the Monte-Carlo readout will tell you the kindest ε is at or near zero — with no future to amortize it, information is a luxury. At 500, an ε of several percent beats loyalty handily, and ε = 0 becomes one of the worst lives on the menu. Same town, same taste buds; only the amount of future changed.
horizon is the whole game
The young are told to explore as if it were a virtue of the young; the bandit says it is simply correct accounting when the horizon is long, because every discovery is multiplied by the evenings left to enjoy it. And the same arithmetic, run in reverse, acquits the old. The eighty-year-old returning to the same café is not ossifying; with a short horizon, an unknown option's upside has almost no time to compound, while its downside is paid in full tonight. What we sneer at as set in their ways is often information, correctly priced.
The readout below the strategies makes this exact: as the horizon stretches from 10 evenings to 500, the best exploration rate climbs from zero to a healthy wander. One caution runs the other way — the horizon that matters is the one attached to the decision, not to you. A move, a career change, a divorce resets the arms and hands you a long horizon in the middle of a short life. People err here in both directions: exploring at eighty like a tourist with no flight home, exploiting at thirty as if the flight were tonight.
optimism as strategy
ε-greedy explores at random — charmingly, stupidly at random, re-tasting a place it already knows is dreadful. There is a better instinct, and it has a formula. UCB — the optimist — scores every restaurant by its average so far plus a bonus for how rarely it's been tried, then goes wherever that optimistic total is highest. It literally treats under-sampled options as better than they have looked. This is called optimism in the face of uncertainty, and it is curiosity implemented as an algorithm: doubt, converted into a budget line.
Optimism is self-correcting: if the bonus lures you somewhere bad, the visit itself shrinks the bonus, and the error retires. Pessimism is self-sealing: a bad first impression is never revisited, so it is never corrected — the one restaurant, or person, that got an unlucky audition stays condemned on one night's evidence. In the readout, the optimist pays a small tax at short horizons and wins by a widening margin at long ones. Grudges are cheap in December; in April they are ruinous.
the mapping
| Mathematics | Life |
|---|---|
| an arm | A restaurant, a career, a city, a person — any option whose quality can only be learned by living it. |
| pulling an arm | An evening spent, unrepeatable: the coin you feed the machine is time, and there are no refunds. |
| exploration | Sampling the unknown at the cost of the known-good — paying tonight for a sharper map tomorrow. |
| exploitation | Harvesting the best you've found: the beloved table, the trusted friend, the craft you're already good at. |
| horizon H | The time left to use what you learn — the multiplier on every discovery, and the real difference between twenty-five and eighty. |
| regret | The gap between the life you lived and the best arm — known, if ever, too late to matter. |
where the metaphor tears
The instrument's qualities are frozen; life's are not. The beloved restaurant loses its chef; the dull job becomes interesting under a new boss; the difficult friend mellows. Mathematicians call these restless bandits, and their lesson amends the whole page: a settled conclusion has a shelf life, so some re-exploration of what you've already decided — the written-off cuisine, the estranged brother — is always due, at any age.
The model assumes each μ is a property of the restaurant, waiting to be discovered. But sampling reshapes the sampler: the palate learns to love what it practices, the tenth opera is better than the first because you changed. Payouts in a life are partly written by the visits themselves — which means exploration doesn't just measure value, it creates it, and the neat separation of learning from earning dissolves.
The bandit learns only from its own pulls; you have friends. Reviews, recommendations, a sister's report from the city you've never tried — social information partially decouples exploring from spending your own evenings. It is cheaper than first-hand sampling, and noisier in a particular way: their μ is not your μ. The crowd can tell you where to look; it cannot tell you what it will feel like to live there.