El Farol problem

The El Farol problem is a model of coordination under uncertainty: when a popular bar becomes unpleasant if too many people attend, each person must decide whether to go by anticipating the choices of everyone else. The problem’s importance lies in its counterintuitive result: even when no individual strategy can reliably predict the crowd, the group can collectively approach the outcome it wants without a central authority directing anyone’s choice.

The coordination problem

The model begins with a familiar predicament. A local bar is enjoyable when it is not crowded and miserable when it is. If everyone expects a busy night and stays home, the bar will be empty and those who go will have a good time. If everyone expects it to be quiet and turns up, it will be packed. The decision is therefore interdependent: what makes sense for one person depends on what everyone else is expected to do, and those other people face the same calculation.

A central planner could, in principle, assign who may enter each week and keep attendance at the desired level. But that would interfere with people’s freedom to choose their own course of action. Any workable solution must instead arise from the mixture of individual decisions. The challenge is not simply to forecast attendance; it is to coordinate choices when the forecast itself changes what people choose.

Arthur’s El Farol model

In the early 1990s, economist Brian Arthur named the problem after El Farol, a bar in Santa Fe that could become overcrowded on nights featuring Irish music. In his model, the bar had a capacity threshold of 60 percent: below that level, everyone who attended would have fun; above it, no one would. People went only when they expected attendance to remain below the threshold.

Because no single answer could be deduced mathematically, Arthur gave the simulated bar-goers different forecasting strategies. Some expected attendance to match the previous week; others looked to the last night they had personally visited, averaged the attendance of recent weeks, or predicted that this week would reverse last week’s outcome. Even someone who had not gone could learn how crowded the bar had been.

Arthur then ran computer experiments simulating one hundred weeks of attendance. The agents’ strategies produced sharp, irregular fluctuations rather than a stable pattern. The bar sometimes exceeded the 60 percent threshold for three or four weeks in a row; at other times, it was below capacity for four of five weeks. No strategy worked indefinitely: approaches that succeeded for a time eventually had to be abandoned.

Those fluctuations had real costs within the model. On some Fridays the bar was too crowded for anyone to enjoy it; on others, people stayed home who would have had a good time if they had gone. Yet across the hundred weeks, average attendance was exactly 60 percent—the level at which the largest possible number could enjoy the bar without anyone having a bad time. Individual misfires did not prevent the crowd, over time, from reaching its preferred average.

A different route through experience

A few years later, engineers Ann M. Bell and William A. Sethares approached the problem with a less elaborate strategy. Rather than assigning bar-goers a variety of forecasts about attendance, they gave them all the same simple rule: return when recent visits had been enjoyable; stay away when they had been unpleasant.

These agents knew less than Arthur’s. They did not try to reason through what other people were thinking, and they did not know how crowded El Farol had been on nights they themselves had missed. Their guide was their own experience of the bar. After enough time had passed for each to learn whether a visit had been good or bad, attendance settled just below 60 percent of capacity. That result fell a little short of the ideal a perfectly informed central planner could achieve, but it came from people following their own experiences rather than coordinating through a shared forecast.

The contrast between the experiments is revealing. Arthur’s agents repeatedly revised among diverse strategies as conditions changed; Bell and Sethares’s agents relied on a common, locally grounded feedback rule. Neither method eliminated uncertainty or made every Friday successful. Both nevertheless produced group outcomes close to the desired level, by different paths.

Local knowledge and collective intelligence

The experiments illustrate a broader tension in collective decision-making. Decentralization encourages independence and specialization, and allows people close to a problem to draw on knowledge rooted in their particular experience. But it cannot guarantee that useful information discovered in one part of a system will reach the rest. A crowd, company, or intelligence agency therefore faces two demands: preserve the local character of individual knowledge while finding ways to make it collectively useful.

The El Farol examples show that independence need not mean disorganization. People whose decisions depend on one another can still produce a collectively intelligent result, even without an all-powerful planner and even when no individual can count on being right each time. In the simple feedback version, participants do not need a complete picture of the crowd; their own recent experience is enough to help move attendance toward the shared target. The outcome is not perfect, but it approaches the balance the group collectively wants.

The model’s broader lesson is not that every decentralized group will coordinate successfully. It is that collective intelligence can arise from the interaction of varied or locally informed decisions, while centralized control carries costs of its own. The practical problem is to let individuals retain access to what they know while creating ways for separate judgments and signals to shape the group’s result.