Coordination problem

A coordination problem is a situation in which people must align their actions even though each person’s best choice depends on what others will do. It is harder than estimating a fact that exists independently of the group: the participants’ decisions help create the very outcome they are trying to predict. The book treats coordination as a recurring challenge in everyday conventions, markets, traffic, and organizations, and argues that independent local judgment can be valuable only when people also have workable ways to align or aggregate it.

What Makes Coordination Difficult

When Francis Galton’s fairgoers estimated an ox’s weight, they were judging a reality outside the group. By contrast, in the El Farol problem, people decide whether to visit a bar whose desirability depends on how many others make the same choice. Their decisions alter the outcome they are trying to anticipate, creating a self-referential loop without an external reference point to settle it. Under those conditions, even arriving at a roughly correct average attendance is striking.

The book compares two approaches to El Farol. In one, computer agents make forecasts of attendance; in another, associated with Ann M. Bell and William A. Sethares, people rely on their own experience rather than trying to predict everyone else. After people had time to learn from their visits, attendance in the latter experiment settled just below 60 percent of the bar’s capacity. That was slightly less effective than an ideal central planner, but it was a collectively intelligent result produced without each participant needing to calculate the group’s behavior.

The stable attendance had a cost: people sorted into frequent regulars and rare visitors. Assuming everyone took equal pleasure in a visit, this arrangement was utilitarianly reasonable: more than half the group could go nearly every week, usually without encountering a crowded bar. But a sizable portion rarely went and often had a bad experience when they did. The comparison therefore does not yield a simple winner. Both solutions were surprisingly good, yet the better one depends on how the group weighs reliable access for regulars against a more even distribution of opportunity.

Shared Focal Points and Conventions

Coordination does not always require explicit communication or a central authority. In a 1958 experiment, Thomas C. Schelling asked law students from New Haven where they would meet an unknown partner in New York City. A majority independently selected the information booth at Grand Central Station. When asked what time they would choose if the date were known but the hour were not, almost all said noon. The familiar place and conventional time acted as focal points: shared expectations gave people a plausible way to match choices despite having no means to consult one another.

Schelling found the same tendency in simpler matching tasks. When paired participants had to choose “heads” or “tails” and succeed by matching, 36 of 42 chose heads. Asked to select one square from a box of 16, 60 percent selected the upper-left square. Asked to name a positive number, 40 percent said one. These choices show how people can converge even when the apparent range of options is large, because some options stand out as natural meeting points.

Conventions also coordinate behavior in ordinary public life. A single queue brings people waiting for a service into one line, with the person at the front served first; the alternative, separate lines for separate checkout counters, makes comparisons and choices more burdensome and frustrating. A queue may look more daunting because everyone is gathered together, but on average a single line serves people faster. Its fairness also rests on an egalitarian convention: permitting people to sell their places would let some exchange waiting time for money and others money for time, but would violate the norm that governs the queue. More broadly, voluntarily observed conventions allow people to coordinate without coercion and without spending excessive thought or effort on each interaction.

Markets, Local Knowledge, and Information

The free market is presented as a way to solve a vast coordination problem: moving resources to the places where they are wanted, at a cost that reflects what it takes to provide them. In the ordinary supply chain, shoppers buy orange juice that grocers stocked without knowing exactly who would arrive; the juice was packaged days earlier from oranges picked weeks earlier, by participants who may never know the final customers. Shoppers, grocers, wholesalers, packagers, and growers act on partial, local knowledge and their own interests. If coordination succeeds, goods move from producers able to supply them cheaply to people who want them strongly, without any participant seeing the whole system or knowing beforehand what the right result will be.

Decentralization helps make this possible because people close to a problem may possess knowledge grounded in a particular place, job, or experience: knowledge that cannot readily be summarized and passed along. Local decisions can encourage specialization and independence while increasing the information available across a group. But decentralization alone is not enough. Useful knowledge discovered in one part of a system may never reach another. Groups need ways to preserve local expertise while making what people learn collectively useful; the example of Google’s use of knowledge supplied by millions of webpage operators illustrates how dispersed contributions can be aggregated without erasing their origins.

The same distinction matters in institutions. Before September 11, the U.S. intelligence community had divisions of labor among agencies, a structure that could support specialized analysis and diverse perspectives. Its central failure, as described in the book, was not decentralization itself but the absence of a means to aggregate information and judgments. Richard Shelby’s proposal for a central intelligence agency addressed the fragmentation problem by putting information under a central authority, but risked sacrificing the independence, diversity, and local knowledge that decentralization provides. The proposed alternative was aggregation rather than centralization: linking agency databases, for example, could help information travel while agencies retained autonomy. Even two years after September 11, the government lacked one unified watch list drawing on all parts of the intelligence community.

When Decentralized Systems Fail

Road traffic shows how individually sensible choices can produce a collective breakdown. In 2002, about 250,000 vehicles entered central London on a typical day, competing for space on narrow, winding streets alongside a million public-transport commuters. Average traffic speed stayed below 10 miles per hour, and on a bad day fell to about 3. In February 2003 the city introduced a £5 charge for entering central London between 7 a.m. and 6:30 p.m.; a driver recorded by one of 230 cameras without paying faced an £80 fine. The plan was intended to reduce congestion by 20 percent and raise £180 million annually for public transportation. Its underlying logic was that a driver imposes delay costs on others without paying for them; the toll was an attempt to make that cost count in the driver’s decision.

Pricing, however, addresses only some causes of traffic jams. A highway lane may carry about 1,800 cars an hour under the usual two-second following distance, or as many as 2,400 on a fast-moving California freeway. Reducing traffic below capacity can prevent jams caused by sheer volume, but tolls cannot remove the effects of accidents, construction, entrances and exits, road curves, inclines, or a slow truck. Traffic also has a dynamic instability: as vehicles slow and brake to preserve safe distances, the reaction spreads backward through the packed cars. Lane changes fill available gaps and can bring lane speeds together at a lower level. A small disturbance may then trigger stop-and-go traffic, which is difficult to dissolve; the jam can travel backward even as vehicles move forward out of it.

This is a coordination failure among decentralized individuals following basic rules—avoid the car ahead, change lanes when possible, and drive as fast as safety permits. Each driver wants to get ahead, while also benefiting if traffic as a whole moves quickly. Unlike successful matching at a focal point, these aims do not automatically produce an aligned outcome. The contrast clarifies the book’s central claim: dispersed decision-making can solve collective problems, but it does not guarantee that individual actions will combine into a coherent result.

Coordination Through Organization and Collaboration

The same tension appears inside corporations. A rigid hierarchy can obstruct the flow of knowledge from people doing a job to those making decisions. In one American company, a new product idea had to traverse 223 possible paths through the bureaucracy before acceptance. At a General Motors plant in Van Nuys, a worker had installed a bracket for two years without being told that it held the car’s sunshade. The details illustrate the cost of separating decisions from the people who possess practical knowledge of the work. The book’s alternative is to let decisions about local problems be made, as far as possible, by people close to them, rather than sending every issue up a hierarchy and every solution back down.

The global search for the cause of SARS offers a contrasting example of decentralized work with effective aggregation. In March 2003, the World Health Organization brought 11 laboratories from countries including France, Germany, Japan, the United States, Hong Kong, Singapore, and China into a collaborative research project. The laboratories held daily teleconferences, posted microscope images and analyses online, and exchanged virus samples. They could investigate the same material in parallel, compare results, and eliminate candidate causes that appeared in some patients but not others. By March 21, researchers in Hong Kong and the United States had separately isolated promising viruses; over the following weeks, laboratories detected a coronavirus across a wide range of SARS samples and sequenced it. Experiments involving infected monkeys in the Netherlands provided further evidence. On April 16, about a month after the collaboration began, the network announced that the coronavirus caused SARS.

Cynthia Goldsmith, an electron microscopist at the Centers for Disease Control and Prevention in Atlanta, was the first to spot the coronavirus, but the book distinguishes that observation from establishing the cause of the disease. The conclusion required weeks of work by laboratories around the world, including research that ruled out other possibilities. No individual lab was in charge of the whole effort; the WHO organized the network, while the laboratories shared relevant data and worked out how to divide the investigation. The discovery was therefore collective: parallel expertise became decisive because information and judgments could be compared and combined.

The Common Ground of Collective Judgment

Science extends this logic beyond explicit research partnerships. Its collective character depends on publishing findings and making data available for others to inspect, challenge, and use in new hypotheses and experiments. Failed hypotheses can help by showing other researchers where not to look. Scientific knowledge is not only cumulative, built on predecessors, but collective, because contemporaries depend on one another’s work as well. The book’s image of standing on the shoulders of giants thus points to a continuing exchange, not merely a succession of isolated discoveries.

A scientific result does not become accepted simply because one person has produced it. The coronavirus caused SARS before the scientific community announced the finding, but its standing as an established scientific conclusion depended on other researchers scrutinizing and accepting the evidence. Laboratories could then build possible tests and vaccines on that shared conclusion. This process is not a literal vote, and most scientists will not repeat every experiment. They rely on other researchers’ data and on a wider body of untested practices and knowledge. Replication and the exposure of fraud remain possible, but routine scientific work also requires trust that competitors will handle and report evidence fairly.

Coordination, then, is not a single recipe for agreement or centralized control. Shared conventions can provide focal points; markets can connect people acting on partial knowledge; organizations can preserve local initiative while transmitting useful information; and scientific communities can turn independent investigations into accepted common knowledge. Each arrangement faces the same underlying difficulty: people must retain enough independence to contribute what they know, while finding a way for their separate choices and judgments to add up to something they could not achieve alone.