A taxonomy of coordination problems

by damiensnyder9 min read22nd Aug 20211 comment

27

Coordination / CooperationRationality
Frontpage

The LessWrong coordination/cooperation tag describes coordination as:

the challenge of distinct actors being able to jointly choose their actions to achieve a favorable outcome. [...] A closely related concept is that of cooperation – multiple actors choosing their actions in ways that maximize collective value despite the temptation of greater short-term individual gain by acting to the detriment of the group/other actors.

For the purposes of this post, I will define a "coordination problem" as a problem such that a coordinated group will likely create a good solution, but without coordination, an individual is unlikely to take steps toward that solution. A classic example of a coordination problem is a stag hunt, because the group can create a good solution (by all choosing stag), but until that problem is solved, an individual is more likely to choose rabbit. An example of a non-coordination problem is convincing people to get vaccinated. This is not a coordination problem, because most people are incentivized to get vaccinated regardless of coordination. (Those who reject vaccinates are still usually incentivized to get vaccinated, though they may believe otherwise.)

Coordination problem is a useful category, but it describes several distinct types of obstacles, which require different approaches to solve. I have identified five types of obstacles, but this is not the only way to categorize them. I am open to suggestions for adding or amending categories.

Taxonomy

1. Participants cannot communicate easily

Some coordination problems could be easily solved if participants could more easily talk to one another. For example: If I tried to meet my friends at a restaurant, but we couldn't message or visit each other beforehand, it would be pretty hard. We wouldn't know where everyone else would go or when, so we would have to hope to get lucky.

An obvious solution to problems like this is to facilitate communication better. If my friends and I could text, we could figure out restaurant plans easily. If this is impossible, another solution is to have a Schelling point solution that everyone can fall back on. This can be coordinated beforehand, or in a pinch you can guess by trying to model everyone else's thinking. For example, maybe there's a restaurant that my friends and I have been to a lot, so if I go there at 7 I'll probably run into some of them. This might not be the best restaurant we could go to, but at least it's better than everyone failing to find each other at all.

If one-way communication is available, then a person who can communicate one-way could dictate details to everyone else. They might not choose the best option available, but relying on what communication does exist is usually better than everyone just guessing.

2. Tragedy of the commons / prisoner's dilemma

This category includes problems where individuals are incentivized to act to the detriment of the group. For example, if I am willing to steal from people, I can have more stuff. But this is to the detriment of the group as a whole: stealing from people creates more problems for everyone else than benefits for me. If everyone acts on their incentives and starts stealing, the society would be worse than if no one stole. (In a prisoner's dilemma, stealing would be "defecting".)

The main solution is to punish defection so that it is no longer the best option. The citizens could create a police force that puts thieves in jail; or, they could enforce a system of reputation, where people who steal are looked down upon and given worse treatment. Reputation is generally the easier way to punish defection, but it doesn't work at the scale of cities or nations, because no one can remember the reputations of a million people. Thus governments are the main solution to coordination problems at large scales.

If some person or organization can directly control the actions of other participants, then they can unilaterally force everyone to cooperate. But cooperation might not be worth the risk of handing over control of your actions to another party.

3. Stag hunts

These are similar to prisoner's dilemmas, with one difference: if there is sufficient buy-in from other participants, an individual is not incentivized to defect. In the classic stag hunt problem, if everyone else chooses stag, an individual player is also best off choosing stag. So past a certain threshold of coordination, the optimal solution enforces itself. The main difficulty is getting the initial buy-in from enough people to reach that threshold.

Assurance contracts are not very widely used, but are designed to solve exactly this problem. In an informal setting, like in the stag hunt problem, players can plan a specific time to all switch to hunting stag, so they don't all have to guess whether it will be too risky to try switching individually. Sometimes players worry that others will not follow through on the commitment, and choose not to follow through either. To prevent this, other players should reassure everyone that the switch is definitely going to work as planned, even to the point of lying about how confident they are. Making a strong precommitment, like throwing away your rabbit-hunting gear, can project even more confidence.

Another way to get everyone on the same page is to delegate individual decisions to a single leader, who can then force everyone to switch at once. This only works, of course, if the leader has sufficient power or buy-in.

4. Desire to coordinate is punished

This is similar to a stag hunt, but has the extra challenge that trying to coordinate is itself punished. For example: if most people believe bullying is mean, but you look uncool for saying that, the people who believe bullying is mean will not say so, so bullying will continue to happen.

One solution is to anonymously poll people, to expose the existence of a silent majority who hates bullying. Alternatively, someone fearless could advocate against bullying despite the consequences. If everyone found out that others thought the same way, they would be less afraid to share their anti-bullying stance, so people would start to intervene against bullying.

This problem has some informative real life examples. For example, public opinion on LGBT rights flipped in a relatively short span. Before the flip, publicly being LGBT or supporting gay rights would damage your reputation. But some prominent advocates raised visibility and garnered support for the movement, encouraging more LGBT people to "come out". At some critical mass of publicity, public support for LGBT rights grew rapidly (around 2005–2020), because there was no fear for reputation anymore. Expanding LGBT rights is not solely a problem of demonstrating that support exists, because support had to grow overall. But it does demonstrate that raising visibility can make it safer to show support for an issue.

Another coordination problem in that vein is of popular revolts against oppressive leaders. Until the regime is overthrown, revolutionaries will be silenced or jailed if they are found out, and revolutionary communication will be suppressed. In many oppressive states, revolutionaries form secret organizations with secret communication networks to circumvent that. (In the previous example, underground gay bars partly served this role.) If possible, revolutionary leaders will seek asylum in friendly countries and lead efforts from abroad. Fearless but less noteworthy members of the movement can become prominent through martyrdom (being prominently jailed or killed) or by sheer numbers. Innuendos can spread revolutionary messages without being punished, such as in the buildup to the Tiananmen Square protests. In some cases a particular region of the country can become a bastion of the movement, where opposition can operate safely. If the resistance demonstrates enough power and support, then eventually ordinary citizens will become less afraid and join in.

Countless sexual predators have remained in positions of power because of this coordination problem. Accusing a powerful person of sexual assault is bad for the victim (whose career advancement may be threatened), and often insufficient on its own to remove the predator. But if multiple accusers step up, they can force the issue. The MeToo movement could be seen as a watershed moment for many coordination problems like this, where several victims at once coordinate to expose the predator's crimes.

5. Battle of the sexes / negotiations

A battle of the sexes is a game where two players would like to choose the same option as the other, but each player wants to coordinate on their own preferred option. This is a simplistic way of expressing a negotiation: multiple parties want to make a transaction happen, but each party wants terms more favorable to themselves. For example: some people want to host a party at Alex's house, but some people want the party to be at Beth's house instead. Both groups would like to party with everyone, even if they don't get it at their preferred location.

The primary goal is to make sure everyone coordinates on one option, but as a secondary goal, everyone should coordinate on the best option. The simplest strategy is for the group to discuss it until everyone accepts the option with the most support. This might involve compromise—say, having this party at Beth's but the next one at Alex's—or persuasion, if one side can be persuaded. If the group is too large to effectively discuss, or wants to decide quickly, they could put it to a vote. If a vote is unsatisfactory, perhaps because some people care more than others, the group could consult a leader or an impartial judge, and resolve to do whatever they decide.

Political primaries serve as a solution to this sort of coordination problem. Without coordination, voters with a common interest would fracture their vote in a general election, because they all individually prefer different candidates. A primary coordinates everyone in the party to vote for a single candidate, and as a secondary goal tries to ensure the best candidate is nominated. Likewise, a general election ensures that multiple politicians do not have to share a single office, and given that constraint attempts to choose the best politician.

(Some other interesting writing on this problem: Nadia Eghbal, Vitalik Buterin)

Patterns

One solution that applies to most coordination problems is to have a leader direct everyone. Then no one has to coordinate; they only have to do what the leader says. The leader could ask for input, but ultimately they make the decisions. Sidestepping coordination saves a lot of difficulty, which is why just about every organization has a hierarchy. (Presidents, generals, CEOs, managers, etc.) Hierarchy is difficult to avoid in any group of more than a handful of people.

Some of these problems require different solutions based on scale of the group coordinating. Reputation can enforce norms in a small community, but large communities need legal systems. Consensus-building works for small groups, but voting is necessary for large ones. Assurance contracts (mostly theoretically) solve stag hunts for markets too large to verbally coordinate.

As I came up with examples, I realized that many of the issues I thought were coordination problems were really other types of problems. "People being wrong" problems, "the result isn't worth the effort" problems, "no one cares" problems, and "the group would be better off splitting" problems are related categories that can seem like the same thing, but aren't.

Having written this taxonomy, I will attempt to be more specific when I mention a "coordination problem". Specifying a "stag hunt problem" or a "tragedy of the commons problem" makes it more clear what types of solutions might apply, and what other problems might relate to this one.

Do let me know if you think I missed a category, or used poor definitions. "The categories were made for man" and all that.

27

1 comments, sorted by Highlighting new comments since Today at 10:52 PM
New Comment

I appreciate this post's broad strokes. I think it's worth distinguishing between different types of coordination failure, because their causes and solutions are quite different.

(This sort of makes me want to make an alternate version of the post called "Tabooing 'Coordination Failure'" whose thesis is less 'these are the particular types of coordination failure' and more 'be more clear about what you're talking about when complaining about it.')

One solution that applies to most coordination problems is to have a leader direct everyone. Then no one has to coordinate; they only have to do what the leader says. The leader could ask for input, but ultimately they make the decisions. Sidestepping coordination saves a lot of difficulty, which is why just about every organization has a hierarchy. (Presidents, generals, CEOs, managers, etc.) Hierarchy is difficult to avoid in any group of more than a handful of people.

I'd phrase this differently – a leader is an instance of a coordination solution, not an absence of coordination.