14 Multiagent Systems

14.8 Social Impact

The multiagent version of the prisoner’s dilemma is the tragedy of the commons, named because common resources (such as the air we breathe) do not become a priority for anyone acting to maximize their utility, making everyone worse off, as in the following numerical example.

Example 14.22.

Suppose there are 100 people and there is an action that each person can choose (such as driving an internal combustion engine car) that gives them 10 units of utility but costs the environment 100, which, shared among everyone, corresponds to −1 for everyone; otherwise each agent can do an action with utility 1 for them and 0 for the environment (such as walking). When each person considers driving, they have a reward of 10 minus their share of the environment, which is −1, so driving has a utility for them of 9, so they are better off driving. This is independent of what anyone else does. So then everyone drives, with a driving reward of 1000 minus 10,000, which is the utility for the environment, giving a total utility of −9000, which is −90 for each person. Everyone would be much better off if no one drove, even though individually each person has a great utility by driving.

One way to solve that is to implement a VCG mechanism, where each person pays a tax. In this case, the mechanism would specify that each person would have to pay a tax of 99, which corresponds to the cost of their action on the others. Then each person would be rational to not drive.

One problem with the scenario in the last example is to convince people to implement the tax. After all, the only reason they are not driving is because of the tax, and they want to drive. One major problem is to compute the cost of changing the environment, where the cost may occur a long time into the future, so depends on how future values are discounted, and where the impact may be in other locations, and so depends on how others’ suffering and loss of habitable land should be measured. This is a key problem for efforts to restrict greenhouse gas emissions to mitigate climate change. It occurs at the individual level as well as at the country level, with countries not reducing their carbon emissions because it is better for them not to.

AI can potentially help in a number of ways. It can make better predictions of the future costs, however, it cannot determine how these future values are discounted, as this depends on society’s values (and each person might value the future differently). It might be able to help by promoting posts on social media that are known to be true, not promoting posts known to be false, and only promoting posts that have appropriate caveats, rather that maximizing engagement, which is known to lead to promoting extreme views [Acemoglu et al., 2021]. However, actually determining whether a post is true or false is very difficult.

Perrault et al. [2020] describe a research agenda about multiagent systems to address complex societal problems. They present many successful deployed systems in three application areas: public safety and security, wildlife conservation, and public health in low-resource communities. Much of their work is framed in the context of Stackelberg security games. A two-player Stackelberg security game is played between a defender and an attacker. Using limited intervention resources, the defender moves first to protect a number of targets from the attacker. Perrault et al. [2020] show how AI can play an important role in fighting social injustice and improving society.