Robots use game theory to understand how we think – Marketplace

This is part of our Econ Extra Credit project, where we read anintroductory economics textbookprovided by the nonprofitCore Econtogether with our listeners.

Robots are doing their part during the COVID-19 pandemic. They whiz around our hospitals and grocery stores, taking patient temperatures, sanitizing high-touch surfaces and restocking shelves and they help minimize the amount of human contact required to accomplish it all.

Part of a robots job is navigating around the people it encounters.

The thing about these environments our roads, our hospital floors is that theyre not empty. Theyre full of people. And so these robots need to choose their actions carefully, said Anca Dragan, a computer science professor at the University of California, Berkeley. Because you dont want these robots bumping into people.

Which means that robots need to get smart enough about our actions to anticipate what were going to do next. They need to understand us, to a certain extent a tall order if we dont always understand ourselves.

To model human behavior for these robots, engineers might start with the basic goals that motivate us as we move about public spaces: We walk down the produce aisle looking for asparagus, or were trying to get through an intersection before the light turns red.

This is very much borrowed from economics, where we model human behavior there as responding to these incentives or utilities, Dragan told Marketplace Morning Report host David Brancaccio.

But our environment can influence our behavior. If a car turns into our lane up ahead, we might slow down, even if it means missing that light.

The easiest thing to do is what traditional economics does, which is to assume that people are these perfect little game players; theyre perfect optimizers; theyre perfect utility maximizers and were not.

Enter game theory, a bedrock principle of behavioral economics.

Game theory comes from the fact that neither the human nor the robot make decisions about what to do in pursuit of their goals in a vacuum. Theyre aware of each other, Dragan said. There comes this tension: What people do influences what the robot does; what the robot does influences what people do. Its this back and forth. Its kind of a negotiation.

The tools of game theory allow robots to account for the ways people might change course because of the presence or actions of others. And anticipating human action efficiently enables robots to make progress toward their own goals, too.

But game theory relies on rational actors and sometimes, human behavior can seem irrational to the onlooker.

One thing that makes all of this very challenging is that once you think about it in this game theoretic terms, the easiest thing to do is what traditional economics does, which is to assume that people are these perfect little game players; theyre perfect optimizers; theyre perfect utility maximizers and were not. And behavioral economics has told us that for a long time, Dragan said.

Take, for example, somebody who sells their stocks when the stock market crashes. Instead of assuming that they are fundamentally irrational, or they dont care about losing money, we might instead consider the beliefs that could have motivated the action. We might need to tweak our model. Selling your stocks when the market is down appears far more rational if you think the market will fall even further in the future.

Dragan says it is important to teach robots how to follow this sort of reasoning.

They should look at an action and not impart their own understanding of the world on the person, [or] expect the person to be rational under the robots view of the world, she said.

Itll help robots understand us better, so they can help us better.