Michael Kearns on Game Theory and Machine Learning

Lex Fridman asks Michael Kearns how game theory helps explain outcomes that arise from interactions among people.

Kearns calls game theory a mathematical framework for reasoning about collective outcomes among individuals. He cites the prisoner’s dilemma, in which individual incentives can lead to an outcome worse than cooperation.

Algorithmic game theory studies large populations and complex incentives, including how platforms might predict or influence outcomes.

Kearns highlights Nash’s work on equilibrium, while clarifying that an equilibrium’s existence does not guarantee that real-world behavior will converge to it.

He describes connections between game theory and no-regret learning. Under suitable models, interacting learners can approach equilibrium; stability alone does not make an outcome good or bad.

Navigation apps illustrate the interaction: each driver seeks a fast route based partly on what other drivers are doing. The system can nudge users toward individually attractive choices.

Kearns says theory and limited traffic studies suggest that an equilibrium of individual route choices can have greater total travel time than another arrangement. A different collective outcome would require different mechanisms.

He extends the question to social platforms and shopping: machine learning predicts traffic, products or content for individual users, but collective welfare is a separate question.