Strategic interactions in the real world such as decision/voting often exhibit local structure in terms of how players are guided by or respond to each other. In other words, different agents make rational moves in response to their neighboring agents leading to locally stable configurations such as Nash equilibria. We design new mathematical tools at the intersection of game theory and machine learning for qualitative and quantitative learning and modeling of such games. We investigate conditions under which stable configurations exist, quantify the hardness of recovering the neighborhoods of agents, design estimation algorithms for meaningful structure recovery and quantitative prediction of equilibria.
Learning Tree Structured Potential Games