Strategy Transfer Learning via Parametric Deviation Sets

Accurately reasoning about agents’ actions in strategic settings is a challenging artificial intelligence task. Many difficulties in learning to perform such reasoning arise due to the uncertainty in both agents’ motives and the strategic games being played. In this paper, we address the problem of learning from observations of the agents’ behavior in some games to predict play in different, but related, games. We introduce a deviation-based strategy prediction approach that, by also using game outcome features to describe different games in a common language, enables generalized strategy learning.