A Model for Reasoning about People's Social Preferences

Kobi Gal
Artificial Intelligence Research Group
Division of Engineering and Applied Sciences
Harvard University

Computers today are commonplace. As a result, they interact with people 
in diverse environments (e.g., e-commerce, dialog systems, multi-robot 
systems). In social environments, people's behavior is affected by a 
multitude of factors. For example, research in behavioral economics has 
established that many people exhibit a preference for choices that 
benefit others well as themselves. In order for computer agents to 
interact well with people, they must be able to represent and to reason 
about these factors within a computational framework.

This talk will present a model that represents and learns people's 
social preferences in games. The model predicts how a human player is  
likely to react to different actions of another player, and these
predictions are used to determine the best possible strategy for that 
player. Data collection and evaluation of the model were performed on a 
negotiation game in which humans played against each other and against 
computer models playing various strategies. Our computer player 
trained on human data outplayed computer players that employed
traditional game theoretic strategies, as well as humans. It also
generalized to play people and game situations it had not seen before.