Technical Program

Paper Detail

Paper: PS-1B.7
Session: Poster Session 1B
Location: Symphony/Overture
Session Time: Thursday, September 6, 18:45 - 20:45
Presentation Time:Thursday, September 6, 18:45 - 20:45
Presentation: Poster
Paper Title: Gaussian Process Models Characterize Other-Regarding Strategies Over Multiple Timescales in a Dynamic Social Game
Manuscript:  Click here to view manuscript
Authors: Kelsey McDonald, Duke University, United States; William F. Broderick, New York University, United States; Scott Huettel, John Pearson, Duke University, United States
Abstract: A primary aim of computational neuroscience is to produce models of human behavior that meaningfully address population-level variability. Previous approaches to strategic interaction have used games with clearly-defined turns and limited choices, since these problems are amenable to tractable mathematical analysis. However, most real-world decisions are dynamic, involving simultaneous, coevolving decisions by each agent. Here, using a competitive game in which participants control the dynamics of an on-screen avatar against either another human or a computer opponent, we show that it is possible to quantify this dynamic coupling between agents. Despite the complexity of this behavior, modern nonparametric modeling methods can address the challenges posed by high-dimensional decision problems. We used Gaussian Processes to model the joint distributions of players' actions and identities (human or computer) as a function of game state. We show that this approach offers a natural set of metrics for quantifying instantaneous strategies, and that these metrics are linked to the models’ hyperparameters. Moreover, because opponent identity is part of the joint distribution, we can differentiate between effects due to opponent identity and effects due to game context. Our approach facilitates analysis at multiple timescales and suggests new classes of tractable paradigms for assessing human behavior.