Q Duong, MP Wellman, S Singh, and M Kearns
Factored models of multiagent systems address the complexity of joint behavior by exploiting locality in agent interactions. History-dependent graphical multiagent models (hGMMs) further capture dynamics by conditioning behavior on history. The challenges of modeling real human behavior motivated us to extend the hGMM representation by distinguishing two types of agent interactions. This distinction opens the opportunity for learning dependence networks that are different from given graphical structures representing observed agent interactions. We propose a greedy algorithm for learning hGMMs from time-series data, inducing both graphical structure and parameters. Our empirical study employs human-subject experiment data for a dynamic consensus scenario, where agents on a network attempt to reach a unanimous vote. We show that the learned hGMMs directly expressing joint behavior outperform alternatives in predicting dynamic human voting behavior, and end-game vote results. Analysis of learned graphical structures reveals patterns of action dependence not directly reflected in the original experiment networks.
A previous version was presented at the Fifth SNA-KDD Workshop on Social Network Mining and Analysis, August 2011.