Learning automated agents from historical game data via tensor decomposition

Abstract

War games and military war games, in general, are extensively played throughout the world to help train people and see the effects of policies. Currently, these games are played by humans at great expense and logistically require many people to be physically present. In this work, we describe how to automatically create agents from historical data to replace some of the human players. We discuss why game-theoretic approaches are inappropriate for this task and the benefits of learning such agents. We formulate a tensor decomposition formulation to this problem that is efficiently solvable in polynomial time. We discuss preliminary results on real world data and future directions.