Learning the Sequential Coordinated Behavior of Teams from Observations

The area of agent modeling deals with the task of observing other agents and modeling their behavior, in order to predict their future behavior, coordinate with them, assist them, or counter their actions. Typically, agent modeling techniques assume the availability of a plan- or behavior-library, which encodes the full repertoire of expected observed behavior. However, recent applications areas of agent modeling raise challenges to the assumption of such a library, as agent modeling systems are increasingly used in open and/or adversarial settings, where the behavioral repertoire of the observed agents is unknown at design time. This paper focuses on the challenge of the unsupervised autonomous learning of the sequential behaviors of agents, from observations of their behavior. The techniques we present translate observations of the dynamic, complex, continuous multi-variate world state into a time-series of recognized atomic behaviors. This time-series is then analyzed to find repeating subsequences characterizing each team. We compare two alternative approaches to extracting such characteristic sequences, based on frequency counts and statistical dependencies. Our results indicate that both techniques are able to extract meaningful sequences, and do significantly better than random predictions. However, the statistical dependency approach is able to correctly reject sequences that are frequent, but are due to random co-occurrence of behaviors, rather than to a true sequential dependency between them.