CHARACTERIZING AND PREDICTING AGENTS VIA MULTI-AGENT EVOLUTION

A method of predicting the behavior of software agents in a simulated environment involves modeling a plurality of software agents representing entities to be analyzed, which may be human beings. Using a set of parameters that governs the behavior of the agents, the internal state of at least one of the agents is estimated by its behavior in the simulation, including its movement within the environment. This facilitates a prediction of the likely future behavior of the agent based solely upon its internal state; that is, without recourse to any intentional agent communications. In the preferred embodiment the simulated environment is based upon a digital pheromone infrastructure. The simulation integrates knowledge of threat regions, a cognitive analysis of the agent's beliefs, desires, and intentions, a model of the agent's emotional disposition and state, and the dynamics of interactions with the environment. By evolving agents in this rich environment, we can fit their internal state to their observed behavior. In realistic wargame scenarios, the system successfully detects deliberately played emotions and makes reasonable predictions about the entities' future behavior.