Building Multiagent Behaviors from Observation

Multiagent behaviors play a vital role in coordinated robotics, scientific simulations, and surveillance systems. Training applications, for example, require authentic movements from the autonomous agents that work with and compete against the user. Similarly, surveillance systems must develop a model for normal group behaviors before they can identify unusual circumstances. In the case of mobile agents, the behaviors must generate movements for each individual based on its perception of the world. Our system automatically develops behaviors for autonomous mobile agents from the observation of the movements of real multiagent groups. It learns by watching. With these behaviors, we demonstrate a multiagent system that anticipates the future actions of the observed system and retargets behaviors to control different groups in novel circumstances.