Behavior-analysis and -prediction for agents in real-time and dynamic adversarial environments

We present an approach for recognition and subsequent prediction of spatio-temporal patterns in a physical real-time environment. The motivation is to provide a domain-independent approach for the analysis of agent's behavior in adversarial multi-agent scenarios. The goal is to create an opponent- specific model, which is used for behavior prediction. We develop a framework for representing a set of hierarchically structured facts, events and actions using temporal logic. Recognition, learning, and prediction is performed using a probabilistic approach utilizing Bayesian Networks. The system is applied to the domain of the RoboCup 3D Simulation League and evaluated with regard to the recognition-, prediction- and realtime capabilities.