As multi-agent systems grow in complexity and diversity, they become increasingly difficult to design. Agents are described in terms of their behaviour, typically trained by an expert who prepares knowledge representations or training data for supervised machine learning.
To reduce development time, agents could learn by observing the behaviour of other agents. This thesis describes an effort to train a RoboCup soccer agent by capturing data from existing players, generating a knowledge representation, and using a real-time scene recognition system. The trained agent later exhibits behaviour traits similar to the observed agent and can appear to completely imitate the behaviour of the original; the process requires little human intervention.
Experiments are performed using three agents of varying complexity. The “scene” knowledge description format, and simple scene matching algorithm, are limited to imitation of stateless and deterministic agent behaviours. Future work includes improving the matching algorithm and developing higher-level behaviour models.