From Basic Agent Behavior to Strategic Patterns in a Robotic Soccer Domain

The paper presents an algorithm for multi-agent strategic modeling (MASM). The method applies domain knowledge and transforms sequences of basic multi-agent actions into a set of strategic action descriptions in the form of graph paths, agent actions, roles and corresponding rules. The rules, constructed by machine learning, enrich the graphical strategic patterns, which are presented in the form of graph paths. The method was evaluated on the RoboCup Soccer Server Internet League data. Tests showed that the constructed rules successfully captured some decisive offensive moves and some major defense flaws, although the description itself was a bit awkward and needed interpretation by a human expert. Povzetek: Predstavljen je sistem, ki si uči strateških vzorcev obnašanja iz enostavnega opazovanja gibanja agentov v domeni robotskega nogometa.