Discovering strategic multi-agent behavior in a robotic soccer domain

Abstract

This paper presents a method for multi-agent strategic modeling (MASM) applied in a robotic soccer domain. The method transforms multi-agent action sequences into a visual graph-based diagram, called action graph. Graph nodes are further augmented with additional domain knowledge. Using hierarchical clustering, action graph nodes are merged by utilizing domain-specific distance function. This step results in an abstract graphical model of agent behavior. Then, sub-graphs describing relevant agent behavior are used as input for association rule mining algorithm. The final output of MASM are strategic action concepts in the form of abstract action graph and association rules.