Mining of agent behaviors

This paper presents a pattern based design method of agent interaction protocols, which are essential in most of multi-agent systems. Since most protocols define complex sets of interactions, we introduce an action pattern to represent abstract properties of a set of interactions. By focusing on hierarchical structures of agent interactions, an action pattern is expressed in an ordered tree with structure variables. Since a variable can be substituted for adequate sub-interactions, the pattern becomes a canonical representation of a class of interactions. In the proposed method, we design a protocol by reusing action patterns. We also show a datamining method for automatically obtaining patterns from histories of interactions.