Monitoring group behavior in goal-directed agents using co-efficient plan observation

Purposeful, time- and cost-oriented engineering of Multi-Agent Systems (MAS) requires developers to understand the relationships between the numerous behaviors exhibited by individual agents and the resulting global MAS behavior. While development methodologies have drawn attention to verification and debugging of single agents, software producing organizations need to validate that the MAS, as a cooperative system exhibiting group behavior, is behaving as expected. Recent research has proposed techniques to infer mathematical descriptions of macroscopic MAS behavior from microscopic reactive and adaptive agent behaviors. In this paper, we show how similar descriptions can be adjusted to MAS composed of goal-directed agent architectures. We argue that goal-hierarchies found in Requirements Engineering and Belief Desire Intention (BDI) architectures are suitable data structures to facilitate a stochastic modeling approach. To enable monitoring of agent behaviors, we introduce an enhancement to the well-known capability concept for BDI agents. So-called co-efficient capabilities are a novel approach to modularize crosscutting concerns in BDI agent implementations. A case study applies co-efficient plan observation to exemplify and confirm our modeling approach.