Discovering process models for multi-agent systems from event logs

A process model discovered from an event log of a multi-agent system often does not fully cover certain viewpoints of its architecture. We consider those concerned with the structure of a model explicitly reflecting agent behavior and interactions. The direct discovery from an event log of a multi-agent system may result in an unclear model structure and over-generalizations of agent behavior. We suggest applying a compositional approach that yields architecture-aware process models of multi-agent systems. An event log of a multi-agent system is filtered by the behavior of individual agents. Then, a multi-agent system model is a composition of agent models discovered from filtered logs. We use an intermediate model, called an interface pattern, specifying agent interactions and representing the architecture of a multi-agent system. We design a collection of specific interface patterns modeling typical agent interactions. An interface pattern provides an abstract specification of interactions and has a part corresponding to the behavior of each agent. We use structural transformations based on morphisms to map agent models discovered from filtered logs on the respective parts in an interface pattern. If such a mapping exists, we guarantee that a composition of agent models preserves their soundness. We conduct a series of experiments to evaluate the compositional approach. Experimental results confirm the improvement in the structure of process models discovered using the compositional approach compared to those discovered directly from event logs. Keywords: Multi-agent systems, event logs, process discovery, Petri nets, composition