An Approach for the Qualitative Analysis of Open Agent Conversations

This paper presents an approach for the qualitative analysis of data obtained from past communicative interactions in an open multiagent system. Such qualitative analysis focuses on the use of high-level agent communication languages to infer theories about agents with mental states which are normally not accessible for the outside observer. The inference of these theories, or context models, is based on logging semantic data available from protocol execution traces and using this information as samples for the application of data mining algorithms. These context models can be applied both by system developers and agents themselves at run-time for various tasks, e.g. to predict future agent behaviour, to support the process of ontological alignment in communication, or to assess the trustworthiness of agents. An implementation of the approach presented is also given, the ProtocolMiner tool, which automates the building of context models from arbitrary protocol executions.