Achieving semantic interoperability in multi-agent systems: A dialogue-based approach

Software agents sharing the same ontology can exchange their knowledge fluently as their knowledge representations are compatible with respect to the concepts regarded as relevant and with respect to the names given to these concepts. However, in open heterogeneous multi-agent systems, this scenario would be very unlikely, because it would require all involved system developers to reach consensus on which ontology to use. Furthermore, different agents may regard different concepts as relevant which causes their ontologies to differ in granularity and scope. In such an environment, the agents must possess the right conversational skills to effectively exchange information even when the speaker's ontology is only approximately translatable to the hearer's ontology. Furthermore, the agents must be able to autonomously establish an ontology translation by exchanging parts of their ontologies. In this thesis, we propose a layered communication protocol in which the agents gradually build towards a semantically integrated system by establishing minimal and effective shared ontologies. We will use the formal notions of sound and lossless communication to state the requirement that sufficient information should flow between the agents in a correct manner. The communication protocol detects when communication is ineffective and applies techniques for ontology exchange to build a shared ontology of minimal size. In this way, the agents exchange ontological information on an as-need basis. Agents first try to cope with the situation as it is; when communication fails to be effective, the agents seek a minimal solution which solves their communication problem at hand. The communication mechanism consists of three layers. The upper layer of the protocol is the Normal Communication Protocol (NCP) which deals with the kind of social interaction that agents normally exhibit when no ontology problems exist in the system. Every conversation starts in the NCP layer. If the agents fail to understand each other, the agents switch to the middle layer in the protocol which is the Concept Definition Protocol (CDP). In this layer, the agents explain the meaning of a concept to each other by exchanging concept definitions. The meaning of a concept is explained in terms of other concepts. If the communication difficulties are so severe that the agents do not even understand each other's concept definitions, the agents switch to the lowest layer in the protocol, i.e. the Concept Explication Protocol (CEP). In CEP, the agents exchange the meaning of a concept using non-symbolic communication, e.g. by pointing to examples of the concept. We tested our system, called ANEMONE, in three ways. Firstly, we provide a mathematical analysis. By formalizing the communication protocol, we can give solid proofs that it possesses the desirable properties. Secondly, we perform simulation experiments. By making the agents communicate in a simulation environment, we can analyze whether the agents indeed build a minimal communication vocabulary. Thirdly, we describe a case study with heterogeneous internet news agents. We show how these agents successfully exchange information on news articles, despite initial difficulties caused by heterogeneous ontologies.