RESUBMISSION: Management of Distributed Knowledge Sources for Complex Application Domains

In order to realise a truly distributed knowledge-based system not only the knowledge processing step has to be carried out in a distributed way, but also the knowledge acquisition step. This paper’s focus1 lies on the distributed knowledge sources of the SEASALT architecture [Reichle et al., 2009a] and their management and (optimised) querying using a Coordination Agent [Bach et al., 2008]. Within SEASALT knowledge modularisation is realised in the Knowledge Line that is based on the principle of product lines as it is known from software engineering. We apply this to the knowledge in knowledge-based systems, thus splitting rather complex knowledge in smaller, reusable units (knowledge sources). Moreover, the knowledge sources contain different kinds of information as well as there can also be multiple knowledge sources for the same purpose. Therefore each source has to be described in order to be integrated in a retrieval process which uses a various number of knowledge sources. A so called Knowledge Map organises all available knowledge sources that can be accessed by a Coordination Agent that creates individual requests and combines information. The term Knowledge Map originates in Davenport’s and Prusak’s work on Working Knowledge [Davenport and Prusak, 2000] in which they describe a Knowledge Map from the organisational point of view mapping human experts in a large organisation or company. We transfer this concept to an intelligent agent framework that coordinates different knowledge sources. The Coordination Agent navigates through the Map and subsequently queries the individual knowledge sources and thus creating an individual path through the map[ReichleSchmehl, 2008]. There are dependencies between knowledge sources, a dependency exists if one source’s output serves as another’s input and thus enforces a subsequent query. Since the dependencies between knowledge sources can take any form, the Knowledge Map is implemented as a graph where each knowledge source is represented by a node and directed edges denote the dependencies. Retrieval paths are computed based on the information a user gives in an individual query and the properties of the knowledge sources. Our current implementation provides an a-priori computation of the retrieval path using a modified Dijkstra algorithm to determine an optimal route over the graph. Considering knowledge sources, different characteristics, and aspects on which to assess knowledge source properties come to mind. The possible properties can refer to content (e.g. quality or topicality) as well as metainformation (e.g. answer speed or access limits). In detail