One of the key advantages of our approach, and confirmed by our case studies, is that we have implemented a set of aggregation functions that can be used from a menu (for example in the RuleTool) when scripting a set of rules. The user can therefore easily use these functions to look at different ways of taking different parts of the background knowledge into account (e.g. should certain sources of information be ignored, should unanimity in sources be sought or is a majority sufficient, what are the exceptions to certain parts of the background knowledge, etc). Each aggregation function is implemented as an aggregation predicate in the knowledgebase, and so can be called via a condition in a fusion rule. This means aggregation is context-dependent. The basic idea of an aggregation predicate is explained in the following paper. Details on some basic forms of aggregation are discussed and compared in the webpages on the weather case study. Click on "Weather" on the main menu.
We have developed further kinds of agregation predicates for particular kinds of knowledge fusion, in particular for "Uncertainty", "Inconsistency", "Expectations", and "Time". Click on the menu main on the left for more details on these. Of course the range of aggregation predicates is open-ended. Some are generic, others are domain specific. However, the key requirement is that they need to be formally defined with well-understood logical properties, and a transparent semantics. Ad hoc approaches to aggregation are counter-productive in larger-scale knowledge fusion tasks.
Contact a.hunter@cs.ucl.ac.uk or +44
20 7679 7295.
Back to Fusion Rule Technology homepage.