Recently several people have studied the effects of multiple sources of knowledge (see [2] for a survey). All approaches share the problem of conflict resolution which is one of the issues tackled by the two probabilistic approaches examined in this paper.
Gams et al., [6] made several experiments with several knowledge bases when classifying new instances. They tried two different strategies to obtain the classification : best-one which uses the opinion with highest confidence factor (this is a strategy similar to ours) and the majority strategy where confidence factors add up in order to reach a conclusion. This latter strategy represents a kind of combination of different opinions. The authors made extensive experiments on artificial domains and the results showed that the best-one strategy scored better whenever few knowledge bases were used. When the number of knowledge bases increased the majority strategy was tbetter. These results seem to suggest that if flexible matching were introduced (which would increase the potential number of opinions) the probabilistic combination strategies examined in this paper might perform better.
Brazdil and Torgo [3] used different learning algorithms to generate several knowledge bases which were combined into one using a kind of best quality strategy. This work suggested that good results could be obtained with this simple strategy, but no comparisons were made with other possible combination strategies.