The experiments carried out did not admit rules whose conditional parts were not completely satisfied (i.e. flexible matching). It would be interesting to see how accuracy would be affected if flexible matching were used.
The experiments could be extended to other datasets and other learning systems. Experiments with some existing ILP systems are under consideration. This later extension requires not only parallel evidence combination methods (as presented in the paper), but also sequential combination methods to cover the case of rule chaining.
The main extension should be to broaden the range of methods used to combine rules. Some effort could be invested towards the use of a model which does not exhibit the limitations of conditional independence [8] that both certainty factors and degrees of sufficiency and necessity suffer from. Some experiments could be done with Dempster-Shaffer [4] theory of evidence and Pearl's belief networks [10].