Most work in inductive learning tends to discuss the learning method details, but little attention is paid to the problem of how the learned rules are used. This paper shows that different problem solving strategies can lead to very different accuracy results. This clearly indicates the importance of these strategies when comparing performance of learning systems.
Our experiments used an attribute-based learning system to generate theories which were then tested with different problem-solving strategies. This problem is however extensible to other types of learning systems. In general, whenever different sources of knowledge are used (including in multi-strategy learning systems) we need a method for conflict resolution.
Experiments were made on three real world domains. Their goal was to observe if different classification strategies could lead to different results. The following strategies were used : two well known expert systems approaches, MYCIN [11] certainty factors and PROSPECTOR's [5] odds), together with the best rule strategy.
The next section describes briefly the inductive system used in the experiments. Section 3 presents the different strategies and section 4 the experiments carried out.