Abstract
Within ILP much effort has been put into designing methods that are complete for hypothesis finding. However, it is not clear whether completeness is important in real-world applications. This paper uses a simplified version of grammar learning to show how a complete method can improve on the learning results of an incomplete method. Seeing the necessity of having a complete method for real-world applications, we introduce a method called ⊤-directed theory co-derivation, which is shown to be correct (ie. sound and complete). The proposed method has been implemented in the ILP system MC-TopLog and tested on grammar learning and the learning of game strategies. Compared to Progol5, an efficient but incomplete ILP system, MC-TopLog has higher predictive accuracies, especially when the background knowledge is severely incomplete.