Successful application of Machine Learning to certain real-world situations sometimes requires to take into account relations among objects. Inductive Logic Programming, being based on First-Order Logic as a representation language, provides a suitable learning framework to be adopted in these cases. However, the intrinsic complexity of this framework, added to the complexity of the specific application context, often requires pure induction to be supported by various kinds of meta-information on the domain itself and/or on its representation in order to prune the search space of all possible definitions. Indeed, avoiding the exploration of paths that do not lead to any correct solution can greatly reduce computational times, and hence becomes a critical issue for the performance of the whole learning process. In the current practice, providing such information is often in charge of the human expert. It is also a difficult and error-prone activity, in which mistakes are highly probable because of a number of factors. This makes it desirable to develop procedures that can automatically generate such information starting from the same observations that are input to the learning process. This paper focuses on a specific kind of meta-information: the types used in the description language and their related domains. Indeed, many learning systems known in the literature are able to exploit (and sometimes require) such a kind of knowledge to improve their performance. An algorithm is proposed to automatically identify types from observations, and detailed examples of its behaviour are given. An evaluation of its performance in domains with different characteristics is reported, and its robustness with respect to incomplete observations is studied.
展开▼
