Automatic Induction of Abduction and Abstraction Theories from Observations

摘要

Traditional Machine Learning approaches are based on single inference mechanisms. A step forward concerned the integration of multiple inference strategies within a first-order logic learning framework, taking advantage of the benefits that each approach can bring. Specifically, abduction is exploited to complete the incoming information in order to handle cases of missing knowledge, and abstraction is exploited to eliminate superfluous details that can affect the performance of a learning system. However, these methods require some background information to exploit the specific inference strategy, that must be provided by a domain expert. This work proposes algorithms to automatically discover such an information in order to make the learning task completely autonomous. The proposed methods have been tested on the system INTHELEX, and their effectiveness has been proven by experiments in a real-world domain.