Inductive Learning From Input And Output

As a research area at the intersection of Machine Learning and Logic Programming,Inductive Logic Programming has not only the theoretical knowledge of Logic Programming but the methods and applications of Machine Learning as well.Also,Inductive Logic Programming learns the first-order rule from background knowledge by using the theoretical knowledge of Logic Programming,and this can help Inductive Logic Programming avoid two main limits in Machine Learning: the limit of knowledge representation and the limit of using background knowledge.Since Stephen H.Muggleton first introduced the term "Inductive Logic Programming" in 1991,Inductive Logic Programming has developed for over two decades.In this two decades,many mature inductive learning systems had been designed by researchers,and these systems had contributed a lot to many research fields.Nevertheless,existed work on Inductive Logic Programming rarely considers at the input and output of examples.This paper aims to find a novel inductive learning task for the input and output of examples.Also,it will try to design the algorithm for this learning task and then finish the prototype system of this algorithm.First off,according to the basic knowledge and technology of Inductive Logic Programming and Horn Programs,this paper proposes a novel learning task.In this task,an example is a form of < I,O >,where I and O,which are both a set of atoms,stand for input and output respectively.In this case,the task means that the output O is the least Herbrand model of the learned Horn programs when it's given the input I.Secondly,this paper proposes the algorithm for this learning task.While designing the algorithm,this paper adds some new functions into this algorithm,such as least general generalization,checking the over-generalization and so on.Finally,this paper introduces the prototype system for this algorithm and runs some experimentation in this system.