基于高阶逻辑的复杂结构归纳学习研究

Survey of Inductive Learning from Complex Structured Data Based on Higher-order Logic

归纳学习的目的在于发现样例与离散的类之间的映射关系,样例及归纳的映射都需用某个形式化语言描述。归纳学习器采用的形式化语言经历了属性-值语言、一阶逻辑、类型化的高阶逻辑三个阶段,后者能克服前二者在知识表达及学习过程中的很多缺点。本文首先阐述了基于高阶逻辑的复杂结构归纳学习产生的历史背景;其次介绍了基于高阶逻辑的编程语言——Escher的知识描述形式及目前已提出的三种学习方法;复杂结构的归纳学习在机器学习领域的应用及如何解决一些现实问题的讨论随后给出;最后分析了复杂结构归纳学习的研究所面临的挑战性问题。

Inductive learning consists of finding mapping of examples into discrete classes. Examples and induced mapping are represented in some formal language. The formal language employed by inductive learners has undergone three stages, such as, attribute-value language, first-order logic, typed higher-order logic. The latter can overcome many shortcomings in knowledge representation and learning process which belong to the former two. This paper firstly provides a survey of background and context from which inductive learning from complex structured data based on higher-order logic arises. Secondly, how to represent knowledge by means of Escher, which is a typed higher-order logic programming language is demonstrated and three kinds of learning algorithm is introduced. The application of inductive learning from complex structured data based on higher-order logic in machine learning and how to solve some practical problems with which is discussed in following. Lastly, several challenging researching problems are identified.