基于解释的人工神经网络集成学习方法

基于知识的人工神经网络是集成学习方法领域中最主要的研究方向，它利用领域知识决定神经网络的初始拓扑结构和联接权值的设置。一方面利用领域知识加速了网络的学习，另一方面利用神经网络的鲁棒性减少了领域理论不完善和数据噪声的影响，从而较好的解决了不完善领域理论的学习问题。本文对这方面的几个主要系统ＫＢＡＮＮ、ＲＡＰＴＵＲＥ进行了分析和评价，指出了它们在网络拓扑结构的动态改变、深层网络的加速学习和结果网络到规则的转译等方面的弱点，并因此提出了一种有效的集成学习方法——基于解释的人工神经网络。

Explanation-based artificial neural network integrated learning approach

nowledge--based artificial neural network is the main topic in the area of integrated learning method research. It integrates existing knowledge into neural networks by defining the network topology and setting initial link weights. It accelerates network learning by using domain theory. Because of the robust of neural network it minimizes the errors that arise when using imperfect domain theories. This article analyzes and evaluates some systems such as KBANN and RAPTRUE. It also points out some defects such as changing the network architecture, learning of deep network, translating from the network result to rules, and presents a new integrated learning method named EBANN which can overcome the difficulties metioned above.