Research On Extraction Of Rules From Binary Neural Networks

A binary neural network (BNN) applies to problems in Boolean space,Extraction of rules is a important research area of it. When a BNN has been trained,because the learning algorithm for it is various,some binary neurons perhaps belong to a kind of linearly separable (LS) series,and some others perhaps belong to another kind of LS series. So it is very significative for extraction of rules from a BNN that the general judging methods and logical meanings of all those LS series are studied.Firstly,the research actualities and some basic theories are introduced,then by discussing the KT method,MofN rule expression and the difference between rules of implication and rules of equivalence,a WTA method based on analysis of the weights and threshold in a binary neuron is proposed. Hypercubes,Hamming spheres,SP functions and cartesian spheres are several important LS series in Boolean space. In this paper,the general judging and constructing methods of them in a BNN are built. Some new concepts such as LEM rules,GEM rules and cartesian spheres are defined. Moreover,cartesian spheres are proved to be a kind of LS series. The logical meanings of Hamming spheres and cartesian spheres are also analyzed.Based on analyzing the relationship between linear separability and a connected set in Boolean space,the particular effect of a restraining neuron in extraction of rules from a BNN is discussed,and that effect is explained through a example called a MIS problem in Boolean space. In this paper,a pattern match learning algorithm of BNNs is proposed. When a BNN has been trained by the algorithm,all the binary neurons of hidden layer belong to one or more LS series,If the logical meanings of those LS series are clear,the knowledge in the BNN can be dug out...