Covering Algorithm Based on Neighborhood Search and its Applications

The theory and applications of artificial neural networks have developed rapidly since the mathematical model of neuron was presented, but the design of network structure for a certain problem was a roadblock over a long period of time. In 1990s, the covering algorithm for forward neural network was put forward, this algorithm is a constructive machine learning method, it designs network with sphere neighborhoods built upon sample data. The theoretical analysis indicates that the general ability of this algorithm is in the inverse ratio of the number of covering domains. Some techniques is performed while constructing the sphere neighborhood set to simplify the network, such as changing the sphere neighborhoods center to the barycentre of points covered by a domain, deleting the points covered by a sphere neighborhood. In this paper, a new constructive algorithm which combines the covering algorithm and neighborhood search is presented and applied to some databases on the web site of UCI and text categorization, the results show that the algorithm can reduce the number of sphere neighborhoods with higher degree of classifying accuracy.