In order to overcome the shortcoming of the longtime training and the frail generalization power of classical neural networks

this paper proposes a new covering classification algorithm based on constructive neural networks. The algorithm starts with the sample data directly and clustering analysis is executed on a hypersphere to find a sample with the max density

and then the intersection between the positive half-space of the hyperplane and sphere

called“sphere neighborhood”

is obtained

by which the training problem of neural networks may be transformed into the covering problem of point sets. Thus the new algorithm can reduce the traditional learning complexity. At the same time

the optimization of the neural network is also considered and computer simulation results show that the proposed neural network is quite efficient.