An improved algorithm for Kohonen´s self-organizing feature maps

A modified algorithm is presented for the learning by self-organizing topology-preserving maps to improve the piecewise-correct problem that arose frequently with the original self-organizing maps. The problem is generally caused by two dominant factors existing in the learning procedure of the original algorithm. One is the initial-sequence-order problem. The present algorithm efficiently reduces the influence of these two factors and successfully guides the network to form a topologically correct map. The proposed algorithm adopts a dynamic network that allows cells to be inserted and deleted, and it adds the Coulomb effect to the learning factor. Simulation results indicate that the modified algorithm performs well in learning the mapping of a two-dimensional input vector distribution using a one-dimensional network