Geun Bae Song, Man Keun Han, Haing Sei Lee

The KIPS Transactions:PartB , Vol. 8, No. 3, pp. 237-242, Jun. 2001
10.3745/KIPSTB.2001.8.3.237,   PDF Download:

Abstract

Like the classical generalized Lloyd algorithm (GLA), the Kohonen learning algortihm (KLA) algorithm approaches a solution of a cost function with a gradient descent step. Therefore, for complex high-dimensional vector quantization problems, KLA are entrapped in local minima although it is efficient and convenient to use. This paper proposes a Cauchy training method, a kind of simulated annealing method, in order to overcome the local minimum entrapment problem of the KLA. But its training speed is too slow. So this paper proposes another method which consists of cascade connected the Cauchy training and the KLA. As a result, the cascaded method not only overcomes local minima like the Cauchy training but also takes less training time than it.

Statistics
Cite this article