Recursive least-squares approach to self-organizing maps

The code update of the self-organizing map is formulated as a parameter estimation problem, which is then solved by the recursive least-squares method. If the lattice or the size of the neighborhood is altered during the organization, the estimation problem becomes time variant. Time weighting is used in the recursive least-squares algorithm. The proposed approach produces a simple code update formula which preserves the parallelism and computational simplicity of the original algorithm based on stochastic approximation. The computational burden of the code update is increased by 2n flops per iteration, where n is the number of the code vectors, compared to the original code update formula. Each code vector has a separate update equation with a scalar gain provided directly by the algorithm. The properties of the algorithm presented are illustrated by simulation examples which demonstrate that the recursive least-squares method yields fast convergence of the code vectors.<
>