Approximating rail locomotive dynamics using the SOCM network

Abstract

We demonstrate the self-organising continuous map (SOCM), a novel use for the self-organising map/learning vector quantisation network that widens the scope of the SOM architecture. We use the SOM/LVQ network as a distribution service, apportioning an equal quantity of work to a number of intelligent nodes. Advantages include improved accuracy, effective and balanced multi-processing for small cluster systems, and potentially large reductions in training and recall times. The example problem chosen uses neural networks to model force dynamics of a coal train. The SOCM configuration used consists of a SOM network where each node is a backpropagation (BP) network. We show that the collection of as few as two BP networks gives at least a 30% reduction in approximation error when compared to the original BP network. We discuss how the SOCM approach could be used in other areas of artificial intelligence, including evolutionary systems, parallel processing, error balancing, hybrid networks, and online training.