Temporal associative memory and function approximation with the self-organizing map

Abstract

We propose an unsupervised neural modelling technique, called vector-quantized temporal associative memory (VQTAM), which enables Kohonen's self-organizing map (SOM) to approximate nonlinear dynamical mappings globally. A theoretical analysis of the VQTAM scheme demonstrates that the approximation error decreases as the SOM training proceeds. The SOM is compared with standard MLP and RBF networks in the forward and inverse identification of a hydraulic actuator. The simulation results produced by the SOM are as accurate as those produced by the MLP network, and better than those produced by the RBF network; both the MLP and the RBF being supervised algorithms. The SOM is also less sensitive to weight initialization than MLP networks. The paper is concluded with a brief discussion on the main properties of the VQTAM approach.