Multiple Neural Network Model Interpolation

This paper presents an efficient method for extracting a multi-model interpolation function from a nonlinear system. The multi-model interpolation function consists of couple simplified time-varying models in neural-network structure to dynamically approximate the nature of the physical phenomena to be interpolated and extrapolated. The purpose of using the multi-model interpolation function is to perform a real-time approximation. This paper demonstrates the interpolation in a simulated environment, the underwater acoustic transmission loss generated from the NAVY-standard acoustic propagation-loss model ASTRAL, which is not suited to real-time operation. The interpolation includes initial learning period that is on the order of 20 minutes (more or less time depends on the size of the parameter intervals and the complexity of the ocean environment), and the subsequent interpolation speed will be measured in fractions of a second, a several orders-of-magnitude improvement over conventional calculations. In addition, for the example presented here, the interpolation error is within 1% of the actual transmission-loss value in a root-mean-square (RMS) sense.