Speed-up of learning in second order neural networks and its application to model synthesis of electrical devices

Abstract

We use neural networks to approximate the terminal behaviour of electrical devices, maintaining the parameter dependencies. To accelerate the approximation time, we have improved the adaption rule by an adaptive evaluation of the learning parameters on the base of second-order sigma-pi neurons. The network paradigm is then automatically transformed either into a netlist of an electrical subcircuit (for example, SPICE-simulation) or into a mathematical description language (for example, a behavioural simulator like SABER). Examples demonstrate the very accurate representation of nonlinear electrical devices for circuit simulation.