An automated Hammerstein recurrent neural network for dynamic applications

This paper presents an automated Hammerstein recurrent neural network (HRNN) associated with a self-construction learning algorithm capable of building the network with a compact state-space representation from the input-output measurements of dynamic systems. The proposed HRNN is constituted by two connectionist networks-a static nonlinear network cascaded with a linear dynamic network. The self-construction algorithm is devised to automate the HRNN construction process via three mechanisms: an order determination scheme, a weight initialization method, and a parameter optimization method. With the learning algorithm, trial and error on the selection of network sizes or parameter initialization can be totally exempted. Computer simulations on nonlinear dynamic system identification validate that the proposed HRNN can closely capture the dynamical behavior of the unknown system with a compact network size.