Various recurrent neural network architectures for solving the problems of parameter estimation in dynamical systems are presented. The architectures based on precomputation of weight/bias information (Hopfield neural network), direct gradient computation with and without normalisation and output error method are developed. A typical computer simulation result is given.