Modified relaxation method for solution of continuous recurrent neural networks

The derivation of a modified relaxation algorithm is presented followed by demonstration examples. The algorithm converges very well for continuous recurrent neural networks with both low and high gain neurons. This enables one to simulate recurrent Hopfield networks with both "soft" and "hard" continuous activation functions. The algorithm is suitable for large systems since the computational effort is proportional only to the system size, in contrast to the commonly used Newton-Raphson method where power relationships exist.<
>