Alternative Dynamic Network Structures for Non-linear System Modelling- Camera ready

Hopfield Neural Networks have been used as universal identifiers of non-linear systems, because of their inherent dynamic properties. However the design decision of the number of neurons in the Hopfield network is not easy to make, in order for the network model to have the necessary complexity, extra neurons are required. This poses a problem since the role of the states that these neurons represent is not clear. Adding a hidden layer in the Hopfield network model increases the complexity of the model without posing the extra states problem. Alternatively breaking the problem down by having different interconnected Hopfield networks modeling each state, also increase the complexity of the problem. A comparison between the three approaches (traditional Hopfield, Hopfield with a hidden layer, and multiple interconnected Hopfield networks) indicates equivalence between the three structures, but with the alternative cases having increased connectivity in the feedback matrix, and limited connectivity in the weight matrices.