Extension of a classical error functional and structure modification of continuous Hopfield neural networks

This paper addresses the problem of training multiple trajectories by means of continuous Hopfield neural networks in identification a control model of the financial flows of the Polish economy. There are two drawbacks of the networks learning procedure when solving this problem. First, a strong network sensitivity to small changes in the network weights as a result of multiple, nonlinear connections between its variables. Second, a poor quality of the network mapping resulting from the finiteness of the learning set describing unique properties of that system. To overcome these constraints, a few modifications of the basic learning procedure have been proposed. The crucial idea here considers extension of a classical error functional to three forms of penalty term, depending on the number of available data and the structure modification.