Efficient recurrent neural network training incorporating a priori knowledge

A new technique for efficient training of Hopfield network models using ite rative training algorithms is described and demonstrated. This technique is useful for producing stable Hopfield networks, using recently derived resu lts concerning stability conditions for the Hopfield network. Since the mod ified training algorithm ensures the stability of the network, no off-line verification of stability is required. This technique can be applied to Hop field networks of any size, and therefore is tested for three different ran domly selected sizes. Additionally, the advantage of utilising a priori inf ormation about the plant is also tested and the results are compared with t hose cases where no such information is available. This information is used to determine critical dynamic properties of the network model, necessary f or the network's ability to generalise. In all tests, the networks are trai ned with modified genetic algorithms, using different initial starting poin ts for the algorithm. (C) 2000 IMACS/Elsevier Science B.V. All rights reser ved.