Practical polynomial expansion of input data can improve neurocomputing results

Multi-layered perceptron (MLP) neurocomputing networks may be slow-learners and produce nonanalytical results; however, when presented with appropriately conditioned input data MLPs offer very good generalization results for complex classification problems. The author recounts some experiences from, and techniques used in, applying MLPs and backpropagation to find successful and better solutions to otherwise difficult practical problems, which often require the determination of real-valued numbers. The most successful contributor has been the application of a polynomial expansion technique to the input data which has helped to reduce network size, speed up the training and improve accuracy in the neurocomputing results.<
>