A new class of neural networks suitable for intelligent control

The authors deal with a class of nonlinear neural networks which are capable of forming decision regions of higher complexity than the multilayer perceptron. These nonlinear neural nets have been used in many applications, from nonlinear classification to intelligent control. They have the same feedforward topology as the multilayer perceptron, except that, in certain nodes, a nonlinear operation is performed on the inputs before the sigmoidal function is applied. In many cases, these nonlinear nets have demonstrated higher classification accuracy than comparably sized multilayer perceptrons. In addition, they require a smaller number of nodes to achieve the same accuracy as the multilayer perceptron, thereby reducing the required interconnections among processing elements in circuit implementations. The generation of an appropriate training set from the target function is also discussed. These training procedures can make additional improvements in the performance of the neural net by increasing the classification accuracy and reducing the required training time. Simulation results are presented