A feedforward neural network with function shape autotuning

In this paper, a novel shape-tunable feedforward neural network is proposed. Based on the steepest descent method, an autotuning algorithm that enables the proposed neural network to possess the ability of automatic shape-tuning is derived Due to the ability of auto-shaping, the flexibility and nonlinearity capacity of the neural network is increased significantly. Furthermore. the novel feature of automatic shaping prevents the nonlinear neurons from saturation, and therefore the scaling procedure. which is usually unavoidable for the traditional fixed-shape neural networks, becomes unnecessary. Simulation results indicate that the proposed shape-tunable neural network gives better agreement than the traditional fixed-shape one does, even though fewer nodes are used. Moreover, the convergence properties are more superior. To demonstrate the capability of the proposed shape-autotuning neural networks to a great extent. we adopted it as a learning-type direct controller. Some related problems were studied. Copyright (C) 1996 Elsevier Science Ltd