engleski

Discrete time neural network synthesis using interaction activation functions

Abstract - A new very fast algorithm for synthesis of discrete-time neural networks (DTNN) is proposed. For this purpose the following concepts are employed: (i) introduction of interaction activation functions, (ii) time-varying DTNN weights distribution, (iii) time-discrete domain synthesis and (iiii) one-step learning iteration approach.. The proposed DTNN synthesis procedure is useful for applications to identification and control of nonlinear, very fast, dynamical systems. In this sense a DTNN for a nonlinear robot control is designed. As the contributions of the paper, the following items can be cited. A nonlinear, discrete-time state representation of a neural structure was proposed for one-step learning. Within the structure, interaction activation functions are introduced which can be combined with input and output activation functions. A new very fast algorithm for one step learning of DTNN is introduced, where interaction activation functions are employed. The functionality of the proposed DTNN structure was demonstrated with the numerical example where a DTNN model for a nonlinear robot control is designed. This DTNN model is trained to imitate a nonlinear robot control algorithm, based on the dynamics of the full robot model of RRTR-structure. The simulation results show the satisfactory performances of the trained DTNN model.

Recurrent neural networks; interaction activation functions; time-discrete domain synthesis; one-step learning; nonlinear dynamical systems; nonlinear robot control.

nije evidentirano