engleski

Adaptive recurrent neural networks for robot control

Abstract - Effects that are generally accepted to take place in real neurones are pulse-coded inputs and outputs (input and outputs activation functions), a dynamic transmission function of the synapse (time-varying weights and interaction activation functions), and capacitive behaviour of the cell membrane (time-delay of signals). This paper follows the ideas of the synthesis of the biological NN prototypes in order to close in the behaviour of the real biological neurones, as close as possible. In this sense a new very fast algorithm for synthesis of adaptive recurrent discrete-time neural networks is proposed, where the following concepts are employed: (i) time-varying NN weights distribution, (ii) one-step learning iteration approach, and (iii) adaptation of interaction weight matrix. Following the proposed NN synthesis procedure the adaptive recurrent NN for an adaptive nonlinear robot control is designed.

Adaptive robot control; recurrent neural networks; one-step learning

nije evidentirano