Analog Recurrent networks for signal tracking and control applications
Faculty of Computing, Health and Science
School of Engineering and Mathematics / Centre for Communications Engineering Research
Techniques of realizing CMOS continuous-time analog recurrent neural networks are studied in this paper. We discuss the application of an analogue recurrent neural network to learn and track the dynamics of an industrial robot. The observations made from this study suggest that RNNs (similar to those in Fig. 1) can be applied to the control of real systems that manifest complex properties - specifically, high-dimensionality, non-linearity and requiring continuous action. Examples of these real systems include aircraft control, satellite stabilization, and robot manipulator control. We conclude that robust controllers of partially observable (non-Markov). systems require real-time electronic systems that can be designed as single-chip Integrated Circuits (CMOS IC). This paper explored such techniques and identified suitable circuits. The synaptic weights are modeled as variable gain cells that can be implemented with a few MOS transistors. For the specific purpose of demonstrating the trajectory learning capabilities, a periodic signal with varying characteristics is used. The developed architecture, however, allows for more general learning tasks typical in applications of identification and control. On-line versions of the synaptic update can be formulated using simple CMOS circuits. The simulated network contains interconnected recurrent neurons ·with continuous-time dynamics. The system emulates random-direction descent of the error as a multidimensional extension to the stochastic approximation. To achieve unsupervised learning in recurrent dynamical systems we propose a synapse circuit which has a very simple structure and is suitable for implementation in VLSI.