Real-time neural controller implemented on parallel architecture

A problem with simulation of multilayer neural network on transputer array is described in this article. The decomposition and mapping on given architecture is proposed as well as a simple message passing scheme. Practical experiments with inverted pendulum plant are described including useful hints for setting neural network architecture and tuning neural networks parameters. Then neural controller is designed and criterion function is defined for given control problem. Finally the real-time aspects of the controller design are outlined.