Experimental study of direct adaptive SPSA control system with diagonal recurrent neural network controller

Abstract

A direct adaptive simultaneous perturbation stochastic approximation (SPSA) control system with a diagonal recurrent neural network (DRNN) as controller was examined by simulation. Different hidden number DRNNs were used in the SPSA system to study the relationship between the performance and neural network architecture and parameters. Results were compared with those of a SPSA using a forward neural network (FNN) controller. Study shows that a direct adaptive SPSA control system with DRNN has a simpler architecture, a smaller size of parameter vector and a faster convergence rate. The system has a steady-state error and is sensitive to SPSA coefficients and termination condition. For real-time trajectory control purposes, further improvement of direct adaptive SPSA approach is required.