NEUROCONTROL OF NONLINEAR DYNAMICAL-SYSTEMS WITH KALMAN FILTER TRAINED RECURRENT NETWORKS

Although the potential of the powerful mapping and representational ca pabilities of recurrent network architectures is generally recognized by the neural network research community, recurrent neural networks ha ve not been widely used for the control of nonlinear dynamical systems , possibly due to the relative ineffectiveness of simple gradient desc ent training algorithms. Recent developments in the use of parameter-b ased extended Kalman filter algorithms for training recurrent networks may provide a mechanism by which these architectures will prove to be of practical value. This paper presents a decoupled extended Kalman f ilter (DEKF) algorithm for training of recurrent networks with special emphasis on application to control problems. We demonstrate in simula tion the application of the DEKF algorithm to a series of example cont rol problems ranging from the well-known cart-pole and bioreactor benc hmark problems to an automotive subsystem, engine idle speed control. These simulations suggest that recurrent controller networks trained b y Kalman filter methods can combine the traditional features of state- space controllers and observers in a homogeneous architecture for nonl inear dynamical systems, while simultaneously exhibiting less sensitiv ity than do purely feedforward controller networks to changes in plant parameters and measurement noise.