Connectionist learning control systems: submarine depth control

Control system design for vehicles with highly nonlinear, time-varying, or poorly modeled dynamics is considered. The use of connectionist systems as learning controllers is proposed. The ability of connectionist systems to approximate arbitrary continuous functions (e.g., control laws) overcomes the usual memory-intensive nature of learning systems. The backpropagation algorithm is extended to allow the connectionist system to learn to function as a closed-loop controller and to force the dynamics of the closed-loop system to match the prespecified dynamics of a reference system. An example of the application of this algorithm to the depth control of an autonomous underwater vehicle is included