The state-sampled controller: Trainable-network control using traditional design and analysis principles with an application to small-satellite attitude control

The State-Sampled Network is a trainable network capable of implementing complex control functions using low cost computers. It is similar to the CMAC network developed by James Albus in 1975, but is based on digital image processing techniques and multidimensional Fourier transforms of control functions. All linear and a wide variety of nonlinear control laws can be shown to be band-limited, thus supporting Albus' premise of smooth control functions. The result is ideal spatial sampling of control functions, thereby reducing the number of network weights.; Training of the State-Sampled Controller, which is a State-Sampled Network used as a controller, can be achieved by supervised learning, parallel controller imitation and optimization, mathematical sampling, and sequential quadratic programming to yield a one-step-ahead optimal controller. Training goals can be mathematically determined by quadratic cost functions. Training is typically faster than CMAC, can be performed in real time, and requires substantially less memory.; Software simulation and hardware tests illustrate design methods and characteristics of the State-Sampled Controller. Controllers for inverted and cantilevered pendulums and a three-axis attitude controller for a small satellite are offered as design examples.