Neural network architectures for active suspension control

The authors describe several methods which can be utilized to train a neural network to control both linear and nonlinear active suspension systems. It is demonstrated that the networks can simply be trained to emulate existing controllers, be trained to become an adaptive controller for a plant by utilizing a second network which has learned to emulate the plant, or be trained to control an unmodeled system through the provision of a properly chosen performance function and error signal. The latter case is quite interesting in that the neural network learns to control an unknown system guided by a training signal developed from heuristic knowledge about generic plant properties and qualitative performance functions. Comparisons of the neural net controllers and conventional controllers are discussed in terms of performance and efforts needed to obtain reasonable performance from both approaches