Model-free controller design for discrete-valued input systems based on autoencoder

Switching control is an effective control technique for control systems equipped with low-resolution actuators. The controller design problem for this class of control system can be formulated as the construction of a mapping between the observed outputs and the discrete inputs, that is, the construction of a switching surface. The mapping can be learned by neural network; however, the training result is sensitive to the initial weights, especially when a redundant structure of the network is selected. In this paper, a controller design method based on a neural network with autoencoder is discussed. An autoencoder learns the identity mapping at each layer. As a result, the output from each layer automatically encodes the feature vectors. The trained weight is used as a suitable initial weight for overall supervised learning. Numerical simulations show that the proposed method can reduce stochastic variance and avoid overfitting, especially for redundant neural controllers.