Neural network controller based on the rule of bang-bang control

Applying neural networks or fuzzy systems to the field of optimal control encounters the difficulty of locating adequate samples that can be used to train the neural networks or modify the fuzzy rules such that the optimal control value for a given state can be produced. Instead of an exhaustive search, this work presents a simple method based on the rule of bang-bang control to locate the training samples for time optimal control. Although the samples obtained by the proposed method can be learned by multilayer perceptrons and radial basis networks, a neural network deemed appropriate for learning these samples is proposed as well. Simulation results demonstrate the effectiveness of the proposed method.