A recurrent neural net approach to one-step ahead control problems

In this paper, we present a recurrent neural net technique to provide control actions for nonlinear dynamic systems. In most current neural net control approaches, two nets are usually required. One acts as a system emulator, and the other one is a controller network. Rather than using two nets, our system requires only one net which is the system emulator. In our proposed system, a neural net is used to learn the forward dynamics of the system, and the control signal is evolved from the output of the same net with use of an equation of motion. There is no need to learn the control law from another neural net, such as a system inverse net. The use of the proposed algorithm is illustrated with an example