Abstract
We propose two unsupervised learning control schemes for a nonlinear system based on a performance function, which is given by the total cost without presenting a target signal. A feedback neurocontroller self-optimizes the system through minimizing the performance function. The first uses a generalized back-propagation method which calculates the sensitivity of the performance function with respect to the cennection weights. The second uses difference approximation of the sensitivity which is given by suitable perturvations for the connection weights.
We apply these two schemes for controlling an inverted pendulum whose initial position is hanging. Each scheme accomplishes stabilization of the pendulum at the center of the rail. The neurocontroller acquires a strategy of accerative oscillation round the hanging position to raise the pendulum within a short rail, and it damps the oscillation and leads the pendulum to the center of the rail by a skillful balance about the standing position.
