A neural-net approach to supervised learning of pole balancing

Abstract

It is shown how artificial neural nets can be used to solve a difficult learning problem. The task is to balance a pole that is hinged to a movable cart by applying either a left or a right force to the cart. The control process consists of developing pattern formations to give the required motor drive control. The latter is implemented with a connectionist net of the Rumelhart semilinear feedforward type. At each instant in time, the values of a training set of the system's state variables are processed into a single pattern which in turn is applied to the input layer of the connectionist net. The response, at the output layer of the net, is used as the control signal for that instant, During the learning period, the system is controlled by a human operator and the neural net learns to mimic human control by backpropagating the human's decisions through the network and updating the synaptic weights. The authors test the approach and conclude that the neural-net embedded rule is more effective in relation to the other methods used, especially in its ability to respond to changing system parameters.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>