ABSTRACT
Neural network is a term describing a computer technique that uses a large number of simple processing elements. The processing elements, called neurons or nodes, are highly interconnected and operate in parallel. Each connection has an associated weight that determines the amount of interaction between the neurons. This physical structure is inspired by the brain, albeit an extremely simplistic representation. Neural nets attempt to mimic properties of the brain such as learning and generalization. There are many different types of neural networks. Depending on the control application, whether it be for system identification, modelling, or as a controller, a certain category of network may be more appropriate. The major feature of a neural network is its ability to learn. Supervised Learning is a process where the input conditions and desired outputs are presented to the network. The network attempts to determine the association between the inputs and outputs by adjusting the weights, as opposed to developing a mathematical or analytical relationship. Depending on the ability of the net to learn the relationship, the net can have generalization properties. Hence an adequate result can be generated for new input conditions. It is their generalization and learning properties that makes neural networks an attractive tool for intelligent control. The current work involves using the multilayer feedforward type of network, Rumelhart and McClelland [5]. Despite being relatively simple it is a powerful network with supervised learning and generalization capabilities. A number of approaches for implementation of neural networks for control have been developed. These can be summarized into the following categories Karasi [6], Barto [7]:
