(Dr. Nader Sadegh, advisor)
"Design of an Online Trained Neural Network to Control an Unknown Plant"
The problem that is posed is to control a plant that is previously unknown, using the technology of artificial neural networks. The plant can be any system with inputs and outputs, such as a robotic arm or a motor. The controller has to learn the plantís behavior and keep track of it in order to be more and more efficient. A new kind of neural network is being formulated as a solution to this problem. This controller is trained online, i.e. directly during the control phase, and it keeps what it has learnt about the system inside its synaptic weights. It also has to be able to deal with changes in the plantís behavior, and adapt to it. To provide with some flexibility in the learning phase, a routine is derived, that is capable of recursively implementing the Levenberg-Marquard algorithm. A first control algorithm is derived, that is able to control a 1-dimensional plant, with one input and one output. It uses a finite difference approximation to estimate what the output of the plant will be, given a certain input. This algorithm is then extended to the 2-dimensional case, and is able to estimate the Jacobian of a system with two inputs and two outputs, involving more complicated computations. Then, a general algorithm is finally derived, that is able to deal with any dimension, with the final aim of being able to control dynamic systems.