| A neural network classifier or non-linear discriminant analysis network is sought. Classical network neurons are aggregations of a weight multiplied by an input value and then controlled via an activation function; with the addition of a bias node within each network's layer. This thesis learns everything within the neuron using a variant of Genetic Programming called Gene Expression Programming (including the connections between neurons) for a fixed number of layers and a fixed number of neurons per layer. That is, this thesis does not explicitly use weights or activation functions within a neuron, nor does it construct an explicit bias node. More precisely, this thesis claims that "weights and activation functions do not need to be specified explicitly when learning the functions associated to neurons within a neural network, as is done in the classical case". Promising results are reported for the 1 and 2 class cases. For example, the investigated one class problem is for the determination of underground caves, in which, for some measuring locations, cave membership is known and for others it is not. The multiple class cases are harder and deserve further investigation, but, even so, some good results have been published. |
