Abstract:
This thesis defines the procedure of identifying optimum neural network design by incorporating data mining. Identify patterns in the training dataset and establish relationships in the training dataset used to train the neural network. Then the information obtained will be used to determine the architecture of the artificial neural network.
There is no evidence supporting any method to determine the optimum ANN architecture. Contemporary approaches are restricted and consume a lot of time. Proven successful approaches offer a solution to given problems but under a given environment. No verifiable theory exists, explaining the design of an ANN.
This scientific research related to artificial intelligence seeks to utilize pattern recognition methods to define the structure of an ANN. It involves clustering methods to group training dataset to identify some common features, that can be grouped depending on given conditions. The results obtained through clustering as far as multilayer ANN structure is concerned.
A regression model is adopted to increase how predictable the projected is, by adopting results from grouping to define the structure of ANN. Depending on hypothesis testing through F-test for regression, the conclusion is arrived at; the neurons in hidden layers and the quantity of hidden layers themselves depend on the factors that are considered through the projected clustering method.
The proposed method reduces the time allocated for network design. This method can make the design simple and easy and possible for a Non-specialist designer. The focus was to develop a fast solution (in terms of learning iterations) maintaining also an acceptable efficiency.
Keywords: Artificial intelligence, machine learning, neural networks architecture, data mining, clustering methods, multi-layer neural network, pattern recognition, regression analysis.
