| The design of Artificial Neural Networks (ANNs) architecture can be a complex and time-consuming task. Typically many trials are involved to select the desired architecture of an ANN for a desired solution or the expected error. In recent years, there has been considerable interest in solving problems in mechanical engineering using an ANN. Design and production problems in this field include large linear computations and complex non-linear equations, the solutions of which traditionally take a long time to discover. This concept of avoiding lengthy computation time can also be used for control applications. In this thesis, a simple and novel procedure for ANN layer partition is proposed. Following this procedure, here two novel Genetic Algorithms (GA) approaches are presented to optimize the connection weights of ANNs in order to prevent a lengthy trial-and-error training process. The first approach, an original variant (in the sense of being based in the proposed layer partition) of a standard GA implementation to optimize connection weights of an ANN produced very promising results. The second approach, a truly novel GA implementation produced even better results than that of the first approach. To verify the methodology, the practical problem of designing the spindle of a cylindrical grinding machine is solved through the proposed ANN and GA implementations. The connection weights of the ANN are extracted and optimized through the proposed GA approaches. The hypothesis designed to optimize the connection weights of an ANN proved to be practical as the experimental results perfectly matched with the expected results. |
