| Key objectives of this thesis are: (a) To formulate efficient neural network modeling algorithms that can facilitate automatic generation of accurate RF and microwave neural models, and (b) To develop robust neural modeling techniques that would enable everyday users to learn, apply and gain experience with neural network without a special expertise in this field.;Major contributions of the thesis include the Automatic Multilayer Selection (AMS) technique and the combination of the proposed AMS technique with an existing automatic model generation (AMG) algorithm to automatically develop a compact neural network model. A microwave component model can be created starting with a simple neural network structure and then proceeding with neural network training in a systematic manner. During each stage, the AMS algorithm utilizes the training error criteria to automatically adjust the number of layers or the number of neurons in each layer of the neural network structure and consequently uses AMG algorithm to train a model to meet a user-desired accuracy. By combining the proposed AMS with Automatic Data Generation (ADG) and AMG algorithms, a more efficient and automated modeling framework is developed to generate neural network models that accurately match training data, with minimal human intervention. (Abstract shortened by UMI.). |
