Robust training of microwave neural models

Neural networks recently gained attention as a fast and flexible vehicle to microwave modeling and design. Neural network models can be developed by learning from microwave data, through a process called training. The trained models can be used during microwave design to provide instant answers to the task they learnt. This article addresses certain key challenges in developing RF/microwave neural models. An iterative multistage (IMS) approach including a macro-level process and a stage-level process is proposed. At the macro-level, the IMS decomposes the complicated original task into several simpler subtasks or stages and at the stage-level, the IMS utilizes a variety of neural network structures and effective training techniques, including several existing techniques and a new Huber quasi-Newton (HQN) technique. The proposed HQN allows for the IMS approach to model only smooth portion of the problem behavior in one of the training stages, ignoring sharp/sudden variations. The advantages of the proposed microwave-oriented modeling techniques are demonstrated through examples.