Balanced Neural Architecture Search and Its Application in Specific Emitter Identification
Mingyang Du, Xikai He, Xiaohao Cai, Daping Bi
Abstract
Since the performance of a single neural network can vary unexpectedly corresponding to different classification tasks, the network with fixed structure may lack the flexibility and often lead to overfitting or underfitting. To deal with the limitation of the fixed neural network structure on classifying radar signals in different electromagnetic environments, we propose an automatic network architecture search (NAS) mechanism and apply it in specific emitter identification (SEI). In this paper, a ‘`block-cell’' structure and a controller based recurrent neural network (RNN) are utilized to design models automatically according to external environment. Like some state-of-the-art neural networks, the combination of some operations like convolutional layers and pooling layers with various sizes, builds the search space. A progressive strategy is utilized to design child models from simple to complicated. The controller then evaluates child models and chooses the most promising ones as candidates to build the final network. A balance function, which considers both the model accuracy and efficiency when designing networks, is also proposed to replace the validation accuracy as the reward to feed back to the controller. Simulations and experiments indicate that the proposed NAS mechanism with the balance function outperforms the conventional algorithms in classifying radar signals in different environments.