Automatic Modulation Recognition Using Neural Architecture Search
Shengyun Wei, Shun Zou, Feifan Liao, Weimin Lang, Wenhui Wu · 2019
Deep learning method has been proposed to deal with modulation recognition tasks instead of likelihood-based and feature-based approaches recently, but many researchers usually suffer from complex architecture engineering as designing neural networks still requires extensive expert knowledge. Neural architecture search, which uses a neural network to tune or compose other neural network architectures automatically, could be convenient for many researchers to find good deep neural networks for different data at hand. In this paper, we proposed a neural architecture search based modulation recognition framework. We use Bayesian optimization and genetic algorithm to search convolutional neural network architecture for synthetic datasets generated with GNU radio. Four existing approaches, including convolutional Neural Network (CNN), Inception Modules (Inception), Residual networks (Resnet) and Long Short-Term Memory (LSTM) based recognition algorithms, are involved for performance comparison. Simulation results indicate the effectiveness of the proposed framework, and the proposed approaches achieve significantly higher performance without manual feature engineering and architecture engineering.