Multi-gene Genetic Programming Based Modulation Classification Using Multinomial Logistic Regression

Automatic modulation classification (AMC) acts as a critical role in cognitive radio network, which has many civilian and military applications including signal demodulation and interference identification. In this paper, we explore a novel feature based (FB) AMC method using multi-gene genetic programming (MGGP) and multinomial logistic regression (MLR) jointly with spectral correlation features (SCFs). The proposed scheme includes two phases. In the training phase, MGGP generates various mappings to transform SCFs into new features and MLR selects some highly distinctive new features as MGGP-features and the mappings as feature optimization functions (FOFs). Meanwhile the corresponding MLR based classifier is output. In the classification phase, SCFs are transformed by the FOFs and the trained classifier identifies signal formats with MGGP-features. Compared to traditional FB methods, simulation results demonstrate that our proposed method yields satisfactory performance improvement and achieves robust classification, especially at lower SNR and fewer number of samples.