Two-Subnet Scheme Based on Conditional GAN for Channel Agnostic End to End Communication

This paper proposes a two-subnet scheme based on conditional generative adversarial nets (CGAN) for channel agnostic end to end communication to improve the bit-error rate (BER) performance with lower complexity. Specifically, a two-subnet scheme for joint optimization of transceiver is first introduced. The entire end to end communication system is implemented by neural networks, and the receiver is composed of two subnets: the equalization subnet R
Equ
and the decoding subnet R
Dec
. Then, a two-stage training strategy is proposed. In training stage I, the transmitter T is directly concatenated to the subnet R
Dec
to address the challenge of updating transmitter parameters through unknown channels. In training stage II, the generator from the subnet R
Equ
which adopts CGAN is deployed to accomplish the estimation and equalization. Simulation results demonstrate that the two-stage training strategy can effectively recover long transmission sequences, and the BER performance precedes both the traditional method based on expert knowledge and the state-of-the-art end-to-end method. Furthermore, the proposed method greatly reduces the retraining complexity since merely the subnet R
Equ
need retrain when the channel scenarios change. It should be noted that the proposed method can be easily applied to any OFDM systems or other practical channels. This paper also develops the parameters updating mechanism.