Application of C1DAE-ANIL in End-to-End Communication of IRS-Assisted UAV System

Due to the uninterrupted flight of the unmanned aerial vehicles (UAV) and the varying phase of the intelligent reflecting surface (IRS) in the IRS-assisted UAV communication system, it is very difficult to estimate the channel information of the system accurately. The traditional solutions try to model the whole communication system as an autoencoder including encoder, channel, and decoder. However, the autoencoder needs to be retrained when the environment changes and the generalization ability is lousy. To solve above problems, a novel C1DAE-ANIL algorithm is proposed to estimate the channel information of the IRS-assisted UAV communication system. On one hand, one-dimensional convolution is introduced to the traditional multi-perceptron autoencoder (MLPAE) for enhancing the encoding and decoding performances of the one-dimensional convolutional autoencoder (C1DAE). On the other hand, the meta-learning is applied to the C1DAE, and the almost no inner loop (ANIL) algorithm is used to find an optimal initial parameter vector of the C1DAE to realize the rapid training in dynamic environments. Compared with the existing contenders, the simulation results show that the end-to-end communication model using the C1DAE-ANIL algorithm not only improves the accuracy of channel estimation but also greatly accelerates the training process in dynamic environments.