Abstract
The emergence of agentic artificial intelligence (AI) necessitates reliable physical communication links to enable high-reliability and low-latency collaboration among autonomous agents, particularly in high-mobility scenarios which suffer severe Doppler effect. Affine frequency division multiplexing (AFDM) is a chirp-based multi-carrier waveform designed to address doubly-selective channels under high-mobility conditions. In this paper, we propose the joint channel estimation and signal detection (JCESD) methods for achieving high-reliable AFDM transmission considering both integer and fractional Doppler cases. For the integer Doppler case, we introduce a threshold-based channel estimation method and propose graph neural network (GNN)-based detector, which is motivated by the sparsity of the effective channel matrix. For the fractional Doppler case, we propose a complex-valued fully-connected neural network (CVFCNN) framework-aided JCESD scheme utilizing a complexvalued neural network to solve the inter-pilot-data interference (IPDI) problem. Specifically, we estimate subchannels individually and reconstruct the channel matrix based on the channel characteristics inherent to AFDM, thereby avoiding overlapping between different subcarriers. Furthermore, our proposed complex-valued GNN is trained to simultaneously enhance the performance of both channel estimation and signal detection. Comprehensive evaluations are conducted under diverse transmission configurations to assess the channel estimation and signal detection capabilities of the proposed algorithm. The simulation results indicate that our method yields a 2∼4 dB improvement in BER performance over other state-of-the-art schemes.