DRL-based intelligent resource allocation for physical layer semantic communication with IRS

Abstract

Semantic communication and intelligent reflection surface (IRS) are considered to be promising technologies to solve the scarce spectrum resource problem for the sixth-generation (6G) communication networks. However, there is few research on semantic resource allocation for IRS-enhanced communication networks, which leverages the efficient spectrum utilization of both semantic communication and IRS. In this paper, the resource allocation problem in the IRS-assisted semantic communication network is investigated, and effective semantic spectral efficiency (ES-SE) is defined considering desired semantic similarity for downstream semantic tasks. For the purpose of maximizing the ES-SE, the selection of DeepSCs, allocation of subchannels, reflection array elements of the IRS and transmit beamforming of the base station (BS) are jointly optimized. Considering the necessity of real-time performance and full-link intelligence, a two-stage intelligent approach using dueling double deep Q networks (D3QN)-soft actor critic (SAC) is proposed to tackle the tough resource allocation problem with non-linear programming and coupled variables. Simulation results validate the effectiveness of our designed IRS-assisted semantic communication scheme and demonstrate the superior performance of our proposed intelligent approach.