Deep Reinforced Feature Compression and Channel Equalization for Semantic Communications

In the sixth-generation (6G) wireless communication system, semantic communication (SC) aims to improve the data transmission rate and processing efficiency. SC can provide enhanced efficiency for data transmission, especially when the signal-to-noise ratio (SNR) is low, compared to conventional wireless communication systems, by employing deep learning techniques. However, these systems often had limitations in the lack of consideration of dynamic compression and nonlinear channel effects. In order address these challenges, deep reinforcement learning (DRL)-based dynamic feature compression and channel equalization methods are proposed. To this end, two strategies for dynamic compression are proposed and evaluated for the favorable situations in each strategy. In the proposed SC framework, the transmission efficiency can be enhanced by adaptively adjusting the vector length or selectively transmitting image segments. Further, the proposed method can optimize task-oriented performance accuracy under nonlinear channel conditions and maintain transmission efficiency by mitigating worse channel effects. The simulation results demonstrated that the proposed SC system can reduce computational complexity by about 40% and improve average performance by around 20%, respectively, compared to state-of-the-art methods under fading channel conditions.