Residual Vector Quantized Diffusion Model-Based Joint Source-Channel Coding for Task-Oriented Agent Communication

In this paper, a novel joint source-channel coding (JSCC) scheme based on the residual vector quantized diffusion model is proposed for task-oriented agent communication. To make the agent communication compatible with existing digital communication systems, the feature vectors encoded by the parameterized encoder are quantized using multiple learnable codebooks, forming the quantized features. Furthermore, to improve the reliability of quantized feature transmission, a discrete channel denoising diffusion model (DCDDM) is designed. The DCDDM can learn the distribution of the sender agent’s quantized features, and then utilize this learned knowledge to correct the receiver agent’s received quantized features. Based on the corrected feature vectors, the parameterized decoder recovers task-related information for the receiver agent’s subsequent decision-making. Finally, the model parameters of the proposed scheme are trained within the reinforcement learning framework. In addition, we design a task-oriented agent communication proof-of-concept prototype. Simulation and experiment results show that the proposed scheme has better generalization capability than the existing schemes. That is, with the proposed scheme, the transmission of task-related information is more reliable, thus the multi-agent systems can collaborate better under different channel conditions.