Abstract
As communication technologies evolve, the AI agent communication network (ACN) is expected to play a critical role in enabling future wireless services. These networks must meet diverse connectivity demands, provide enhanced intelligence, and provide reliable communication support for AI agents. In this work, we propose an integrated computation-communication (ICC) architecture that jointly addresses baseband signal processing and neural network (NN) inference within a shared computing framework. A unified modeling approach is established by combining multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) channel representation with neural network–based application modeling. On the communication side, we derive spectral efficiency and channel capacity expressions that incorporate quantization error, transmit power, and noise, thus characterizing hardware-induced constraints on channel performance. On the computation side, we analyze the model scaling in terms of depth, width, and input resolution. Building upon this framework, we formulate the joint resource allocation challenge as a Markov Decision Process (MDP). We introduce a learning-driven allocation algorithm that uniquely combines Gibbs sampling with a Proximal Policy Optimization (PPO) based reinforcement learning agent to solve this MDP. This enables the system to dynamically distribute resources across communication and computation, adapting to variations in datasets, Signal-to-Noise Ratio (SNR) regimes, model configurations, and scenario constraints. The proposed approach improves both task completion time and prediction accuracy compared to conventional fixed resource allocation schemes. Theoretical analysis and simulation results validate the effectiveness of the proposed architecture, highlighting its potential for efficient cross-layer co-design in next-generation ACNs.
Funding Information
  • National Natural Science Foundation of China (62331009)

This publication has 28 references indexed in Scilit: