Optimizing Generative AI Networking: A Dual Perspective with Multi-Agent Systems and Mixture of Experts

TL;DR

The paper addresses the high computational burden and limited adaptability of centralized GenAI in next-generation networks. It proposes a hybrid framework that combines Multi-Agent Systems (MAS) for dynamic, distributed task coordination with Mixture of Experts (MoE) for expert-driven execution, culminating in a novel MoE-PPO approach for 3D object generation and data transfer. Key contributions include a comprehensive architectural and applicability analysis of MAS and MoE for GenAI in networking, plus a case study demonstrating improved performance and cost efficiency under diverse network conditions. The work highlights the potential of MAS-MoE hybrids to enable scalable, adaptive GenAI-enabled networking and points toward semantic-communication-oriented research as a practical long-term impact.

Abstract

In the continued development of next-generation networking and artificial intelligence content generation (AIGC) services, the integration of multi-agent systems (MAS) and the mixture of experts (MoE) frameworks is becoming increasingly important. Motivated by this, this article studies the contrasting and converging of MAS and MoE in AIGC-enabled networking. First, we discuss the architectural designs, operational procedures, and inherent advantages of using MAS and MoE in generative AI to explore its functionality and applications fully. Next, we review the applications of MAS and MoE frameworks in content generation and resource allocation, emphasizing their impact on networking operations. Subsequently, we propose a novel multi-agent-enabled MoE-proximal policy optimization (MoE-PPO) framework for 3D object generation and data transfer scenarios. The framework uses MAS for dynamic task coordination of each network service provider agent and MoE for expert-driven execution of respective tasks, thereby improving overall system efficiency and adaptability. The simulation results demonstrate the effectiveness of our proposed framework and significantly improve the performance indicators under different network conditions. Finally, we outline potential future research directions.

Figures (4)

• Figure 1: Structure and function of MoE and MAS for GenAI. MoE framework is described as a dynamic network where a central gating mechanism delegates tasks to specialized experts, focusing on feedback for performance optimization and benefits such as precision and customization. MAS is described as a collaborative environment where agents use shared protocols for interaction, decision-making, and context awareness.
• Figure 2: The Structure of system model and proposed solution framework. Multiple NSPs (each NSP responsible for a different type of food) collaborate to collect, process, and tramsit 3D objects of food. The left part of the figure shows the process from collecting various 2D images from different angles for each NSP to generate these images into a 3D object. The right part introduces MoE framework, in which multiple experts, guided by a gating network, optimize specific subtasks, such as image selection and power allocation, to improve efficiency and reduce overall transmission costs.
• Figure 3: The efficacy of the multi-agent-enabled MoE framework for the 3D object generation and transmission system. Each NSP operates as an independent agent, with the first expert responsible for selecting the most conducive images for 3D reconstruction. These selected images are transformed into 3D objects, transferred as point clouds by a second expert specializing in data transmission.
• Figure 4: (a) The cumulative reward versus the number of episodes; (b) The cost consumption versus the number of NSPs.