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. This article explores the synergy between MAS and MoE in AIGC-enabled networking, discussing their architectural designs, operational procedures, and inherent advantages. We propose a novel multi-agent-enabled MoE-proximal policy optimization (MoE-PPO) framework for 3D object generation and data transfer scenarios. The framework utilizes MAS for dynamic task coordination among network service provider agents and MoE for expert-driven execution, improving system efficiency and adaptability. Simulation results demonstrate that the proposed framework significantly improves performance indicators under varying network conditions.
The Challenge: Resource-Hungry GenAI in Next-Gen Networks
Generative AI (GenAI) models, such as large language models and text-to-image generators, are transforming digital services but present significant hurdles for deployment in B5G and 6G networking. These models are notoriously resource-intensive, often requiring billions of parameters and massive computational power.
Key challenges include:
- Unmanageable Complexity: The scale of GenAI models makes them difficult to execute on resource-constrained edge devices.
- Low Adaptability: Centralized models often fail to react quickly to dynamic network conditions or diverse user requirements.
- Performance Bottlenecks: Centralized processing can lead to data transfer "traffic jams," resulting in high latency for real-time applications like the Metaverse or AR/VR.
A Hybrid Solution: MAS and MoE
To address these issues, this research moves away from "one-size-fits-all" centralized models toward a distributed, specialized architecture. This is achieved by combining two powerful paradigms:
- Multi-Agent Systems (MAS): Acts as the "coordination layer." Multiple autonomous agents—such as different network service providers (NSPs)—collaborate and adjust their policies based on real-time environmental feedback.
- Mixture of Experts (MoE): Acts as the "execution layer." Instead of a single monolithic neural network, the system employs a gating mechanism to activate only the most relevant specialized sub-models (experts) for a specific task, maintaining high capacity with low active parameter counts.
The MoE-PPO Framework for 3D Generation
The core technical contribution is the MoE-Proximal Policy Optimization (MoE-PPO) framework. This framework is particularly effective for bandwidth-intensive tasks like 3D object generation. By generating and transmitting a single 3D object (approx. 7MB) rather than multiple high-resolution 2D images (approx. 20MB), the system significantly reduces bandwidth requirements while maintaining immersion.
In this framework, each NSP acts as an independent agent containing specialized experts. For 3D reconstruction, one expert is optimized for image selection (choosing optimal 2D angles), while another handles power allocation for data transmission. To accelerate learning, the system utilizes "max-propagation" to compute the largest generalized advantage estimation across all agents, ensuring the network learns from the most successful actions.
Performance Evaluation
Simulation results indicate that the MoE-PPO framework reaches optimal performance approximately 40% faster than standard multi-agent PPO benchmarks. Furthermore, as the number of service providers increases, the MoE-PPO approach consistently maintains the lowest total operating cost compared to greedy or random strategies.
In 3D reconstruction tests, the system achieved high fidelity across multiple metrics. For instance, a coffee cup reconstruction (NSP 2) achieved a PSNR of 5.196 and a Surface Quality score of 8.337, demonstrating that computational efficiency does not come at the expense of visual accuracy.
Future Directions
While the MAS-MoE hybrid shows great promise, several areas for future research remain:
- Multimodal Complexity: Integrating text, video, and audio seamlessly into the same architecture.
- Autonomous Recalibration: Developing systems that can dynamically swap or re-train MoE components as mission requirements change.
- Semantic Communication: Optimizing for the transfer of "meaning" and "intent" rather than just raw data, ensuring content remains intact in noisy smart city environments.
Summary updated