Mixture of Experts for Network Optimization: A Large Language Model-enabled Approach

TL;DR

The paper addresses the challenge of optimizing diverse user requirements in 6G networks where training separate DRL models is costly. It introduces an LLM-enabled Mixture-of-Experts framework that coordinates a pool of DRL experts via an LLM-based gate to formulate objectives, select relevant experts, and fuse their inferences, thereby reducing the need for new task-specific models. The approach is validated on a maze navigation task and an NSP utility maximization scenario, showing improved decision quality, adaptability, and energy efficiency over conventional gate-network MoE configurations. This work highlights a scalable, user-centric path for network optimization by harnessing the reasoning capabilities of LLMs to manage expert DRL ensembles on edge infrastructure.

Abstract

Optimizing various wireless user tasks poses a significant challenge for networking systems because of the expanding range of user requirements. Despite advancements in Deep Reinforcement Learning (DRL), the need for customized optimization tasks for individual users complicates developing and applying numerous DRL models, leading to substantial computation resource and energy consumption and can lead to inconsistent outcomes. To address this issue, we propose a novel approach utilizing a Mixture of Experts (MoE) framework, augmented with Large Language Models (LLMs), to analyze user objectives and constraints effectively, select specialized DRL experts, and weigh each decision from the participating experts. Specifically, we develop a gate network to oversee the expert models, allowing a collective of experts to tackle a wide array of new tasks. Furthermore, we innovatively substitute the traditional gate network with an LLM, leveraging its advanced reasoning capabilities to manage expert model selection for joint decisions. Our proposed method reduces the need to train new DRL models for each unique optimization problem, decreasing energy consumption and AI model implementation costs. The LLM-enabled MoE approach is validated through a general maze navigation task and a specific network service provider utility maximization task, demonstrating its effectiveness and practical applicability in optimizing complex networking systems.

Figures (6)

• Figure 1: Network optimization strategies. Part A demonstrates the drawbacks of training distinct AI models for different user requirements, emphasizing the costs of excessive AI model deployment. Part B presents our LLM-enabled MoE approach, using a limited set of DRL models to efficiently address a variety of user tasks.
• Figure 2: Workflow of the proposed LLM-enabled MoE framework: Upon receiving a text-based description of their requirements, the LLM processes and reasons about the user's needs, identifying the experts necessary for the task at hand and determining their decision-making weights.
• Figure 3: LLM-enabled MoE framework demonstration in a maze navigation task. An ensemble of DRL models trained on diverse tasks serves as a set of expert models accessible to the LLM. The LLM analyzes and infers user tasks, leveraging combinations of expert models to address the final objectives.
• Figure 4: System model. Part A shows the Service Market Model, illustrating the interaction between the NSPs and Users. Part B represents the Wireless Network Model, wherein we consider a BS with $M$ antennas providing services to a user device. Part C shows various user requirements under different scenarios, which affect the payment structure in Part A and the optimal power allocation strategies in Part B.
• Figure 5: Mixture of mission experts rewards and mission completion rate comparison.

Theorems & Definitions (2)

• Proposition 1
• Proposition 2