Federated Prompt-based Decision Transformer for Customized VR Services in Mobile Edge Computing System

TL;DR

The paper tackles resource allocation for customized VR services in MEC by proposing a QoE metric that blends latency, user attention, and preferred resolutions. It reframes the allocation problem as offline RL and introduces FedPromptDT, a federated, prompt-based Decision Transformer that generalizes across heterogeneous user environments. FedPromptDT uses FL to pre-train a DT model offline and employs prompts (environmental and user-preferred cues) to adapt online without re-training, achieving superior QoE and stability compared to baselines. The approach demonstrates strong cross-environment generalization, privacy preservation, and adaptability to varying numbers of users and levels, indicating practical viability for scalable MEC VR services.

Abstract

This paper investigates resource allocation to provide heterogeneous users with customized virtual reality (VR) services in a mobile edge computing (MEC) system. We first introduce a quality of experience (QoE) metric to measure user experience, which considers the MEC system's latency, user attention levels, and preferred resolutions. Then, a QoE maximization problem is formulated for resource allocation to ensure the highest possible user experience,which is cast as a reinforcement learning problem, aiming to learn a generalized policy applicable across diverse user environments for all MEC servers. To learn the generalized policy, we propose a framework that employs federated learning (FL) and prompt-based sequence modeling to pre-train a common decision model across MEC servers, which is named FedPromptDT. Using FL solves the problem of insufficient local MEC data while protecting user privacy during offline training. The design of prompts integrating user-environment cues and user-preferred allocation improves the model's adaptability to various user environments during online execution.

Figures (7)

• Figure 1: Illustration of the FL-based MEC system. The system takes FL to enhance the generalization of its decision model on MEC servers. The decision model is considered a Decision Transformer (DT) that tokens states, actions, and returns of MEC servers using their corresponding linear embedding layers to predict actions for resource allocation.
• Figure 2: (a) Illustration of the attention-based VR video content for the $k$-th user served by $e$-th edge server at time slot $t$. (b) Different users have diverse eye gaze locations while viewing the same 360° VR video content from Xu2018video360, indicating heterogeneous attention tile size. (c) Different user levels are required to process different tile sizes, resulting in heterogeneous computation and communication requests.
• Figure 3: Illustration of the FedPromptDT-empowered MEC system. Digital twin allows the MEC system to monitor the system's real-time state, facilitate perceiving user environments, and collect historical data. During online execution, states, actions, and returns are tokenized by their corresponding linear embedding layers and added with episodic timestep encoding. These tokens are fed into the pre-trained FedPromptDT model to autoregressively predict actions with the prompt based on current user environments. During local training in FL, the MEC server collects a batch that concatenates the prompt and training trajectories in each local iteration. After that, it updates FedpromptDT iteratively on various user environments throughout the local training process. More details for offline training and online execution are in Section \ref{['sec:OfflineFL_online_ex']}.
• Figure 4: (a): MEC MA rewards on different user numbers; (b): MEC MA rewards on different user levels.
• Figure 5: (a): MEC MA rewards on different QoE and hfQoE thresholds in ($\textbf{P}_0$); (b): MEC MA rewards on different MEC bandwidth and frequency capability.