Distributed Linear Bandits for Freshness-Aware Preprocessing in MEC-Enabled Digital Twin Networks

Mariam Yahya, Setareh MaghsudiarXiv 2026 · arXiv

TL;DR

This paper introduces a collaborative framework to keep Digital Twins (DTs) synchronized with physical IoT systems without drowning the network in raw data. By combining Multi-access Edge Computing (MEC) for data preprocessing with a distributed "linear bandit" learning algorithm, the system identifies the most efficient CPU allocation strategies that satisfy strict data freshness requirements (Age of Information) across multiple task classes.

What Problem Does This Paper Address?

Digital Twins—virtual replicas of physical systems—are only as good as the data they receive. In massive IoT networks, transmitting every bit of raw sensor data to a central DT server creates a massive bottleneck, leading to high latency and "stale" virtual models.

Preprocessing data at the edge (on MEC servers) can reduce this volume, but it introduces a new challenge: how should an MEC server divide its limited CPU power among different types of tasks? Some tasks might be computationally heavy but less time-sensitive, while others are light but require immediate updates to maintain freshness. Because the actual volume and complexity of these tasks are unpredictable (stochastic), finding the "sweet spot" for resource allocation is a moving target.

Methods

The authors propose a two-stage "mutual support" framework between the DT and MEC servers:

  1. Feasibility Filtering (The DT's Job): The DT server uses known bounds on workloads and freshness requirements to define a set of "freshness-feasible" CPU allocation strategies. These are strategies guaranteed to keep the Age of Information (AoI)—the time elapsed since the last successful update was generated—within acceptable limits.
  2. Collaborative Optimization (The MECs' Job): MEC servers act as agents in a distributed linear bandit problem. They treat different CPU allocation strategies as "arms" in a multi-armed bandit scenario.

The core innovation is the use of the DistLinGapE (Distributed Linear Gap-based Exploration) algorithm. Unlike standard bandit problems where agents learn in isolation, MEC servers share their observations about preprocessing delays. Because the problem is modeled as a linear bandit, the servers can use the mathematical structure of the task classes to infer the performance of one allocation strategy based on results from another, significantly speeding up the learning process.

Fig. 1: The system model.

Datasets

The study evaluates the framework using a simulated environment rather than a static public dataset, which is standard for testing dynamic resource allocation algorithms. The setup includes:

  • Environment: A 600m x 600m area with 4 MEC servers and 16 IoT clusters.
  • Task Diversity: 8 distinct task classes with varying computational requirements (cycles per bit) and freshness thresholds (ranging from 350ms to 400ms).
  • Workload: Stochastic task generation using Bernoulli distributions and uniformly distributed task sizes (40-60 kbytes).
  • Network: 30 Mbit/s uplink for IoT devices and 100 Mbit/s backhaul for MEC-to-DT communication.

Key Insights and Ideas

  • Mutual Support is Key: The DT provides the "safety rails" by defining what is feasible, while the MEC servers provide the "intelligence" by learning what is optimal. This separation of concerns simplifies the learning problem.
  • Linearity Reduces Exploration: By modeling the expected delay as a linear function of the CPU allocation and task features, the system doesn't have to try every possible strategy. It can "guess" the performance of similar strategies, which is crucial in time-sensitive environments.
  • The Power of Collaboration: The most significant "aha" moment is the near-linear speedup. By sharing scalar feedback (the total normalized delay) rather than raw workload data, MEC servers can reach a consensus on the best strategy much faster than if they worked alone, without needing to share private, raw data.

Key Results

  • Faster Convergence: DistLinGapE identified the optimal CPU allocation strategy in an average of 1,354 rounds, compared to 5,127 rounds for servers learning independently.
  • Efficiency Gains: The learned strategy reduced the mean class execution time to 38.34 ms, a significant improvement over uniform allocation (47.74 ms) or random selection (47.96 ms).
  • Scalability: The "learning speedup" was nearly proportional to the number of participating MEC servers. For instance, using 8 collaborating servers resulted in a 6.96x speedup compared to a single agent.
  • Freshness Guaranteed: Because the learning happens within a "freshness-feasible" set provided by the DT, the system maintained Age of Information requirements even during the exploration/learning phase.

Limitations and Future Work

  • Homogeneity Assumption: The current model assumes MEC servers have similar computing capacities and face similar average workloads. In real-world deployments, servers might be heterogeneous, which would require more complex collaborative algorithms.
  • Fixed Allocation per Slot: The CPU rates are fixed for the duration of a time slot. The authors suggest that dynamic, intra-slot redistribution of unused CPU cycles could further reduce delays.
  • Stationary Statistics: The framework performs best under stationary task-class assignments. Adapting to rapidly shifting network topologies or task distributions remains a challenge for future research.

Summary updated

Abstract

Digital twins (DTs) require timely updates from Internet of Things (IoT) devices to accurately represent the state of the corresponding physical systems, but transmitting large data volumes can compromise freshness. Moreover, different task classes have distinct computational demands and freshness requirements, while stochastic task generation and sizes complicate resource allocation. In this paper, we propose a freshness-aware framework in which multi-access edge computing (MEC) servers preprocess IoT data, reducing transmission volume and DT-server processing load. The framework establishes mutual support between DT and MEC servers. The DTs determine feasible CPU-allocation strategies guaranteeing class-specific age-of-information (AoI) requirements under specified workload, delay, and update-generation bounds. MEC servers collaboratively identify the feasible strategy minimizing expected freshness-normalized preprocessing delay. We formulate CPU-strategy selection as a fixed-confidence distributed best-arm identification problem in linear bandits, with MEC servers as agents and CPU allocations across task classes as arms. We employ DistLinGapE, a distributed collaborative learning method, to coordinate exploration and share observations. Numerical results show that the method identifies the optimal feasible strategy and reduces preprocessing delay when deployed under stationary task-class assignments and workload statistics. Collaboration yields an identification speedup nearly proportional to the number of participating MEC servers in the evaluated settings.
BibTeX
@misc{yahya2026distributed,
  author = {Mariam Yahya and Setareh Maghsudi},
  title = {{Distributed Linear Bandits for Freshness-Aware Preprocessing in MEC-Enabled Digital Twin Networks}},
  year = {2026},
  eprint = {2611.05115},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05115},
}
Figure